# Vector engine benchmark: eshoponweb Run 2026-10-06T14:27:24Z (run-all). Command line: ``` ~/gvb-work/lanes/v7-final/src/GenericVectorBuilder.Bench/bin/Release/net10.0/GenericVectorBuilder.Bench.dll run-all --pipeline eshoponweb --queries golden --seed 702 --out ~/ForClaude/GenericVectorBuilder/bench-results ``` - Machine: bench-host, Intel(R) Xeon(R) CPU E5-1620 v3 @ 3.50GHz (8 logical CPUs), 62.7 GiB RAM, GPU NVIDIA GeForce RTX 3060, 12288 MiB, Ubuntu 24.04.5 LTS, kernel 6.8.0-142-generic, .NET 10.0.12, load average at start 3.30 3.89 3.49 - Data: 524 vectors x 1024 dims, collection `gvbbench_eshoponweb`. Read READ-ONLY from GenericVectorBuilder.dbo.gvb_eshoponweb in ChunkId order, vectors as native SqlVector, 0.2 s. - Queries: golden: 20 labelled questions from ~/ForClaude/evalkit/questions_golden.json, embedded with qwen3-emb-0.6b (cached in ~/gvb-data/bench-cache/golden-eshoponweb-68df42ca69efa079.json, no embedding calls). Top 10. Throughput at concurrency 1, 8 for 20 s each. - Request speed on a small collection (524 vectors): Measured end to end through each engine's .NET client; at this size it reflects per-request cost including the client library, not index scaling. - Client CPU per search is the CPU time the test's .NET client itself used for each search, measured in the same pass as the figure beside it. Where it is close to the latency, the client library is a large part of what is measured. For an embedded engine (DuckDB, sqlite-vec) the engine runs inside the client process, so its figure is the engine's own CPU time, not client overhead. - Ground truth: brute force over every vector in memory, 0.0 s for all 20 queries. - nDCG@10 of the exact answer itself (the ceiling for this embedder): 0.518 | engine | index | load rows/s | p50 ms | p95 ms | p99 ms | QPS@1 | QPS@8 | client CPU ms/search@1 | client CPU ms/search@8 | recall@10 | nDCG@10 | RAM | disk | |---|---|---|---|---|---|---|---|---|---|---|---|---|---| | opensearch | faiss HNSW float32, no compression, m=16, ef_construction=128, cosinesimil; search k=top, ef_search=100; 1 shard, 0 replicas; graph built at any segment size (approximate_threshold=0); force-merged to one segment after the load | 306 | 1.99 | 2.29 | 2.78 | 490.8 | 1516.0 | 0.46 | 0.50 | 1.000 | 0.518 | 2.57 GiB | 9.48 MiB | | redis | HNSW TYPE FLOAT32 M=16 EF_CONSTRUCTION=128, EF_RUNTIME=100 per query, cosine; exact mode = FLAT index built on first exact query | 11,736 | 0.39 | 0.49 | 0.70 | 2536.6 | 5067.3 | 0.56 | 0.37 | 1.000 | 0.518 | 253.7 MiB | 0 B | | qdrant-hnsw | HNSW m=16 ef_construct=100, hnsw_ef=server default, cosine; indexing_threshold_kb 1 and full_scan_threshold_kb 10 (server defaults are 10,000 each) so a small collection builds and walks its graph | 4,640 | 0.92 | 1.03 | 1.40 | 1076.6 | 3440.1 | 0.74 | 0.52 | 1.000 | 0.518 | 38.41 MiB | 164.1 MiB | | weaviate | HNSW maxConnections(M)=16 efConstruction=128, ef=-1 (dynamic: limit x 8 clamped 100..500), cosine, no quantization; approximate only (no exact mode) | 1,001 | 5.36 | 10.5 | 11.6 | 166.7 | 305.5 | 0.61 | 0.62 | 1.000 | 0.518 | 71.18 MiB | 2.96 MiB | | qdrant | exact scan: the builder's sink sends exact=true on every search, so no HNSW graph is used whether or not Qdrant has built one (see the index state) | 6,919 | 0.88 | 1.02 | 1.39 | 1113.8 | 3280.1 | 0.73 | 0.53 | 1.000 | 0.518 | 35.56 MiB | 196.08 MiB | | mariadb | VECTOR INDEX (HNSW variant) DISTANCE=cosine, M=16 (no ef_construction setting exists), mhnsw_ef_search=100 per statement (the ef 100 most engines here use, so the search effort matches; MariaDB's own default is 20; recall@10 at ef 100 falls as the set grows (random 1024-dimension vectors, measured 2026-10-04: 0.99 at 524, 0.89 to 0.92 at 2,000; an earlier run gave about 0.09 at 100,000)), mhnsw_max_cache_size 4G; exact mode = IGNORE INDEX full scan | 1,035 | 0.62 | 0.71 | 0.85 | 1578.0 | 5919.7 | 0.48 | 0.38 | 1.000 | 0.518 | 170.4 MiB | 22.01 MiB | | vespa | HNSW float32 tensor, prenormalized-angular (cosine), max-links-per-node=16, neighbors-to-explore-at-insert=128; search targetHits=top, ef=100 via exploreAdditionalHits; exact mode = approximate:false; vectors held in memory | 238 | 1.95 | 2.42 | 2.88 | 496.0 | 1505.7 | 0.94 | 0.88 | 1.000 | 0.518 | 2.8 GiB | 0 B | | oracle | HNSW in-memory neighbor graph NEIGHBORS=16 EFCONSTRUCTION=128, EFSEARCH=100 per query, cosine; exact mode = FETCH EXACT FIRST (full scan); Oracle Free caps itself at 2 CPUs (cpu_count 2 in V$PARAMETER, edition FREE in V$INSTANCE, 8 host CPUs in V$OSSTAT NUM_CPUS; the 2 CPU thread limit is Oracle's documented Free edition limit) | 1,955 | 0.93 | 1.12 | 1.87 | 1042.1 | 2655.3 | 0.89 | 0.65 | 1.000 | 0.518 | 2.16 GiB | 0 B | | milvus | HNSW M=16 efConstruction=128, ef=100, metric COSINE, Strong consistency searches; approximate only (no exact mode) | 1,178 | 2.32 | 2.89 | 3.83 | 408.5 | 1258.0 | 0.56 | 0.64 | 1.000 | 0.518 | 213 MiB | 10.86 MiB | | clickhouse | vector_similarity HNSW cosineDistance, quantization bf16, M=16 ef_construction=128, hnsw_candidate_list_size_for_search=256, rescoring off; exact mode = full scan with skip indexes off | 2,160 | 4.94 | 6.00 | 7.74 | 195.8 | 428.5 | 0.88 | 1.06 | 0.965 | 0.519 | 892.7 MiB | 2.57 GiB | | sql | exact VECTOR_DISTANCE cosine, no vector index (full scan) | 586 | 4.02 | 4.57 | 6.02 | 242.9 | 602.2 | 1.22 | 1.20 | 1.000 | 0.518 | 532.2 MiB | 5.15 MiB | | duckdb | HNSW (vss extension) FLOAT[n] metric=cosine m=16 ef_construction=128, ef_search=100 per connection, persistent (hnsw_enable_experimental_persistence=true, checkpoint_threshold=256MB); exact mode = array_cosine_similarity sequential scan; score = 1 - cosine distance; searches run concurrently, one connection per searcher (opened as searchers arrive, at most 32), writes run one at a time and never overlap a search | 1,486 | 3.68 | 4.07 | 5.27 | 269.1 | 576.0 | 5.03 | 6.90 | 1.000 | 0.518 | - | 2.9 MiB | | pgvector | HNSW vector_cosine_ops m=16 ef_construction=128, hnsw.ef_search=100 per query, float32 vector(n), cosine; exact mode = same query with index scans off (sequential scan) | 609 | 0.87 | 1.04 | 1.29 | 1125.7 | 4001.3 | 0.69 | 0.54 | 1.000 | 0.518 | 99.93 MiB | 7.56 MiB | | elasticsearch | HNSW float32, no quantization, m=16, ef_construction=128, cosine; search k=top, num_candidates=100; 1 shard, 0 replicas; force-merged to one segment after the load (at 1,024 dimensions a segment under 1,043 vectors gets no graph) | 416 | 1.33 | 1.52 | 1.76 | 739.4 | 2668.9 | 0.46 | 0.49 | 1.000 | 0.518 | 2.61 GiB | 2.39 MiB | | sql-diskann | DiskANN (preview) via VECTOR_SEARCH, cosine, build {"StartId":"306", "L":"48", "M":"8", "R":"48"}; the exact mode scans the same table | 611 | 3.69 | 4.20 | 5.57 | 263.3 | 682.3 | 1.03 | 1.11 | 0.965 | 0.526 | 529.9 MiB | 4.52 MiB | | chroma | HNSW M=16 ef_construction=128, ef_search=100 (Chroma default), cosine; approximate only (no exact mode) | 820 | 2.11 | 2.38 | 2.63 | 468.9 | 777.2 | 0.47 | 0.47 | 1.000 | 0.518 | 44.93 MiB | 414.16 KiB | | typesense | HNSW float32 (hnswlib), m=16, ef_construction=128, cosine; search k=top, ef=100; exact mode = filter ordinal:>=0 with flat_search_cutoff; index held in memory | 991 | 4.20 | 4.56 | 5.08 | 234.5 | 574.0 | 0.53 | 0.59 | 1.000 | 0.518 | 117.5 MiB | 0 B | | mongodb | vectorSearch index, HNSW maxEdges=16 numEdgeCandidates=128, float32 binData, cosine, numCandidates=20x hits (min 100); exact mode = $vectorSearch exact:true | 2,896 | 1.41 | 1.66 | 1.99 | 693.4 | 2115.3 | 0.31 | 0.28 | 1.000 | 0.518 | 646.8 MiB | 148.12 KiB | | sqlitevec | vec0 brute-force scan, no ANN index (exact), float32, cosine distance, default chunk_size=1024; score = 1 - cosine distance; searches run concurrently, one WAL reader connection per searcher (opened as searchers arrive, at most 32), writes run one at a time and may overlap searches | 4,338 | 1.91 | 2.13 | 2.36 | 514.3 | 1136.4 | 2.04 | 3.51 | 1.000 | 0.518 | - | 9.72 MiB | ## Details per target ### opensearch - Engine: OpenSearch 3.9.0 (k-NN plugin, faiss) (compose) - Index: faiss HNSW float32, no compression, m=16, ef_construction=128, cosinesimil; search k=top, ef_search=100; 1 shard, 0 replicas; graph built at any segment size (approximate_threshold=0); force-merged to one segment after the load - Search settings: approximate_threshold=0, ef_construction=128, ef_search=100, k=top, m=16 - Load: 524 rows in batches of 1000, 1.7 s of upserts (306 rows/s); count matched 0.2 s after the last upsert; index step 0.9 s (force-merged to one segment, graph ready after 0.9 s: every segment searched through its HNSW graph, covering 524 of 524 vectors: 1 segment(s), 0 merge(s) running, profiled probe search: ann_search_count 1 (segments tried through a graph), exact_search_count 0 (of those, scanned instead), approximate_threshold 0; force-merged to one segment on purpose, so one graph answers every search) - Search: 9817 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 0.46 ms, 8 searchers 0.50 ms - Exact mode: 21,950 searches, p50 2.69 ms, p95 3.01 ms, recall 1.000 - RAM: 2.57 GiB (docker stats); disk: 9.48 MiB added by this load (12.51 MiB (whole engine data folder), 3.03 MiB before) - Started by run-all for this measurement with every container created on CPUs 2-3,6-7 (started opensearch.compose.yaml with every container created on CPUs 2-3,6-7: gvb-opensearch cpuset 2-3,6-7 (read back from docker inspect)); stopped afterwards. - Engine files (SHA-256, first 12 hex digits): [opensearch.compose.yaml 028b44c40129]. This run created the engine from these files. - Index after the load: ready, 524 of 524 indexed (every segment searched through its HNSW graph, covering 524 of 524 vectors: 1 segment(s), 0 merge(s) running, profiled probe search: ann_search_count 1 (segments tried through a graph), exact_search_count 0 (of those, scanned instead), approximate_threshold 0; force-merged to one segment on purpose, so one graph answers every search). Durability: Every acknowledged bulk request is fsynced to the translog before the answer: index.translog.durability=request, the OpenSearch default, which neither this sink nor opensearch.compose.yaml overrides (OpenSearchReadinessTests reads it back from the live index). By that setting a process crash or power loss loses no acknowledged write; this is read from the setting, not shown by pulling power. One node and no replicas, so a lost disk loses the data.. - WARNING: processes outside the benchmark used 0.34 CPUs on average over the last 3 s and 0.34 over the last few seconds when searching began (limit 0.3): the box was busy as the untimed rehearsal started. Each timed pass still waits for a quiet box before its warm-up, and one that ran busy is flagged 'busy box' on its own. The load average 2.83 3.71 3.45 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: default@8 25,808 searches in 30.0 s, exact 10,955 searches in 30.0 s, default@1 14,724 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. default@8: warm-up 15.0 s and 21,935 searches (0 failed) at 8 searchers; QPS of the last windows 1,520, 1,530, 1,527 (windows of at least 2 s and 100 searches); trial of 4,562 searches in 3.0 s at 8 searchers: 1,519 QPS against the settled 1,527 QPS, 1% apart (limit 10%); the timed pass 1,516 QPS, 1% from the settled 1,527 QPS (limit 10%); exact: warm-up 15.0 s and 5,498 searches (0 failed) at 1 searcher; p50 of the last windows 2.680, 2.688, 2.691 ms (windows of at least 2 s and 100 searches); trial of 1,096 searches in 3.0 s at 1 searcher: p50 2.697 ms against the settled p50 2.688 ms, 0% apart (limit 10%); the timed pass p50 2.688 ms, 0% from the settled p50 2.688 ms (limit 10% or 0.1 ms; 0.001 ms); default@1: warm-up 15.0 s and 7,334 searches (0 failed) at 1 searcher; p50 of the last windows 1.991, 1.990, 1.993 ms (windows of at least 2 s and 100 searches); trial of 1,469 searches in 3.0 s at 1 searcher: p50 1.998 ms against the settled p50 1.991 ms, 0% apart (limit 10%); the timed pass p50 1.995 ms, 0% from the settled p50 1.991 ms (limit 10% or 0.1 ms; 0.004 ms). - Pass default@8 after rehearsal: 2026-10-06T14:29:44.932Z to 2026-10-06T14:30:04.935Z (20.0 s), 30,324 searches, 0 failed, p50 5.035 ms, mean 5.274 ms, p99 10.572 ms, 1516.0 QPS (1000/QPS 0.660 ms). - Pass exact after default@8: 2026-10-06T14:30:23.030Z to 2026-10-06T14:31:23.033Z (60.0 s), 21,950 searches, 0 failed, p50 2.688 ms, mean 2.732 ms, p99 3.608 ms, 365.8 QPS (1000/QPS 2.734 ms). - Pass default@1 after exact: 2026-10-06T14:31:41.069Z to 2026-10-06T14:32:01.069Z (20.0 s), 9,817 searches, 0 failed, p50 1.995 ms, mean 2.035 ms, p99 2.782 ms, 490.8 QPS (1000/QPS 2.037 ms). - Passes in the order run: default@8, exact, default@1; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 93,381 sent, 0 failed (warmupErrors). Index after the searches: ready, 524 of 524 indexed (every segment searched through its HNSW graph, covering 524 of 524 vectors: 1 segment(s), 0 merge(s) running, profiled probe search: ann_search_count 1 (segments tried through a graph), exact_search_count 0 (of those, scanned instead), approximate_threshold 0; force-merged to one segment on purpose, so one graph answers every search). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: container address on its Docker network, not the published port (no docker-proxy): container gvb-opensearch (b904b902c985) on network engines_default at container-ip:9200, not through docker-proxy localhost:9201. Open after the passes: localhost:1433 x2 (not this target), container-ip:9200 x1 (this target). - CPU pinning: cpuset from the container's creation when run-all started it on the engine CPUs, else docker update --cpuset-cpus (each container's change line says which), CPUs 2-3,6-7. Changed: container gvb-opensearch: already on CPUs 2-3,6-7 (its cpuset since creation), not changed. Read back: container gvb-opensearch: 55 thread(s) on CPUs 2-3,6-7 (every thread of the container). - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): default@8 3492/3492 MHz, average 3492.1/3492.3, performance, outside load 0.21 before the warm-up, 0.21 during, client CPU 0.499 ms per search; exact 3492/3492 MHz, average 3492.1/3492.5, performance, outside load 0.21 before the warm-up, 0.21 during, client CPU 0.466 ms per search; default@1 3492/3492 MHz, average 3491.2/3492.6, performance, outside load 0.21 before the warm-up, 0.22 during, client CPU 0.458 ms per search. ### redis - Engine: Redis 8.10.2 (query engine, HASH + vector index) (compose) - Index: HNSW TYPE FLOAT32 M=16 EF_CONSTRUCTION=128, EF_RUNTIME=100 per query, cosine; exact mode = FLAT index built on first exact query - Search settings: EF_CONSTRUCTION=128, EF_RUNTIME=100, M=16 - Load: 524 rows in batches of 1000, 0.0 s of upserts (11,736 rows/s); count matched 1.0 s after the last upsert; index step 0.0 s (HNSW index finished after 0.0 s of waiting: FT.INFO indexing 0, percent_indexed 1, hash_indexing_failures 0, flat_buffer_size 0, num_docs 524, hashes counted with SCAN under gvb_gvbbench_eshoponweb: 524) - Search: 50732 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 0.56 ms, 8 searchers 0.37 ms - Exact mode: 158,989 searches, p50 0.38 ms, p95 0.47 ms, recall 1.000 - RAM: 253.7 MiB (docker stats); disk: 0 B added by this load (94.22 MiB (whole engine data folder), 94.22 MiB before) - Started by run-all for this measurement with every container created on CPUs 2-3,6-7 (started redis.compose.yaml with every container created on CPUs 2-3,6-7: gvb-redis cpuset 2-3,6-7 (read back from docker inspect)); stopped afterwards. - Engine files (SHA-256, first 12 hex digits): [redis.compose.yaml 7448634e0d92]. This run created the engine from these files. - Index after the load: ready, 524 of 524 indexed (FT.INFO indexing 0, percent_indexed 1, hash_indexing_failures 0, flat_buffer_size 0, num_docs 524, hashes counted with SCAN under gvb_gvbbench_eshoponweb: 524). Durability: save "300 1" and appendonly no (redis.compose.yaml): an RDB snapshot is written every 5 minutes if at least one key changed, and on a clean stop; there is no append-only log, so a crash loses every write since the last snapshot (up to 5 minutes plus the time a snapshot takes; measured with docker kill, which is SIGKILL, right after a 2,000-vector load: none of the 2,000 hashes were there after the restart, and the restart loaded an older snapshot that still held keys of a collection that had been dropped since). - Outside load when searching began: 0.22 CPUs on average over the last 1 s and 0.22 over the last few seconds (limit 0.3), a quiet box. The load average 2.29 3.28 3.38 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: default@8 151,766 searches in 30.0 s, default@1 75,574 searches in 30.0 s, exact 76,524 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. default@8: warm-up 15.0 s and 75,565 searches (0 failed) at 8 searchers; QPS of the last windows 5,018, 5,082, 5,083 (windows of at least 2 s and 100 searches); trial of 14,920 searches in 3.0 s at 8 searchers: 4,971 QPS against the settled 5,082 QPS, 2% apart (limit 10%); the timed pass 5,067 QPS, 0% from the settled 5,082 QPS (limit 10%); default@1: warm-up 15.0 s and 37,563 searches (0 failed) at 1 searcher; p50 of the last windows 0.403, 0.404, 0.385 ms (windows of at least 2 s and 100 searches); trial of 7,406 searches in 3.0 s at 1 searcher: p50 0.410 ms against the settled p50 0.403 ms, 2% apart (limit 10%); the timed pass p50 0.393 ms, 3% from the settled p50 0.403 ms (limit 10% or 0.1 ms; 0.010 ms); exact: warm-up 15.0 s and 39,558 searches (0 failed) at 1 searcher; p50 of the last windows 0.372, 0.391, 0.381 ms (windows of at least 2 s and 100 searches); trial of 7,897 searches in 3.0 s at 1 searcher: p50 0.385 ms against the settled p50 0.381 ms, 1% apart (limit 10%); the timed pass p50 0.382 ms, 0% from the settled p50 0.381 ms (limit 10% or 0.1 ms; 0.001 ms). - Pass default@8 after rehearsal: 2026-10-06T14:34:13.702Z to 2026-10-06T14:34:33.703Z (20.0 s), 101,354 searches, 0 failed, p50 1.573 ms, mean 1.577 ms, p99 2.263 ms, 5067.3 QPS (1000/QPS 0.197 ms). - Pass default@1 after default@8: 2026-10-06T14:34:51.741Z to 2026-10-06T14:35:11.741Z (20.0 s), 50,732 searches, 0 failed, p50 0.393 ms, mean 0.393 ms, p99 0.695 ms, 2536.6 QPS (1000/QPS 0.394 ms). - Pass exact after default@1: 2026-10-06T14:35:29.799Z to 2026-10-06T14:36:29.799Z (60.0 s), 158,989 searches, 0 failed, p50 0.382 ms, mean 0.376 ms, p99 0.670 ms, 2649.8 QPS (1000/QPS 0.377 ms). - Passes in the order run: default@8, default@1, exact; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 486,773 sent, 0 failed (warmupErrors). Index after the searches: ready, 524 of 524 indexed (FT.INFO indexing 0, percent_indexed 1, hash_indexing_failures 0, flat_buffer_size 0, num_docs 524, hashes counted with SCAN under gvb_gvbbench_eshoponweb: 524). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: container address on its Docker network, not the published port (no docker-proxy): container gvb-redis (bf94aa81ff23) on network gvb-redis_default at container-ip:6379, not through docker-proxy localhost:6379. Open after the passes: container-ip:6379 x1 (this target). - CPU pinning: cpuset from the container's creation when run-all started it on the engine CPUs, else docker update --cpuset-cpus (each container's change line says which), CPUs 2-3,6-7. Changed: container gvb-redis: already on CPUs 2-3,6-7 (its cpuset since creation), not changed. Read back: container gvb-redis: 16 thread(s) on CPUs 2-3,6-7 (every thread of the container). - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): default@8 3492/3492 MHz, average 3492.1/3492.0, performance, outside load 0.20 before the warm-up, 0.20 during, client CPU 0.366 ms per search; default@1 3492/3492 MHz, average 3492.0/3492.0, performance, outside load 0.20 before the warm-up, 0.20 during, client CPU 0.561 ms per search; exact 3492/3492 MHz, average 3492.0/3491.7, performance, outside load 0.20 before the warm-up, 0.20 during, client CPU 0.572 ms per search. ### qdrant-hnsw - Engine: Qdrant 1.17.0 in container gvb-qdrant (image qdrant/qdrant:v1.17.0@sha256:f1c7272cdac52b38c1a0e89313922d940ba50afd90d593a1605dbbc214e66ffb; cpuset 2-3,6-7 from its creation; 3 search thread(s)) (compose) - Index: HNSW m=16 ef_construct=100, hnsw_ef=server default, cosine; indexing_threshold_kb 1 and full_scan_threshold_kb 10 (server defaults are 10,000 each) so a small collection builds and walks its graph - Search settings: ef_construct=100, hnsw_ef=server, m=16 - Load: 524 rows in batches of 1000, 0.1 s of upserts (4,640 rows/s); count matched 0.0 s after the last upsert; index step 2.0 s (indexing_threshold_kb 1, full_scan_threshold_kb 10; status Green, 524 of 524 vectors in HNSW segments (2 segments), waited 2.0 s) - Search: 21533 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 0.74 ms, 8 searchers 0.52 ms - Exact mode: 66,134 searches, p50 0.89 ms, p95 1.02 ms, recall 1.000 - RAM: 38.41 MiB (docker stats); disk: 164.1 MiB (collection folder) - Started by run-all for this measurement with every container created on CPUs 2-3,6-7 (started qdrant.compose.yaml with every container created on CPUs 2-3,6-7: gvb-qdrant cpuset 2-3,6-7 (read back from docker inspect)); stopped afterwards. - Engine files (SHA-256, first 12 hex digits): [qdrant.compose.yaml eb57e55e93c1]. This run created the engine from these files. - Index after the load: ready, 524 of 524 indexed (status Green, optimizer ok, indexed_vectors_count 524 of 524 points, 2 segments, indexing_threshold_kb 1, full_scan_threshold_kb 10; every one of 1 non-empty segments has a complete HNSW graph above its full-scan threshold; no search has run yet, so the walk is proven by the segment settings only; searches counted since the collection was created: unfiltered_hnsw 0, unfiltered_plain 0, unfiltered_exact 0). Durability: What was measured (strace -f on the Qdrant server process, 30 single-point REST upserts per setting, every sync call timed against the request that caused it): with wait=true each upsert had exactly one msync(MS_SYNC) of the write-ahead-log segment inside the request, before the reply (30 of 30); with wait=false the 30 upserts were acknowledged within 0.26 s with no sync call, and the first WAL msync (one call covering all 30 records) came 2.7 s after the last reply, followed by the segment-file flushes. So the log is flushed to disk under both settings, but with wait=false the flush comes after the acknowledgement: an operating-system crash or power cut in that gap loses acknowledged writes, while a crash of the Qdrant process alone should not, because the bytes are already in the kernel's page cache (inferred, not tested). Every upsert here is sent with wait=true in batches of 256 over gRPC; the strace used one point per request, so one flush per batch is inferred, not measured. Segment files are flushed every 5 s; log segments 32 MB, 0 created ahead. /qdrant/config/config.yaml in container gvb-qdrant (the image's own file; its storage keys are listed) sets: storage.collection.quantization = null, storage.collection.replication_factor = 1, storage.collection.vectors.on_disk = null, storage.collection.write_consistency_factor = 1, storage.hnsw_index.ef_construct = 100, storage.hnsw_index.full_scan_threshold_kb = 10000, storage.hnsw_index.m = 16, storage.hnsw_index.max_indexing_threads = 0, storage.hnsw_index.on_disk = false, storage.hnsw_index.payload_m = null, storage.max_collections = null, storage.node_type = Normal, storage.on_disk_payload = true, storage.optimizers.default_segment_number = 0, storage.optimizers.deleted_threshold = 0.2, storage.optimizers.flush_interval_sec = 5, storage.optimizers.indexing_threshold_kb = 10000, storage.optimizers.max_optimization_threads = null, storage.optimizers.max_segment_size_kb = null, storage.optimizers.vacuum_min_vector_number = 1000, storage.performance.max_search_threads = 0, storage.performance.optimizer_cpu_budget = 0, storage.performance.update_rate_limit = null, storage.shard_transfer_method = null, storage.snapshots_config.snapshots_storage = local, storage.snapshots_path = ./snapshots, storage.storage_path = ./storage, storage.temp_path = null, storage.update_concurrency = null, storage.wal.wal_capacity_mb = 32, storage.wal.wal_segments_ahead = 0; no QDRANT__ environment overrides in the server process. Not tested by cutting power; whether the disk's own write cache reaches the media was not checked.. - Outside load when searching began: 0.23 CPUs on average over the last 2 s and 0.23 over the last few seconds (limit 0.3), a quiet box. The load average 2.17 2.91 3.22 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: exact 33,224 searches in 30.0 s, default@1 30,716 searches in 30.0 s, default@8 83,434 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. exact: warm-up 15.0 s and 16,471 searches (0 failed) at 1 searcher; p50 of the last windows 0.898, 0.893, 0.896 ms (windows of at least 2 s and 100 searches); trial of 3,327 searches in 3.0 s at 1 searcher: p50 0.890 ms against the settled p50 0.896 ms, 1% apart (limit 10%); the timed pass p50 0.893 ms, 0% from the settled p50 0.896 ms (limit 10% or 0.1 ms; 0.003 ms); default@1: warm-up 15.0 s and 16,161 searches (0 failed) at 1 searcher; p50 of the last windows 0.912, 0.914, 0.916 ms (windows of at least 2 s and 100 searches); trial of 3,218 searches in 3.0 s at 1 searcher: p50 0.918 ms against the settled p50 0.914 ms, 0% apart (limit 10%); the timed pass p50 0.916 ms, 0% from the settled p50 0.914 ms (limit 10% or 0.1 ms; 0.002 ms); default@8: warm-up 15.0 s and 51,844 searches (0 failed) at 8 searchers; QPS of the last windows 3,442, 3,476, 3,439 (windows of at least 2 s and 100 searches); trial of 10,452 searches in 3.0 s at 8 searchers: 3,482 QPS against the settled 3,442 QPS, 1% apart (limit 10%); the timed pass 3,440 QPS, 0% from the settled 3,442 QPS (limit 10%). - Pass exact after rehearsal: 2026-10-06T14:39:28.993Z to 2026-10-06T14:40:28.994Z (60.0 s), 66,134 searches, 0 failed, p50 0.893 ms, mean 0.905 ms, p99 1.389 ms, 1102.2 QPS (1000/QPS 0.907 ms). - Pass default@1 after exact: 2026-10-06T14:40:47.053Z to 2026-10-06T14:41:07.054Z (20.0 s), 21,533 searches, 0 failed, p50 0.916 ms, mean 0.927 ms, p99 1.403 ms, 1076.6 QPS (1000/QPS 0.929 ms). - Pass default@8 after default@1: 2026-10-06T14:41:25.077Z to 2026-10-06T14:41:45.078Z (20.0 s), 68,807 searches, 0 failed, p50 2.238 ms, mean 2.323 ms, p99 4.012 ms, 3440.1 QPS (1000/QPS 0.291 ms). - Passes in the order run: exact, default@1, default@8; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 248,847 sent, 0 failed (warmupErrors). Index after the searches: ready, 524 of 524 indexed (status Green, optimizer ok, indexed_vectors_count 524 of 524 points, 2 segments, indexing_threshold_kb 1, full_scan_threshold_kb 10; every one of 1 non-empty segments has a complete HNSW graph above its full-scan threshold; 286,165 segment searches walked the graph and none scanned; searches counted since the collection was created: unfiltered_hnsw 286,165, unfiltered_plain 0, unfiltered_exact 119,156). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: container address on its Docker network, not the published port (no docker-proxy): container gvb-qdrant (bda3d52c33a2) on network gvb-qdrant_default at container-ip:6334, not through docker-proxy localhost:16334; container gvb-qdrant (bda3d52c33a2) on network gvb-qdrant_default at container-ip:6333, not through docker-proxy localhost:16333. Open after the passes: container-ip:6334 x1 (this target). - CPU pinning: cpuset from the container's creation when run-all started it on the engine CPUs, else docker update --cpuset-cpus (each container's change line says which), CPUs 2-3,6-7. Changed: container gvb-qdrant: already on CPUs 2-3,6-7 (its cpuset since creation), not changed. Read back: container gvb-qdrant: 24 thread(s) on CPUs 2-3,6-7 (every thread of the container). - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): exact 3492/3492 MHz, average 3492.0/3492.0, performance, outside load 0.29 before the warm-up after waiting 60 s, 0.19 during, client CPU 0.733 ms per search; default@1 3492/3492 MHz, average 3491.5/3492.0, performance, outside load 0.19 before the warm-up, 0.16 during, client CPU 0.737 ms per search; default@8 3492/3492 MHz, average 3492.0/3492.0, performance, outside load 0.17 before the warm-up, 0.24 during, client CPU 0.517 ms per search. ### weaviate - Engine: Weaviate 1.39.8 (single node, REST + GraphQL) (compose) - Index: HNSW maxConnections(M)=16 efConstruction=128, ef=-1 (dynamic: limit x 8 clamped 100..500), cosine, no quantization; approximate only (no exact mode) - Search settings: ef=-1, efConstruction=128 - Load: 524 rows in batches of 1000, 0.5 s of upserts (1,001 rows/s); count matched 0.0 s after the last upsert; index step 0.0 s (HNSW is updated inside every batch, nothing to build; confirmed in 0.0 s: shard iKiSj18BxrvU vectorIndexingStatus READY, vectorQueueLength 0, node status objectCount 0 (refreshed only when the memtable is flushed, so it lags a minute and does not decide readiness), schema shard iKiSj18BxrvU status READY, Aggregate count 524, Aggregate nearVector with objectLimit 524 reached 524, Weaviate reports no count of vectors in the HNSW index, so indexed vectors are not reported) - Search: 3335 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 0.61 ms, 8 searchers 0.62 ms - RAM: 71.18 MiB (docker stats); disk: 2.96 MiB added by this load (5.02 MiB (whole engine data folder), 2.06 MiB before) - Started by run-all for this measurement with every container created on CPUs 2-3,6-7 (started weaviate.compose.yaml with every container created on CPUs 2-3,6-7: gvb-weaviate cpuset 2-3,6-7 (read back from docker inspect)); stopped afterwards. - Engine files (SHA-256, first 12 hex digits): [weaviate.compose.yaml d96b7399662a]. This run created the engine from these files. - Index after the load: ready, ? of 524 indexed (shard iKiSj18BxrvU vectorIndexingStatus READY, vectorQueueLength 0, node status objectCount 0 (refreshed only when the memtable is flushed, so it lags a minute and does not decide readiness), schema shard iKiSj18BxrvU status READY, Aggregate count 524, Aggregate nearVector with objectLimit 524 reached 524, Weaviate reports no count of vectors in the HNSW index, so indexed vectors are not reported). Durability: Weaviate 1.39.8 defaults, weaviate.compose.yaml sets no persistence variable: every object is appended to the LSM write-ahead log with a plain write and no fsync, and the log is fsynced only when its memtable is flushed, 60 seconds after the last write (PERSISTENCE_MEMTABLES_FLUSH_IDLE_AFTER_SECONDS default; measured with strace: 2,080 writes into the objects log during a 2,000-object load, the first fsync 60 s after the last write), so a power cut can lose the last minute of writes; the HNSW commit log is buffered inside the process (77 writes for 2,000 vectors), so a killed process loses the newest vectors from the vector index while their objects survive (measured with docker kill, which is SIGKILL, one second after the load and a restart, two runs each: 0 to 1 of 50, 264 to 267 of 300 and 1,992 to 1,997 of 2,000 stored objects were still reachable through a vector search). - Outside load when searching began: 0.21 CPUs on average over the last 1 s and 0.21 over the last few seconds (limit 0.3), a quiet box. The load average 2.87 2.70 3.01 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: default@8 9,288 searches in 30.0 s, default@1 5,040 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. default@8: warm-up 15.0 s and 4,604 searches (0 failed) at 8 searchers; QPS of the last windows 306, 306, 306 (windows of at least 2 s and 100 searches); trial of 920 searches in 3.0 s at 8 searchers: 305 QPS against the settled 306 QPS, 0% apart (limit 10%); the timed pass 305 QPS, 0% from the settled 306 QPS (limit 10%); default@1: warm-up 15.0 s and 2,492 searches (0 failed) at 1 searcher; p50 of the last windows 5.332, 5.368, 5.354 ms (windows of at least 2 s and 100 searches); trial of 498 searches in 3.0 s at 1 searcher: p50 5.372 ms against the settled p50 5.354 ms, 0% apart (limit 10%); the timed pass p50 5.360 ms, 0% from the settled p50 5.354 ms (limit 10% or 0.1 ms; 0.005 ms). - Pass default@8 after rehearsal: 2026-10-06T14:43:27.188Z to 2026-10-06T14:43:47.203Z (20.0 s), 6,114 searches, 0 failed, p50 24.983 ms, mean 26.178 ms, p99 53.363 ms, 305.5 QPS (1000/QPS 3.274 ms). - Pass default@1 after default@8: 2026-10-06T14:44:05.208Z to 2026-10-06T14:44:25.209Z (20.0 s), 3,335 searches, 0 failed, p50 5.360 ms, mean 5.994 ms, p99 11.555 ms, 166.7 QPS (1000/QPS 5.997 ms). - Passes in the order run: default@8, default@1; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 22,842 sent, 0 failed (warmupErrors). Index after the searches: ready, ? of 524 indexed (shard iKiSj18BxrvU vectorIndexingStatus READY, vectorQueueLength 0, node status objectCount 524 (refreshed only when the memtable is flushed, so it lags a minute and does not decide readiness), schema shard iKiSj18BxrvU status READY, Aggregate count 524, Aggregate nearVector with objectLimit 524 reached 524, Weaviate reports no count of vectors in the HNSW index, so indexed vectors are not reported). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: container address on its Docker network, not the published port (no docker-proxy): container gvb-weaviate (9d1c5a3b2b01) on network gvb-weaviate_default at container-ip:8080, not through docker-proxy localhost:8085. Open after the passes: container-ip:8080 x1 (this target). - CPU pinning: cpuset from the container's creation when run-all started it on the engine CPUs, else docker update --cpuset-cpus (each container's change line says which), CPUs 2-3,6-7. Changed: container gvb-weaviate: already on CPUs 2-3,6-7 (its cpuset since creation), not changed. Read back: container gvb-weaviate: 8 thread(s) on CPUs 2-3,6-7 (every thread of the container). - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): default@8 3492/3492 MHz, average 3492.0/3492.4, performance, outside load 0.21 before the warm-up, 0.24 during, client CPU 0.621 ms per search; default@1 3492/3492 MHz, average 3492.2/3492.6, performance, outside load 0.22 before the warm-up, 0.20 during, client CPU 0.608 ms per search. ### qdrant - Engine: Qdrant 1.17.0 in container gvb-qdrant (image qdrant/qdrant:v1.17.0@sha256:f1c7272cdac52b38c1a0e89313922d940ba50afd90d593a1605dbbc214e66ffb; cpuset 2-3,6-7 from its creation; 3 search thread(s)) (compose) - Index: exact scan: the builder's sink sends exact=true on every search, so no HNSW graph is used whether or not Qdrant has built one (see the index state) - Search settings: exact=true - Load: 524 rows in batches of 1000, 0.1 s of upserts (6,919 rows/s); count matched 0.0 s after the last upsert; index step 1.0 s (status Green, 0 of 524 vectors in HNSW segments (2 segments), waited 1.0 s) - Search: 22277 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 0.73 ms, 8 searchers 0.53 ms - RAM: 35.56 MiB (docker stats); disk: 196.08 MiB (collection folder) - Started by run-all for this measurement with every container created on CPUs 2-3,6-7 (started qdrant.compose.yaml with every container created on CPUs 2-3,6-7: gvb-qdrant cpuset 2-3,6-7 (read back from docker inspect)); stopped afterwards. - Engine files (SHA-256, first 12 hex digits): [qdrant.compose.yaml eb57e55e93c1]. This run created the engine from these files. - Index after the load: ready, 0 of 524 indexed (status Green, optimizer ok, indexed_vectors_count 0 of 524 points, 2 segments, indexing_threshold_kb 10,000, full_scan_threshold_kb 10,000; no index used, exact scan by design (the builder's sink sends exact=true on every search); no graph was built; searches counted since the collection was created: unfiltered_hnsw 0, unfiltered_plain 0, unfiltered_exact 0). Durability: What was measured (strace -f on the Qdrant server process, 30 single-point REST upserts per setting, every sync call timed against the request that caused it): with wait=true each upsert had exactly one msync(MS_SYNC) of the write-ahead-log segment inside the request, before the reply (30 of 30); with wait=false the 30 upserts were acknowledged within 0.26 s with no sync call, and the first WAL msync (one call covering all 30 records) came 2.7 s after the last reply, followed by the segment-file flushes. So the log is flushed to disk under both settings, but with wait=false the flush comes after the acknowledgement: an operating-system crash or power cut in that gap loses acknowledged writes, while a crash of the Qdrant process alone should not, because the bytes are already in the kernel's page cache (inferred, not tested). Every upsert here is sent with wait=true in batches of 256 over gRPC; the strace used one point per request, so one flush per batch is inferred, not measured. Segment files are flushed every 5 s; log segments 32 MB, 0 created ahead. /qdrant/config/config.yaml in container gvb-qdrant (the image's own file; its storage keys are listed) sets: storage.collection.quantization = null, storage.collection.replication_factor = 1, storage.collection.vectors.on_disk = null, storage.collection.write_consistency_factor = 1, storage.hnsw_index.ef_construct = 100, storage.hnsw_index.full_scan_threshold_kb = 10000, storage.hnsw_index.m = 16, storage.hnsw_index.max_indexing_threads = 0, storage.hnsw_index.on_disk = false, storage.hnsw_index.payload_m = null, storage.max_collections = null, storage.node_type = Normal, storage.on_disk_payload = true, storage.optimizers.default_segment_number = 0, storage.optimizers.deleted_threshold = 0.2, storage.optimizers.flush_interval_sec = 5, storage.optimizers.indexing_threshold_kb = 10000, storage.optimizers.max_optimization_threads = null, storage.optimizers.max_segment_size_kb = null, storage.optimizers.vacuum_min_vector_number = 1000, storage.performance.max_search_threads = 0, storage.performance.optimizer_cpu_budget = 0, storage.performance.update_rate_limit = null, storage.shard_transfer_method = null, storage.snapshots_config.snapshots_storage = local, storage.snapshots_path = ./snapshots, storage.storage_path = ./storage, storage.temp_path = null, storage.update_concurrency = null, storage.wal.wal_capacity_mb = 32, storage.wal.wal_segments_ahead = 0; no QDRANT__ environment overrides in the server process. Not tested by cutting power; whether the disk's own write cache reaches the media was not checked.. - Outside load when searching began: 0.13 CPUs on average over the last 1 s and 0.13 over the last few seconds (limit 0.3), a quiet box. The load average 2.11 2.67 2.97 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: default@8 98,475 searches in 30.0 s, default@1 33,437 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. default@8: warm-up 15.0 s and 49,176 searches (0 failed) at 8 searchers; QPS of the last windows 3,273, 3,298, 3,264 (windows of at least 2 s and 100 searches); trial of 9,904 searches in 3.0 s at 8 searchers: 3,298 QPS against the settled 3,273 QPS, 1% apart (limit 10%); the timed pass 3,280 QPS, 0% from the settled 3,273 QPS (limit 10%); default@1: warm-up 15.0 s and 16,660 searches (0 failed) at 1 searcher; p50 of the last windows 0.887, 0.884, 0.888 ms (windows of at least 2 s and 100 searches); trial of 3,340 searches in 3.0 s at 1 searcher: p50 0.883 ms against the settled p50 0.887 ms, 0% apart (limit 10%); the timed pass p50 0.883 ms, 0% from the settled p50 0.887 ms (limit 10% or 0.1 ms; 0.003 ms). - Pass default@8 after rehearsal: 2026-10-06T14:45:52.577Z to 2026-10-06T14:46:12.579Z (20.0 s), 65,609 searches, 0 failed, p50 2.352 ms, mean 2.437 ms, p99 4.241 ms, 3280.1 QPS (1000/QPS 0.305 ms). - Pass default@1 after default@8: 2026-10-06T14:46:30.601Z to 2026-10-06T14:46:50.602Z (20.0 s), 22,277 searches, 0 failed, p50 0.883 ms, mean 0.896 ms, p99 1.385 ms, 1113.8 QPS (1000/QPS 0.898 ms). - Passes in the order run: default@8, default@1; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 210,992 sent, 0 failed (warmupErrors). Index after the searches: ready, 0 of 524 indexed (status Green, optimizer ok, indexed_vectors_count 0 of 524 points, 2 segments, indexing_threshold_kb 10,000, full_scan_threshold_kb 10,000; no index used, exact scan by design (the builder's sink sends exact=true on every search); no graph was built; searches counted since the collection was created: unfiltered_hnsw 0, unfiltered_plain 597,756, unfiltered_exact 0). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: container address on its Docker network, not the published port (no docker-proxy): container gvb-qdrant (4336f91608a1) on network gvb-qdrant_default at container-ip:6334, not through docker-proxy localhost:16334; container gvb-qdrant (4336f91608a1) on network gvb-qdrant_default at container-ip:6333, not through docker-proxy localhost:16333. Open after the passes: container-ip:6334 x1 (this target). - CPU pinning: cpuset from the container's creation when run-all started it on the engine CPUs, else docker update --cpuset-cpus (each container's change line says which), CPUs 2-3,6-7. Changed: container gvb-qdrant: already on CPUs 2-3,6-7 (its cpuset since creation), not changed. Read back: container gvb-qdrant: 24 thread(s) on CPUs 2-3,6-7 (every thread of the container). - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): default@8 3492/3492 MHz, average 3492.0/3492.0, performance, outside load 0.21 before the warm-up, 0.23 during, client CPU 0.531 ms per search; default@1 3492/3492 MHz, average 3492.0/3492.1, performance, outside load 0.21 before the warm-up, 0.21 during, client CPU 0.729 ms per search. ### mariadb - Engine: MariaDB 11.8.9 (InnoDB + VECTOR INDEX) (compose) - Index: VECTOR INDEX (HNSW variant) DISTANCE=cosine, M=16 (no ef_construction setting exists), mhnsw_ef_search=100 per statement (the ef 100 most engines here use, so the search effort matches; MariaDB's own default is 20; recall@10 at ef 100 falls as the set grows (random 1024-dimension vectors, measured 2026-10-04: 0.99 at 524, 0.89 to 0.92 at 2,000; an earlier run gave about 0.09 at 100,000)), mhnsw_max_cache_size 4G; exact mode = IGNORE INDEX full scan - Search settings: DISTANCE=cosine, M=16, mhnsw_ef_search=100 - Load: 524 rows in batches of 1000, 0.5 s of upserts (1,035 rows/s); count matched 0.0 s after the last upsert; index step 0.5 s (VECTOR INDEX is maintained inside every INSERT, nothing to build; confirmed after 2 poll(s) in 0.5 s (poll 1 of 2 said: not ready: a search at mhnsw_ef_search 100 returned only 391 of 524 requested rows through the index (EXPLAIN key 'vec_idx') (information_schema lists vec_idx as INDEX_TYPE VECTOR, EXPLAIN of the default search uses key vec_idx, the server applied mhnsw_ef_search 100 (asked for 100, read back from the server), a search at that effort returned 391 of 524 requested rows through the index; MariaDB has no per-index row count (the graph is a hidden InnoDB table), so indexed vectors are not reported)): information_schema lists vec_idx as INDEX_TYPE VECTOR, EXPLAIN of the default search uses key vec_idx, the server applied mhnsw_ef_search 100 (asked for 100, read back from the server), a search at that effort returned 519 of 524 requested rows through the index; MariaDB has no per-index row count (the graph is a hidden InnoDB table), so indexed vectors are not reported) - Search: 31560 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 0.48 ms, 8 searchers 0.38 ms - Exact mode: 36,041 searches, p50 1.63 ms, p95 1.80 ms, recall 1.000 - RAM: 170.4 MiB (docker stats); disk: 22.01 MiB added by this load (182.79 MiB (whole engine data folder), 160.78 MiB before) - Started by run-all for this measurement with every container created on CPUs 2-3,6-7 (started mariadb-bench.compose.yaml with every container created on CPUs 2-3,6-7: gvbbench-mariadb cpuset 2-3,6-7 (read back from docker inspect)); stopped afterwards. - Engine files (SHA-256, first 12 hex digits): [mariadb-bench.compose.yaml e9a2147e6f91]. This run created the engine from these files. - Index after the load: ready, ? of 524 indexed (information_schema lists vec_idx as INDEX_TYPE VECTOR, EXPLAIN of the default search uses key vec_idx, the server applied mhnsw_ef_search 100 (asked for 100, read back from the server), a search at that effort returned 519 of 524 requested rows through the index; MariaDB has no per-index row count (the graph is a hidden InnoDB table), so indexed vectors are not reported). Durability: innodb_flush_log_at_trx_commit=2 (set in mariadb-bench.compose.yaml for the benchmark's own container gvbbench-mariadb, and in mariadb.compose.yaml for the daily gvb-mariadb, which has the same server settings): the InnoDB redo log is written to the operating system at every commit but fsynced about once a second, so a crash of the mariadbd process loses nothing, while an operating-system crash or power cut can lose the last second of commits; innodb_doublewrite is on (default), the binary log is off, and the vector graph is an InnoDB table under the same log (settings read from the running server with SHOW VARIABLES; the crash behaviour is InnoDB's documented behaviour for this setting and was not tested on either container). - Outside load when searching began: 0.18 CPUs on average over the last 1 s and 0.18 over the last few seconds (limit 0.3), a quiet box. The load average 3.70 3.55 3.29 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: default@1 46,559 searches in 30.0 s, exact 17,925 searches in 30.0 s, default@8 179,461 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. default@1: warm-up 15.0 s and 23,541 searches (0 failed) at 1 searcher; p50 of the last windows 0.612, 0.625, 0.639 ms (windows of at least 2 s and 100 searches); trial of 4,755 searches in 3.0 s at 1 searcher: p50 0.619 ms against the settled p50 0.625 ms, 1% apart (limit 10%); the timed pass p50 0.623 ms, 0% from the settled p50 0.625 ms (limit 10% or 0.1 ms; 0.002 ms); exact: warm-up 15.0 s and 9,016 searches (0 failed) at 1 searcher; p50 of the last windows 1.639, 1.628, 1.635 ms (windows of at least 2 s and 100 searches); trial of 1,780 searches in 3.0 s at 1 searcher: p50 1.632 ms against the settled p50 1.635 ms, 0% apart (limit 10%); the timed pass p50 1.635 ms, 0% from the settled p50 1.635 ms (limit 10% or 0.1 ms; 0.000 ms); default@8: warm-up 15.0 s and 88,497 searches (0 failed) at 8 searchers; QPS of the last windows 5,800, 5,944, 5,946 (windows of at least 2 s and 100 searches); trial of 17,908 searches in 3.0 s at 8 searchers: 5,967 QPS against the settled 5,944 QPS, 0% apart (limit 10%); the timed pass 5,920 QPS, 0% from the settled 5,944 QPS (limit 10%). - Pass default@1 after rehearsal: 2026-10-06T14:48:47.783Z to 2026-10-06T14:49:07.784Z (20.0 s), 31,560 searches, 0 failed, p50 0.623 ms, mean 0.632 ms, p99 0.853 ms, 1578.0 QPS (1000/QPS 0.634 ms). - Pass exact after default@1: 2026-10-06T14:49:25.815Z to 2026-10-06T14:50:25.815Z (60.0 s), 36,041 searches, 0 failed, p50 1.635 ms, mean 1.663 ms, p99 2.191 ms, 600.7 QPS (1000/QPS 1.665 ms). - Pass default@8 after exact: 2026-10-06T14:50:43.847Z to 2026-10-06T14:51:03.848Z (20.0 s), 118,400 searches, 0 failed, p50 1.303 ms, mean 1.350 ms, p99 2.478 ms, 5919.7 QPS (1000/QPS 0.169 ms). - Passes in the order run: default@1, exact, default@8; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 389,442 sent, 0 failed (warmupErrors). Index after the searches: ready, ? of 524 indexed (information_schema lists vec_idx as INDEX_TYPE VECTOR, EXPLAIN of the default search uses key vec_idx, the server applied mhnsw_ef_search 100 (asked for 100, read back from the server), a search at that effort returned 519 of 524 requested rows through the index; MariaDB has no per-index row count (the graph is a hidden InnoDB table), so indexed vectors are not reported). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: container address on its Docker network, not the published port (no docker-proxy): container gvbbench-mariadb (3289eed72802) on network gvbbench-mariadb_default at container-ip:3306, not through docker-proxy localhost:13306. Open after the passes: container-ip:3306 x8 (this target). - CPU pinning: cpuset from the container's creation when run-all started it on the engine CPUs, else docker update --cpuset-cpus (each container's change line says which), CPUs 2-3,6-7. Changed: container gvbbench-mariadb: already on CPUs 2-3,6-7 (its cpuset since creation), not changed. Read back: container gvbbench-mariadb: 12 thread(s) on CPUs 2-3,6-7 (every thread of the container). - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): default@1 3492/3492 MHz, average 3492.8/3491.8, performance, outside load 0.18 before the warm-up, 0.20 during, client CPU 0.482 ms per search; exact 3492/3492 MHz, average 3493.1/3491.3, performance, outside load 0.18 before the warm-up, 0.19 during, client CPU 0.504 ms per search; default@8 3492/3492 MHz, average 3492.0/3492.0, performance, outside load 0.19 before the warm-up, 0.16 during, client CPU 0.376 ms per search. ### vespa - Engine: Vespa 8.754.14 (tensor attribute + HNSW) (compose) - Index: HNSW float32 tensor, prenormalized-angular (cosine), max-links-per-node=16, neighbors-to-explore-at-insert=128; search targetHits=top, ef=100 via exploreAdditionalHits; exact mode = approximate:false; vectors held in memory - Search settings: ef=100, max-links-per-node=16, neighbors-to-explore-at-insert=128, targetHits=top - Load: 524 rows in batches of 1000, 2.2 s of upserts (238 rows/s); count matched 0.0 s after the last upsert; index step 0.1 s (nothing to build, Vespa inserts into the HNSW graph while writing; ready after 0.1 s: HNSW graph returns 524 of 524 vectors: totalCount 524, coverage full True, nearestNeighbor approximate:true with targetHits above the count: has_index True, algorithm "index top k", top_k_hits 524; the graph is updated while each write is applied, so there is no later build step) - Search: 9920 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 0.94 ms, 8 searchers 0.88 ms - Exact mode: 29,894 searches, p50 1.94 ms, p95 2.40 ms, recall 1.000 - RAM: 2.8 GiB (docker stats); disk: 0 B added by this load (2.12 GiB (whole engine data folder), 2.14 GiB before) - Started by run-all for this measurement with every container created on CPUs 2-3,6-7 (started vespa.compose.yaml with every container created on CPUs 2-3,6-7: gvb-vespa cpuset 2-3,6-7 (read back from docker inspect)); stopped afterwards. - Engine files (SHA-256, first 12 hex digits): [vespa.compose.yaml 3cfb702cc29f]. This run created the engine from these files. - Index after the load: ready, 524 of 524 indexed (HNSW graph returns 524 of 524 vectors: totalCount 524, coverage full True, nearestNeighbor approximate:true with targetHits above the count: has_index True, algorithm "index top k", top_k_hits 524; the graph is updated while each write is applied, so there is no later build step). Durability: Vespa's transaction log server fsyncs after each commit: searchlib.translogserver usefsync=true, and an operation is searchable when it is acknowledged: proton documentdb visibilitydelay=0. Neither is overridden by this sink or by the services.xml it deploys; VespaReadinessTests reads both from the live config server. After a crash the node replays the transaction log over its last flushed data and rebuilds the in-memory HNSW graph. By those settings an acknowledged write survives a crash; this is read from the settings, not shown by pulling power. One node and min-redundancy 1, so a lost disk loses the data.. - Outside load when searching began: 0.25 CPUs on average over the last 39 s and 0.26 over the last few seconds (limit 0.3), a quiet box. The load average 5.58 4.10 3.53 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: default@8 17,366 searches in 30.0 s, default@1 13,978 searches in 30.0 s, exact 14,765 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. default@8: warm-up 15.0 s and 20,872 searches (0 failed) at 8 searchers; QPS of the last windows 1,512, 1,538, 1,526 (windows of at least 2 s and 100 searches); trial of 4,436 searches in 3.0 s at 8 searchers: 1,477 QPS against the settled 1,526 QPS, 3% apart (limit 10%); the timed pass 1,506 QPS, 1% from the settled 1,526 QPS (limit 10%); default@1: warm-up 15.0 s and 7,444 searches (0 failed) at 1 searcher; p50 of the last windows 1.949, 1.967, 1.952 ms (windows of at least 2 s and 100 searches); trial of 1,493 searches in 3.0 s at 1 searcher: p50 1.947 ms against the settled p50 1.952 ms, 0% apart (limit 10%); the timed pass p50 1.947 ms, 0% from the settled p50 1.952 ms (limit 10% or 0.1 ms; 0.005 ms); exact: warm-up 15.0 s and 7,484 searches (0 failed) at 1 searcher; p50 of the last windows 1.939, 1.937, 1.938 ms (windows of at least 2 s and 100 searches); trial of 1,513 searches in 3.0 s at 1 searcher: p50 1.930 ms against the settled p50 1.938 ms, 0% apart (limit 10%); the timed pass p50 1.941 ms, 0% from the settled p50 1.938 ms (limit 10% or 0.1 ms; 0.003 ms). - Pass default@8 after rehearsal: 2026-10-06T14:53:49.148Z to 2026-10-06T14:54:09.150Z (20.0 s), 30,117 searches, 0 failed, p50 5.151 ms, mean 5.310 ms, p99 9.207 ms, 1505.7 QPS (1000/QPS 0.664 ms). - Pass default@1 after default@8: 2026-10-06T14:54:27.160Z to 2026-10-06T14:54:47.161Z (20.0 s), 9,920 searches, 0 failed, p50 1.947 ms, mean 2.013 ms, p99 2.880 ms, 496.0 QPS (1000/QPS 2.016 ms). - Pass exact after default@1: 2026-10-06T14:55:05.173Z to 2026-10-06T14:56:05.174Z (60.0 s), 29,894 searches, 0 failed, p50 1.941 ms, mean 2.004 ms, p99 2.760 ms, 498.2 QPS (1000/QPS 2.007 ms). - Passes in the order run: default@8, default@1, exact; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 89,351 sent, 0 failed (warmupErrors). Index after the searches: ready, 524 of 524 indexed (HNSW graph returns 524 of 524 vectors: totalCount 524, coverage full True, nearestNeighbor approximate:true with targetHits above the count: has_index True, algorithm "index top k", top_k_hits 524; the graph is updated while each write is applied, so there is no later build step). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: container address on its Docker network, not the published port (no docker-proxy): container gvb-vespa (da66def04364) on network engines_default at container-ip:8080, not through docker-proxy localhost:8090; container gvb-vespa (da66def04364) on network engines_default at container-ip:19071, not through docker-proxy localhost:19071. Open after the passes: container-ip:8080 x1 (this target). - CPU pinning: cpuset from the container's creation when run-all started it on the engine CPUs, else docker update --cpuset-cpus (each container's change line says which), CPUs 2-3,6-7. Changed: container gvb-vespa: already on CPUs 2-3,6-7 (its cpuset since creation), not changed. Read back: container gvb-vespa: 214 thread(s) on CPUs 2-3,6-7 (every thread of the container). - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): default@8 3492/3492 MHz, average 3492.0/3492.0, performance, outside load 0.19 before the warm-up, 0.28 during, client CPU 0.876 ms per search; default@1 3492/3492 MHz, average 3492.0/3492.2, performance, outside load 0.24 before the warm-up, 0.20 during, client CPU 0.938 ms per search; exact 3492/3492 MHz, average 3492.0/3492.3, performance, outside load 0.23 before the warm-up, 0.19 during, client CPU 0.938 ms per search. ### oracle - Engine: Oracle AI Database Free 23.26 (23ai line, VECTOR FLOAT32) (compose) - Index: HNSW in-memory neighbor graph NEIGHBORS=16 EFCONSTRUCTION=128, EFSEARCH=100 per query, cosine; exact mode = FETCH EXACT FIRST (full scan); Oracle Free caps itself at 2 CPUs (cpu_count 2 in V$PARAMETER, edition FREE in V$INSTANCE, 8 host CPUs in V$OSSTAT NUM_CPUS; the 2 CPU thread limit is Oracle's documented Free edition limit) - Search settings: EFCONSTRUCTION=128, EFSEARCH=100, NEIGHBORS=16 - Load: 524 rows in batches of 1000, 0.3 s of upserts (1,955 rows/s); count matched 0.1 s after the last upsert; index step 3.6 s (HNSW graph holds 524 of 524 rows, change log waiting: 0 inserts and 0 deletes, USER_INDEXES status VALID, plan of the default search: VECTOR INDEX HNSW SCAN > TABLE ACCESS BY INDEX ROWID, index used by 0 queries so far; Oracle Free caps itself at 2 CPUs (cpu_count 2 in V$PARAMETER, edition FREE in V$INSTANCE, 8 host CPUs in V$OSSTAT NUM_CPUS; the 2 CPU thread limit is Oracle's documented Free edition limit); graph rebuilt in 2.3 s) - Search: 20842 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 0.89 ms, 8 searchers 0.65 ms - Exact mode: 23,442 searches, p50 2.49 ms, p95 2.95 ms, recall 1.000 - RAM: 2.16 GiB (docker stats); disk: 0 B added by this load (6.9 GiB (whole engine data folder), 6.9 GiB before) - Started by run-all for this measurement with every container created on CPUs 2-3,6-7 (started oracle.compose.yaml with every container created on CPUs 2-3,6-7: gvb-oracle cpuset 2-3,6-7, gvb-oracle-seed cpuset 2-3,6-7 (read back from docker inspect)); stopped afterwards. - Engine files (SHA-256, first 12 hex digits): [oracle-init/01-vector-memory.sh 6452e4e1953e; oracle.compose.yaml 2a7fd367012f]. This run created the engine from these files. - Index after the load: ready, 524 of 524 indexed (HNSW graph holds 524 of 524 rows, change log waiting: 0 inserts and 0 deletes, USER_INDEXES status VALID, plan of the default search: VECTOR INDEX HNSW SCAN > TABLE ACCESS BY INDEX ROWID, index used by 0 queries so far; Oracle Free caps itself at 2 CPUs (cpu_count 2 in V$PARAMETER, edition FREE in V$INSTANCE, 8 host CPUs in V$OSSTAT NUM_CPUS; the 2 CPU thread limit is Oracle's documented Free edition limit)). Durability: By these settings a committed row should survive a crash or power loss. The sink uses a plain COMMIT and commit_logging, commit_wait and commit_write are unset in the database, so every commit waits for its redo to be written. Measured 2026-10-04: 50 separate client commits raised V$SYSSTAT 'redo synch writes' by 55 (the 50 commits plus the CREATE and DROP of the probe table), and the log writer and the datafile writer hold their files open with O_DSYNC (open flags 02110002, filesystemio_options none). The database runs NOARCHIVELOG (V$DATABASE.LOG_MODE), so redo serves crash recovery only and there is no point-in-time restore. The HNSW graph lives in the 768 MB vector memory pool (oracle-init/01-vector-memory.sh) and is not the durable copy; the table is. Not tested by cutting power; whether the disk's own write cache reaches the media was not checked. CPU: Oracle Free caps itself at 2 CPUs (measured 2026-10-04: cpu_count 2 in V$PARAMETER, edition FREE in V$INSTANCE, 8 host CPUs in V$OSSTAT NUM_CPUS; the 2 CPU thread limit is Oracle's documented Free edition limit; this sink reads the live value when the index state is read).. - Outside load when searching began: 0.21 CPUs on average over the last 6 s and 0.21 over the last few seconds (limit 0.3), a quiet box. The load average 2.43 4.10 3.80 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: default@1 33,731 searches in 30.0 s, exact 12,101 searches in 30.0 s, default@8 93,946 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. default@1: warm-up 15.0 s and 15,682 searches (0 failed) at 1 searcher; p50 of the last windows 0.927, 0.938, 0.932 ms (windows of at least 2 s and 100 searches); trial of 3,104 searches in 3.0 s at 1 searcher: p50 0.941 ms against the settled p50 0.932 ms, 1% apart (limit 10%); the timed pass p50 0.934 ms, 0% from the settled p50 0.932 ms (limit 10% or 0.1 ms; 0.002 ms); exact: warm-up 15.0 s and 5,849 searches (0 failed) at 1 searcher; p50 of the last windows 2.496, 2.495, 2.488 ms (windows of at least 2 s and 100 searches); trial of 1,157 searches in 3.0 s at 1 searcher: p50 2.496 ms against the settled p50 2.495 ms, 0% apart (limit 10%); the timed pass p50 2.490 ms, 0% from the settled p50 2.495 ms (limit 10% or 0.1 ms; 0.005 ms); default@8: warm-up 15.4 s and 43,861 searches (0 failed) at 8 searchers; QPS of the last windows 3,166, 2,620, 2,789 (windows of at least 2 s and 100 searches) (its last 3 windows differed by more than 5%); trial of 9,100 searches in 3.0 s at 8 searchers: 3,032 QPS against the settled 2,789 QPS, 8% apart (limit 10%), so the warm-up was EXTENDED once (30.0 s and 84,126 searches (0 failed) at 8 searchers; QPS of its older 6 windows 2,840 (windows 2,989, 2,415, 3,394, 2,488, 2,929, 2,824) and of its newer 6 2,831 (windows 2,594, 3,204, 2,540, 3,234, 2,534, 2,877), 0.3% apart (limit 5%) (windows of at least 2 s and 100 searches)); second trial of 9,294 searches in 3.0 s at 8 searchers: 3,096 QPS against the settled 2,831 QPS, 9% apart (limit 10%), inside its newer windows' range 2,534 to 3,234 QPS; settled after the extension; the timed pass 2,655 QPS, 7% from the settled 2,831 QPS (limit 10%). - Pass default@1 after rehearsal: 2026-10-06T14:58:20.794Z to 2026-10-06T14:58:40.795Z (20.0 s), 20,842 searches, 0 failed, p50 0.934 ms, mean 0.958 ms, p99 1.870 ms, 1042.1 QPS (1000/QPS 0.960 ms). - Pass exact after default@1: 2026-10-06T14:58:58.815Z to 2026-10-06T14:59:58.817Z (60.0 s), 23,442 searches, 0 failed, p50 2.490 ms, mean 2.557 ms, p99 3.880 ms, 390.7 QPS (1000/QPS 2.560 ms). - Pass default@8 after exact: 2026-10-06T15:01:05.295Z to 2026-10-06T15:01:25.422Z (20.1 s), 53,445 searches, 0 failed, p50 1.675 ms, mean 3.010 ms, p99 3.855 ms, 2655.3 QPS (1000/QPS 0.377 ms). - Passes in the order run: default@1, exact, default@8; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 354,234 sent, 0 failed (warmupErrors). Index after the searches: ready, 524 of 524 indexed (HNSW graph holds 524 of 524 rows, change log waiting: 0 inserts and 0 deletes, USER_INDEXES status VALID, plan of the default search: VECTOR INDEX HNSW SCAN > TABLE ACCESS BY INDEX ROWID, index used by 409414 queries so far; Oracle Free caps itself at 2 CPUs (cpu_count 2 in V$PARAMETER, edition FREE in V$INSTANCE, 8 host CPUs in V$OSSTAT NUM_CPUS; the 2 CPU thread limit is Oracle's documented Free edition limit)). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: container address on its Docker network, not the published port (no docker-proxy): container gvb-oracle (959e18947252) on network gvb-oracle_default at container-ip:1521, not through docker-proxy localhost:1521. Open after the passes: container-ip:1521 x15 (this target). - CPU pinning: cpuset from the container's creation when run-all started it on the engine CPUs, else docker update --cpuset-cpus (each container's change line says which), CPUs 2-3,6-7. Changed: container gvb-oracle: already on CPUs 2-3,6-7 (its cpuset since creation), not changed. Read back: container gvb-oracle: 86 thread(s) on CPUs 2-3,6-7 (every thread of the container). - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): default@1 3492/3492 MHz, average 3492.1/3491.9, performance, outside load 0.22 before the warm-up, 0.23 during, client CPU 0.888 ms per search; exact 3492/3492 MHz, average 3491.6/3491.1, performance, outside load 0.22 before the warm-up, 0.21 during, client CPU 0.956 ms per search; default@8 3492/3492 MHz, average 3129.1/3229.8, performance, outside load 0.26 before the warm-up, 0.22 during, client CPU 0.654 ms per search. - WARNING: clock off its pinned value during oracle default@8: the engine CPUs averaged 3129 MHz, -10.6% against the pinned 3500 MHz; the client CPUs averaged 3230 MHz, -7.7% against the pinned 3500 MHz (limit 1%); this pass is not comparable with passes at the pinned clock. ### milvus - Engine: Milvus 2.6.25 (standalone, embedded etcd, local storage, REST API v2) (compose) - Index: HNSW M=16 efConstruction=128, ef=100, metric COSINE, Strong consistency searches; approximate only (no exact mode) - Search settings: M=16, ef=100, efConstruction=128 - Load: 524 rows in batches of 1000, 0.4 s of upserts (1,178 rows/s); count matched 1.7 s after the last upsert; index step 138.5 s (index Finished, type HNSW (COSINE, {"M":16,"efConstruction":128}), indexedRows 524 of 524 sealed rows, pendingRows 0; stored rows 524; LoadStateLoaded; query node: 1 sealed segment(s) with the index loaded covering 524 rows against 524 stored rows, 0 sealed without it, 0 rows in growing segments; datacoord: 1 live segment(s), 1 flushed and indexed covering 524 rows, 0 delete-log (L0) segment(s) waiting for a compaction, 3 segment(s) compacted away so far; flush, index build and load took 14.6 s, then compaction: 1 job(s) accepted by Milvus, layout ready and unchanged for 75 s, 123.9 s for the whole step; search ledger started, later state reads count the searches against the query node's own counters) - Search: 8171 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 0.56 ms, 8 searchers 0.64 ms - RAM: 213 MiB (docker stats); disk: 10.86 MiB added by this load (211.13 MiB (whole engine data folder), 200.26 MiB before) - Started by run-all for this measurement with every container created on CPUs 2-3,6-7 (started milvus.compose.yaml with every container created on CPUs 2-3,6-7: gvb-milvus cpuset 2-3,6-7 (read back from docker inspect)); stopped afterwards. - Engine files (SHA-256, first 12 hex digits): [milvus.compose.yaml eaefecdc3671]. This run created the engine from these files. - Index after the load: ready, 524 of 524 indexed (index Finished, type HNSW (COSINE, {"M":16,"efConstruction":128}), indexedRows 524 of 524 sealed rows, pendingRows 0; stored rows 524; LoadStateLoaded; query node: 1 sealed segment(s) with the index loaded covering 524 rows against 524 stored rows, 0 sealed without it, 0 rows in growing segments; datacoord: 1 live segment(s), 1 flushed and indexed covering 524 rows, 0 delete-log (L0) segment(s) waiting for a compaction, 3 segment(s) compacted away so far; search ledger since the index finished: the sink sent 0 searches (searchParams.params.ef=100, Strong consistency), the query node counted 0 search requests for this collection, segment searches Sealed +0 and Growing +0, the same sealed segments with the same loaded index builds at both readings, segments compacted away +0). Durability: Writes are not fsynced. Standalone uses the default message queue rocksmq (RocksDB, /var/lib/milvus/rdb_data, mq.type default) and flushed segments go to local-disk object storage (COMMON_STORAGETYPE=local in milvus.compose.yaml). Measured 2026-10-04 with strace over 30 acknowledged single-row upserts and one flush: fdatasync ran only on the embedded etcd files (member/wal, member/snap/db), never on rdb_data or the segment files. A host power loss or kernel crash can lose acknowledged rows that are still in the OS page cache; a Milvus process crash alone should not (inferred, not tested). Metadata in embedded etcd is fdatasynced on every commit.. - Outside load when searching began: 0.20 CPUs on average over the last 60 s and 0.23 over the last few seconds (limit 0.3), a quiet box. The load average 0.77 2.59 3.28 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: default@1 12,236 searches in 30.0 s, default@8 37,588 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. default@1: warm-up 15.0 s and 6,110 searches (0 failed) at 1 searcher; p50 of the last windows 2.340, 2.321, 2.324 ms (windows of at least 2 s and 100 searches); trial of 1,237 searches in 3.0 s at 1 searcher: p50 2.308 ms against the settled p50 2.324 ms, 1% apart (limit 10%); the timed pass p50 2.316 ms, 0% from the settled p50 2.324 ms (limit 10% or 0.1 ms; 0.009 ms); default@8: warm-up 15.0 s and 18,814 searches (0 failed) at 8 searchers; QPS of the last windows 1,248, 1,252, 1,256 (windows of at least 2 s and 100 searches); trial of 3,784 searches in 3.0 s at 8 searchers: 1,260 QPS against the settled 1,252 QPS, 1% apart (limit 10%); the timed pass 1,258 QPS, 1% from the settled 1,252 QPS (limit 10%). - Pass default@1 after rehearsal: 2026-10-06T15:05:17.246Z to 2026-10-06T15:05:37.247Z (20.0 s), 8,171 searches, 0 failed, p50 2.316 ms, mean 2.445 ms, p99 3.830 ms, 408.5 QPS (1000/QPS 2.448 ms). - Pass default@8 after default@1: 2026-10-06T15:05:55.262Z to 2026-10-06T15:06:15.266Z (20.0 s), 25,165 searches, 0 failed, p50 5.815 ms, mean 6.356 ms, p99 14.631 ms, 1258.0 QPS (1000/QPS 0.795 ms). - Passes in the order run: default@1, default@8; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 79,769 sent, 0 failed (warmupErrors). Index after the searches: ready, 524 of 524 indexed (index Finished, type HNSW (COSINE, {"M":16,"efConstruction":128}), indexedRows 524 of 524 sealed rows, pendingRows 0; stored rows 524; LoadStateLoaded; query node: 1 sealed segment(s) with the index loaded covering 524 rows against 524 stored rows, 0 sealed without it, 0 rows in growing segments; datacoord: 1 live segment(s), 1 flushed and indexed covering 524 rows, 0 delete-log (L0) segment(s) waiting for a compaction, 3 segment(s) compacted away so far; search ledger since the index finished: the sink sent 113105 searches (searchParams.params.ef=100, Strong consistency), the query node counted 113105 search requests for this collection, segment searches Sealed +111020 and Growing +0, the same sealed segments with the same loaded index builds at both readings, segments compacted away +0). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: container address on its Docker network, not the published port (no docker-proxy): container gvb-milvus (6b05295a9e29) on network gvb-milvus_default at container-ip:19530, not through docker-proxy localhost:19530; container gvb-milvus (6b05295a9e29) on network gvb-milvus_default at container-ip:9091, not through docker-proxy localhost:9091. Open after the passes: container-ip:19530 x8 (this target). - CPU pinning: cpuset from the container's creation when run-all started it on the engine CPUs, else docker update --cpuset-cpus (each container's change line says which), CPUs 2-3,6-7. Changed: container gvb-milvus: already on CPUs 2-3,6-7 (its cpuset since creation), not changed. Read back: container gvb-milvus: 24 thread(s) on CPUs 2-3,6-7 (every thread of the container). - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): default@1 3492/3492 MHz, average 3492.0/3492.4, performance, outside load 0.22 before the warm-up, 0.22 during, client CPU 0.564 ms per search; default@8 3492/3492 MHz, average 3492.0/3492.2, performance, outside load 0.22 before the warm-up, 0.26 during, client CPU 0.636 ms per search. ### clickhouse - Engine: ClickHouse 26.3.39.7 (MergeTree + vector_similarity index; image-default system logging, each log table with a TTL of 1 hour) (compose) - Index: vector_similarity HNSW cosineDistance, quantization bf16, M=16 ef_construction=128, hnsw_candidate_list_size_for_search=256, rescoring off; exact mode = full scan with skip indexes off - Search settings: M=16, ef_construction=128, hnsw_candidate_list_size_for_search=256 - Load: 524 rows in batches of 1000, 0.2 s of upserts (2,160 rows/s); count matched 0.0 s after the last upsert; index step 6.2 s (1 data part(s), index files on 1 of them, 524 of 524 live rows in indexed parts; 0 merge(s) running, 0 unfinished mutation(s); plan of the default search uses the Skip index vec_idx on 1 of 1 parts; wait took 6.2 s) - Search: 3916 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 0.88 ms, 8 searchers 1.06 ms - Exact mode: 8,741 searches, p50 6.66 ms, p95 7.99 ms, recall 1.000 - RAM: 892.7 MiB (docker stats); disk: 2.57 GiB added by this load (6.79 GiB (whole engine data folder), 4.22 GiB before) - Started by run-all for this measurement with every container created on CPUs 2-3,6-7 (started clickhouse.compose.yaml with every container created on CPUs 2-3,6-7: gvb-clickhouse cpuset 2-3,6-7 (read back from docker inspect)); stopped afterwards. - Engine files (SHA-256, first 12 hex digits): [clickhouse-config/gvb-system-logs.xml 02cc53f0f171; clickhouse.compose.yaml 560e0ae487c6]. This run created the engine from these files. - Index after the load: ready, 524 of 524 indexed (1 data part(s), index files on 1 of them, 524 of 524 live rows in indexed parts; 0 merge(s) running, 0 unfinished mutation(s); plan of the default search uses the Skip index vec_idx on 1 of 1 parts). Durability: Acknowledged inserts are not fsynced. The sink creates plain MergeTree tables, so the MergeTree settings are the defaults: fsync_after_insert 0 and fsync_part_directory 0 (system.merge_tree_settings), and async_insert 1 with wait_for_async_insert 1 is the server default for the gvb login (system.settings), so an insert is acknowledged once its part is written, not once it is on disk. Measured 2026-10-04 with strace over 30 acknowledged single-row inserts (30 Ok rows in system.asynchronous_insert_log): zero fsync, fdatasync or sync_file_range calls, while the control, a table created with fsync_after_insert=1, made 61 fdatasync calls for 5 inserts. A host power loss or kernel crash can lose acknowledged rows that are still in the OS page cache; a ClickHouse process crash alone should not (inferred, not tested).. - Outside load when searching began: 0.21 CPUs on average over the last 7 s and 0.22 over the last few seconds (limit 0.3), a quiet box. The load average 3.84 3.19 3.39 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: default@8 13,056 searches in 30.0 s, default@1 5,834 searches in 30.0 s, exact 4,280 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. default@8: warm-up 15.0 s and 6,560 searches (0 failed) at 8 searchers; QPS of the last windows 470, 466, 438 (windows of at least 2 s and 100 searches) (its last 3 windows differed by more than 5%); trial of 1,333 searches in 3.0 s at 8 searchers: 442 QPS against the settled 466 QPS, 5% apart (limit 10%), so the warm-up was EXTENDED once (30.0 s and 12,161 searches (0 failed) at 8 searchers; QPS of its older 6 windows 427 (windows 462, 324, 410, 472, 435, 456) and of its newer 6 410 (windows 335, 464, 448, 464, 436, 314), 4.0% apart (limit 5%) (windows of at least 2 s and 100 searches)); second trial of 1,269 searches in 3.0 s at 8 searchers: 422 QPS against the settled 410 QPS, 3% apart (limit 10%), inside its newer windows' range 314 to 464 QPS; settled after the extension; the timed pass 428 QPS, 4% from the settled 410 QPS (limit 10%); default@1: warm-up 15.0 s and 2,588 searches (0 failed) at 1 searcher; p50 of the last windows 5.058, 7.957, 5.333 ms (windows of at least 2 s and 100 searches) (its last 3 windows differed by more than 5%); trial of 526 searches in 3.0 s at 1 searcher: p50 5.137 ms against the settled p50 5.333 ms, 4% apart (limit 10%), so the warm-up was EXTENDED once (30.0 s and 5,269 searches (0 failed) at 1 searcher; p50 of its older 6 windows 5.312 ms (windows 7.591, 5.234, 5.334, 5.438, 5.207, 5.150) and of its newer 6 5.130 ms (windows 5.477, 5.324, 5.187, 5.146, 4.951, 4.946), 3.6% apart (limit 5% or 0.1 ms; 0.182 ms) (windows of at least 2 s and 100 searches)); second trial of 571 searches in 3.0 s at 1 searcher: p50 4.987 ms against the settled p50 5.130 ms, 3% apart (limit 10%), inside its newer windows' range 4.946 to 5.477 ms; settled after the extension; the timed pass p50 4.941 ms, 4% from the settled p50 5.130 ms (limit 10% or 0.1 ms; 0.188 ms); exact: warm-up 15.0 s and 2,192 searches (0 failed) at 1 searcher; p50 of the last windows 6.650, 6.718, 6.667 ms (windows of at least 2 s and 100 searches); trial of 421 searches in 3.0 s at 1 searcher: p50 6.700 ms against the settled p50 6.667 ms, 0% apart (limit 10%); the timed pass p50 6.657 ms, 0% from the settled p50 6.667 ms (limit 10% or 0.1 ms; 0.010 ms). - Pass default@8 after rehearsal: 2026-10-06T15:09:07.681Z to 2026-10-06T15:09:27.692Z (20.0 s), 8,574 searches, 0 failed, p50 17.957 ms, mean 18.663 ms, p99 35.279 ms, 428.5 QPS (1000/QPS 2.334 ms). - Pass default@1 after default@8: 2026-10-06T15:10:33.715Z to 2026-10-06T15:10:53.720Z (20.0 s), 3,916 searches, 0 failed, p50 4.941 ms, mean 5.105 ms, p99 7.742 ms, 195.8 QPS (1000/QPS 5.109 ms). - Pass exact after default@1: 2026-10-06T15:11:11.740Z to 2026-10-06T15:12:11.741Z (60.0 s), 8,741 searches, 0 failed, p50 6.657 ms, mean 6.860 ms, p99 10.275 ms, 145.7 QPS (1000/QPS 6.864 ms). - Passes in the order run: default@8, default@1, exact; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 65,255 sent, 0 failed (warmupErrors). Index after the searches: ready, 524 of 524 indexed (1 data part(s), index files on 1 of them, 524 of 524 live rows in indexed parts; 0 merge(s) running, 0 unfinished mutation(s); plan of the default search uses the Skip index vec_idx on 1 of 1 parts). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: container address on its Docker network, not the published port (no docker-proxy): container gvb-clickhouse (d0120a1f26f2) on network gvb-clickhouse_default at container-ip:8123, not through docker-proxy localhost:8123. Open after the passes: container-ip:8123 x1 (this target). - CPU pinning: cpuset from the container's creation when run-all started it on the engine CPUs, else docker update --cpuset-cpus (each container's change line says which), CPUs 2-3,6-7. Changed: container gvb-clickhouse: already on CPUs 2-3,6-7 (its cpuset since creation), not changed. Read back: container gvb-clickhouse: 733 thread(s) on CPUs 2-3,6-7 (every thread of the container). - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): default@8 3492/3492 MHz, average 3492.0/3492.1, performance, outside load 0.25 before the warm-up, 0.22 during, client CPU 1.055 ms per search; default@1 3492/3492 MHz, average 3492.0/3492.4, performance, outside load 0.26 before the warm-up, 0.24 during, client CPU 0.878 ms per search; exact 3492/3492 MHz, average 3492.0/3492.6, performance, outside load 0.25 before the warm-up, 0.24 during, client CPU 0.903 ms per search. ### sql - Engine: Microsoft SQL Server 2025 (RTM-CU9) (KB5122048) - 17.0.5005.3 (X64), Enterprise Developer Edition (64-bit), in container gvb-mssql (image mcr.microsoft.com/mssql/server:2025-CU9-ubuntu-24.04@sha256:2b5b581621126574f3d1f75e78d3eebe8d05aedb59ad0cfdf9aa42cb0634d726; cpuset 2-3,6-7 from its creation; SQL Server counts 4 CPU(s) and runs 4 visible scheduler(s), affinity AUTO) (compose) - Index: exact VECTOR_DISTANCE cosine, no vector index (full scan) - Search settings: description=exact VECTOR_DISTANCE cosine, no vector index (full scan) - Load: 524 rows in batches of 1000, 0.9 s of upserts (586 rows/s); count matched 0.0 s after the last upsert; index step n/a - Search: 4858 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 1.22 ms, 8 searchers 1.20 ms - RAM: 532.2 MiB (docker stats); disk: 5.15 MiB (table and its indexes, reserved pages) - Started by run-all for this measurement with every container created on CPUs 2-3,6-7 (started mssql.compose.yaml with every container created on CPUs 2-3,6-7: gvb-mssql cpuset 2-3,6-7 (read back from docker inspect)); stopped afterwards. - Engine files (SHA-256, first 12 hex digits): [mssql.compose.yaml b7359663cc15]. This run created the engine from these files. - Index after the load: ready, 0 of 524 indexed (no index used, exact scan by design: dbo.gvb_gvbbench_eshoponweb in GvbBench holds 524 rows; sys.vector_indexes lists no index on the table; indexes: PK_gvb_gvbbench_eshoponweb CLUSTERED, IX_gvb_gvbbench_eshoponweb_DocKey NONCLUSTERED; a real search under SET STATISTICS XML ON ran: Clustered Index Scan of GvbBench.gvb_gvbbench_eshoponweb through PK_gvb_gvbbench_eshoponweb (IndexKind Clustered), 524 rows read, no vector index operator, 10 hits). Durability: A commit returns after its transaction-log records are written to disk (SQL Server write-ahead logging; delayed durability is DISABLED); a database created here copies the model database: recovery model FULL, page_verify CHECKSUM; no global trace flags are enabled; mssql.conf of container gvb-mssql (/var/opt/mssql/mssql.conf, on the host at ~/gvb-data/engines/mssql/mssql.conf) does not exist, so it sets: nothing; it has no [control] or [traceflag] entry, so SQL Server's own Linux defaults for flushing writes apply (Microsoft's Linux performance guide names trace flag 3982 as that default; read from the guide, not tested here). Not tested by cutting power; whether the disk's own write cache reaches the media was not checked. Container settings from its environment (names only): MSSQL_AGENT_ENABLED, MSSQL_MEMORY_LIMIT_MB, MSSQL_PID, MSSQL_RPC_PORT, MSSQL_SA_PASSWORD.. - Outside load when searching began: 0.22 CPUs on average over the last 3 s and 0.22 over the last few seconds (limit 0.3), a quiet box. The load average 2.02 4.12 3.97 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: default@1 7,335 searches in 30.0 s, default@8 17,941 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. default@1: warm-up 15.0 s and 3,625 searches (0 failed) at 1 searcher; p50 of the last windows 4.059, 4.073, 4.051 ms (windows of at least 2 s and 100 searches); trial of 725 searches in 3.0 s at 1 searcher: p50 4.034 ms against the settled p50 4.059 ms, 1% apart (limit 10%); the timed pass p50 4.016 ms, 1% from the settled p50 4.059 ms (limit 10% or 0.1 ms; 0.043 ms); default@8: warm-up 15.0 s and 9,185 searches (0 failed) at 8 searchers; QPS of the last windows 602, 620, 619 (windows of at least 2 s and 100 searches); trial of 1,875 searches in 3.0 s at 8 searchers: 623 QPS against the settled 619 QPS, 1% apart (limit 10%); the timed pass 602 QPS, 3% from the settled 619 QPS (limit 10%). - Pass default@1 after rehearsal: 2026-10-06T15:13:47.925Z to 2026-10-06T15:14:07.927Z (20.0 s), 4,858 searches, 0 failed, p50 4.016 ms, mean 4.114 ms, p99 6.017 ms, 242.9 QPS (1000/QPS 4.117 ms). - Pass default@8 after default@1: 2026-10-06T15:14:25.951Z to 2026-10-06T15:14:45.963Z (20.0 s), 12,052 searches, 0 failed, p50 17.471 ms, mean 13.275 ms, p99 20.852 ms, 602.2 QPS (1000/QPS 1.660 ms). - Passes in the order run: default@1, default@8; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 40,686 sent, 0 failed (warmupErrors). Index after the searches: ready, 0 of 524 indexed (no index used, exact scan by design: dbo.gvb_gvbbench_eshoponweb in GvbBench holds 524 rows; sys.vector_indexes lists no index on the table; indexes: PK_gvb_gvbbench_eshoponweb CLUSTERED, IX_gvb_gvbbench_eshoponweb_DocKey NONCLUSTERED; a real search under SET STATISTICS XML ON ran: Clustered Index Scan of GvbBench.gvb_gvbbench_eshoponweb through PK_gvb_gvbbench_eshoponweb (IndexKind Clustered), 524 rows read, no vector index operator, 10 hits). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: container address on its Docker network, not the published port (no docker-proxy): container gvb-mssql (3f8b7a4913d4) on network gvb-mssql_default at container-ip:1433, not through docker-proxy localhost:14330. Open after the passes: container-ip:1433 x9 (this target). - CPU pinning: cpuset from the container's creation when run-all started it on the engine CPUs, else docker update --cpuset-cpus (each container's change line says which), CPUs 2-3,6-7. Changed: container gvb-mssql: already on CPUs 2-3,6-7 (its cpuset since creation), not changed. Read back: container gvb-mssql: 87 thread(s) on CPUs 2-3,6-7 (every thread of the container). - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): default@1 3492/3492 MHz, average 3492.4/3492.6, performance, outside load 0.21 before the warm-up, 0.22 during, client CPU 1.220 ms per search; default@8 3492/3492 MHz, average 3492.0/3492.3, performance, outside load 0.21 before the warm-up, 0.25 during, client CPU 1.204 ms per search. ### duckdb - Engine: DuckDB 1.5.6 + vss b833341 (HNSW, embedded, in process) (embedded) - Index: HNSW (vss extension) FLOAT[n] metric=cosine m=16 ef_construction=128, ef_search=100 per connection, persistent (hnsw_enable_experimental_persistence=true, checkpoint_threshold=256MB); exact mode = array_cosine_similarity sequential scan; score = 1 - cosine distance; searches run concurrently, one connection per searcher (opened as searchers arrive, at most 32), writes run one at a time and never overlap a search - Search settings: checkpoint_threshold=256MB, ef_construction=128, ef_search=100, hnsw_enable_experimental_persistence=true, m=16, metric=cosine - Load: 524 rows in batches of 1000, 0.4 s of upserts (1,486 rows/s); count matched 0.0 s after the last upsert; index step 0.0 s (HNSW is maintained inside every transaction, nothing to build; confirmed in 0.01 s: duckdb_indexes() lists gvb_gvbbench_eshoponweb_hnsw, pragma_hnsw_index_info() counts 524 vectors of 524 rows, EXPLAIN of the default search shows HNSW_INDEX_SCAN on it = True) - Search: 5382 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 5.03 ms, 8 searchers 6.90 ms - Exact mode: 14,016 searches, p50 4.20 ms, p95 4.93 ms, recall 1.000 - RAM: in this process, not measured; disk: 2.9 MiB added by this load (1.13 GiB (engine data folder), 1.13 GiB before) - Index after the load: ready, 524 of 524 indexed (duckdb_indexes() lists gvb_gvbbench_eshoponweb_hnsw, pragma_hnsw_index_info() counts 524 vectors of 524 rows, EXPLAIN of the default search shows HNSW_INDEX_SCAN on it = True). Durability: DuckDB's write-ahead log is fsynced at every commit and no DuckDbSink setting changes that (checkpoint_threshold=256MB only spaces out the checkpoints that write the database file; measured with strace: 42 WAL fsyncs for 40 single-record commits plus 2 setup statements), so an operating-system crash or power cut loses no committed row; measured with kill -9 right after the last commit and a reopen: all 50, 300 and 2,000 rows and an HNSW index that counted the same number were recovered from the log; not tested and documented by DuckDB: the HNSW index is file-backed only through hnsw_enable_experimental_persistence = true, which this sink turns on, and WAL recovery for such custom indexes is not complete, so a crash during a checkpoint or a later commit can damage the index while the rows survive. - Outside load when searching began: not known yet (too little sampled history). Load average 3.56 3.90 3.90 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: exact 7,163 searches in 30.0 s, default@8 17,430 searches in 30.0 s, default@1 8,129 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. exact: warm-up 15.0 s and 3,474 searches (0 failed) at 1 searcher; p50 of the last windows 4.209, 4.252, 4.266 ms (windows of at least 2 s and 100 searches); trial of 698 searches in 3.0 s at 1 searcher: p50 4.230 ms against the settled p50 4.252 ms, 1% apart (limit 10%); the timed pass p50 4.200 ms, 1% from the settled p50 4.252 ms (limit 10% or 0.1 ms; 0.052 ms); default@8: warm-up 15.0 s and 8,659 searches (0 failed) at 8 searchers; QPS of the last windows 580, 576, 573 (windows of at least 2 s and 100 searches); trial of 1,756 searches in 3.0 s at 8 searchers: 584 QPS against the settled 576 QPS, 1% apart (limit 10%); the timed pass 576 QPS, 0% from the settled 576 QPS (limit 10%); default@1: warm-up 15.0 s and 4,051 searches (0 failed) at 1 searcher; p50 of the last windows 3.669, 3.676, 3.635 ms (windows of at least 2 s and 100 searches); trial of 813 searches in 3.0 s at 1 searcher: p50 3.667 ms against the settled p50 3.669 ms, 0% apart (limit 10%); the timed pass p50 3.678 ms, 0% from the settled p50 3.669 ms (limit 10% or 0.1 ms; 0.009 ms). - Pass exact after rehearsal: 2026-10-06T15:16:36.481Z to 2026-10-06T15:17:36.485Z (60.0 s), 14,016 searches, 0 failed, p50 4.200 ms, mean 4.277 ms, p99 5.818 ms, 233.6 QPS (1000/QPS 4.281 ms). - Pass default@8 after exact: 2026-10-06T15:17:54.514Z to 2026-10-06T15:18:14.522Z (20.0 s), 11,524 searches, 0 failed, p50 13.189 ms, mean 13.878 ms, p99 28.090 ms, 576.0 QPS (1000/QPS 1.736 ms). - Pass default@1 after default@8: 2026-10-06T15:18:32.530Z to 2026-10-06T15:18:52.530Z (20.0 s), 5,382 searches, 0 failed, p50 3.678 ms, mean 3.713 ms, p99 5.275 ms, 269.1 QPS (1000/QPS 3.716 ms). - Passes in the order run: exact, default@8, default@1; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 52,173 sent, 0 failed (warmupErrors). Index after the searches: ready, 524 of 524 indexed (duckdb_indexes() lists gvb_gvbbench_eshoponweb_hnsw, pragma_hnsw_index_info() counts 524 vectors of 524 rows, EXPLAIN of the default search shows HNSW_INDEX_SCAN on it = True). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: in this process (embedded), no network: no network address. No open connection seen (no open TCP connection was found). - CPU pinning: none: an embedded engine runs inside the benchmark client process, so it shares the client CPUs, CPUs 0-1,4-5. - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): exact 3492/3492 MHz, average 3492.8/3492.2, performance, outside load 0.20 before the warm-up, 0.22 during, client CPU 6.271 ms per search; default@8 3492/3492 MHz, average 3492.7/3492.0, performance, outside load 0.22 before the warm-up, 0.21 during, client CPU 6.899 ms per search; default@1 3492/3492 MHz, average 3493.1/3492.1, performance, outside load 0.21 before the warm-up, 0.19 during, client CPU 5.026 ms per search. ### pgvector - Engine: PostgreSQL 17 + pgvector 0.8.7 (compose) - Index: HNSW vector_cosine_ops m=16 ef_construction=128, hnsw.ef_search=100 per query, float32 vector(n), cosine; exact mode = same query with index scans off (sequential scan) - Search settings: ef_construction=128, hnsw.ef_search=100, m=16 - Load: 524 rows in batches of 1000, 0.9 s of upserts (609 rows/s); count matched 0.0 s after the last upsert; index step 0.0 s (HNSW is maintained inside every insert, nothing to build; confirmed in 0.0 s: pg_indexes lists gvb_gvbbench_eshoponweb_hnsw (USING hnsw (embedding vector_cosine_ops) WITH (m='16', ef_construction='128')), pg_index valid and ready = True, EXPLAIN of the default search uses Index Scan on it = True, that search returned 10 of 10 rows; PostgreSQL keeps no entry count for an HNSW index, so indexed vectors are not reported) - Search: 22514 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 0.69 ms, 8 searchers 0.54 ms - Exact mode: 26,494 searches, p50 2.21 ms, p95 2.54 ms, recall 1.000 - RAM: 99.93 MiB (docker stats); disk: 7.56 MiB added by this load (374.57 MiB (whole engine data folder), 367 MiB before) - Started by run-all for this measurement with every container created on CPUs 2-3,6-7 (started pgvector.compose.yaml with every container created on CPUs 2-3,6-7: gvb-pgvector cpuset 2-3,6-7 (read back from docker inspect)); stopped afterwards. - Engine files (SHA-256, first 12 hex digits): [pgvector.compose.yaml 27549563c622]. This run created the engine from these files. - Index after the load: ready, ? of 524 indexed (pg_indexes lists gvb_gvbbench_eshoponweb_hnsw (USING hnsw (embedding vector_cosine_ops) WITH (m='16', ef_construction='128')), pg_index valid and ready = True, EXPLAIN of the default search uses Index Scan on it = True, that search returned 10 of 10 rows; PostgreSQL keeps no entry count for an HNSW index, so indexed vectors are not reported). Durability: fsync on, synchronous_commit on, full_page_writes on, wal_sync_method fdatasync (PostgreSQL defaults; pgvector.compose.yaml sets only shared_buffers, maintenance_work_mem and max_wal_size): every commit is flushed to the write-ahead log before it returns and HNSW index changes are WAL-logged, so a crash loses no committed row (max_wal_size 4GB only spaces out checkpoints; measured with docker kill, which is SIGKILL, right after a 2,000-vector load and a restart: all 2,000 rows were there and the HNSW index was valid and used by the default search). - Outside load when searching began: 0.16 CPUs on average over the last 1 s and 0.16 over the last few seconds (limit 0.3), a quiet box. The load average 4.08 4.31 4.08 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: default@8 109,515 searches in 30.0 s, exact 13,259 searches in 30.0 s, default@1 33,987 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. default@8: warm-up 15.0 s and 59,807 searches (0 failed) at 8 searchers; QPS of the last windows 3,938, 4,022, 3,970 (windows of at least 2 s and 100 searches); trial of 12,015 searches in 3.0 s at 8 searchers: 4,004 QPS against the settled 3,970 QPS, 1% apart (limit 10%); the timed pass 4,001 QPS, 1% from the settled 3,970 QPS (limit 10%); exact: warm-up 15.0 s and 6,605 searches (0 failed) at 1 searcher; p50 of the last windows 2.198, 2.214, 2.228 ms (windows of at least 2 s and 100 searches); trial of 1,321 searches in 3.0 s at 1 searcher: p50 2.209 ms against the settled p50 2.214 ms, 0% apart (limit 10%); the timed pass p50 2.210 ms, 0% from the settled p50 2.214 ms (limit 10% or 0.1 ms; 0.005 ms); default@1: warm-up 15.0 s and 16,848 searches (0 failed) at 1 searcher; p50 of the last windows 0.883, 0.877, 0.864 ms (windows of at least 2 s and 100 searches); trial of 3,436 searches in 3.0 s at 1 searcher: p50 0.852 ms against the settled p50 0.877 ms, 3% apart (limit 10%); the timed pass p50 0.870 ms, 1% from the settled p50 0.877 ms (limit 10% or 0.1 ms; 0.008 ms). - Pass default@8 after rehearsal: 2026-10-06T15:20:48.309Z to 2026-10-06T15:21:08.310Z (20.0 s), 80,030 searches, 0 failed, p50 1.952 ms, mean 1.997 ms, p99 3.550 ms, 4001.3 QPS (1000/QPS 0.250 ms). - Pass exact after default@8: 2026-10-06T15:21:26.349Z to 2026-10-06T15:22:26.351Z (60.0 s), 26,494 searches, 0 failed, p50 2.210 ms, mean 2.263 ms, p99 3.507 ms, 441.6 QPS (1000/QPS 2.265 ms). - Pass default@1 after exact: 2026-10-06T15:22:44.375Z to 2026-10-06T15:23:04.376Z (20.0 s), 22,514 searches, 0 failed, p50 0.870 ms, mean 0.886 ms, p99 1.292 ms, 1125.7 QPS (1000/QPS 0.888 ms). - Passes in the order run: default@8, exact, default@1; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 256,793 sent, 0 failed (warmupErrors). Index after the searches: ready, ? of 524 indexed (pg_indexes lists gvb_gvbbench_eshoponweb_hnsw (USING hnsw (embedding vector_cosine_ops) WITH (m='16', ef_construction='128')), pg_index valid and ready = True, EXPLAIN of the default search uses Index Scan on it = True, that search returned 10 of 10 rows; PostgreSQL keeps no entry count for an HNSW index, so indexed vectors are not reported). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: container address on its Docker network, not the published port (no docker-proxy): container gvb-pgvector (39f0cbb37d9b) on network gvb-pgvector_default at container-ip:5432, not through docker-proxy localhost:5432. Open after the passes: container-ip:5432 x8 (this target). - CPU pinning: cpuset from the container's creation when run-all started it on the engine CPUs, else docker update --cpuset-cpus (each container's change line says which), CPUs 2-3,6-7. Changed: container gvb-pgvector: already on CPUs 2-3,6-7 (its cpuset since creation), not changed. Read back: container gvb-pgvector: 6 thread(s) on CPUs 2-3,6-7 (every thread of the container). - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): default@8 3492/3492 MHz, average 3492.0/3492.0, performance, outside load 0.20 before the warm-up, 0.22 during, client CPU 0.538 ms per search; exact 3492/3492 MHz, average 3492.3/3492.1, performance, outside load 0.21 before the warm-up, 0.21 during, client CPU 0.694 ms per search; default@1 3492/3492 MHz, average 3492.3/3492.1, performance, outside load 0.21 before the warm-up, 0.20 during, client CPU 0.687 ms per search. ### elasticsearch - Engine: Elasticsearch 9.5.3 (dense_vector) (compose) - Index: HNSW float32, no quantization, m=16, ef_construction=128, cosine; search k=top, num_candidates=100; 1 shard, 0 replicas; force-merged to one segment after the load (at 1,024 dimensions a segment under 1,043 vectors gets no graph) - Search settings: ef_construction=128, k=top, m=16, num_candidates=100 - Load: 524 rows in batches of 1000, 1.3 s of upserts (416 rows/s); count matched 0.1 s after the last upsert; index step 0.4 s (FAILED after 0.4 s: NO HNSW GRAPH, searches scan all 524 vectors: 1 segment(s), 524 vectors, total_vex_size_bytes 0 (_stats dense_vector), index_options hnsw m=16 ef_construction=128; Lucene builds no graph for a segment this small (measured at 1,024 dimensions: 1,042 vectors none, 1,043 vectors a graph), so default search equals exact search) - Search: 14788 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 0.46 ms, 8 searchers 0.49 ms - Exact mode: 45,908 searches, p50 1.28 ms, p95 1.49 ms, recall 1.000 - RAM: 2.61 GiB (docker stats); disk: 2.39 MiB added by this load (2.55 MiB (whole engine data folder), 172.72 KiB before) - Started by run-all for this measurement with every container created on CPUs 2-3,6-7 (started elasticsearch.compose.yaml with every container created on CPUs 2-3,6-7: gvb-elasticsearch cpuset 2-3,6-7 (read back from docker inspect)); stopped afterwards. - Engine files (SHA-256, first 12 hex digits): [elasticsearch.compose.yaml 582b9a97b8de]. This run created the engine from these files. - WARNING: the index step after the load FAILED (NO HNSW GRAPH, searches scan all 524 vectors: 1 segment(s), 524 vectors, total_vex_size_bytes 0 (_stats dense_vector), index_options hnsw m=16 ef_construction=128; Lucene builds no graph for a segment this small (measured at 1,024 dimensions: 1,042 vectors none, 1,043 vectors a graph), so default search equals exact search). Searched anyway; see the index state. - WARNING: index not ready after the load: NOT ready, 0 of 524 indexed (NO HNSW GRAPH, searches scan all 524 vectors: 1 segment(s), 524 vectors, total_vex_size_bytes 0 (_stats dense_vector), index_options hnsw m=16 ef_construction=128; Lucene builds no graph for a segment this small (measured at 1,024 dimensions: 1,042 vectors none, 1,043 vectors a graph), so default search equals exact search). Measured anyway, so its search numbers may come from a scan or a half-built index. Durability: Every acknowledged bulk request is fsynced to the translog before the answer: index.translog.durability=request, the Elasticsearch default, which neither this sink nor elasticsearch.compose.yaml overrides (ElasticsearchReadinessTests reads it back from the live index). By that setting a process crash or power loss loses no acknowledged write; this is read from the setting, not shown by pulling power. One node and no replicas, so a lost disk loses the data.. - WARNING: processes outside the benchmark used 0.53 CPUs on average over the last 2 s and 0.53 over the last few seconds when searching began (limit 0.3): the box was busy as the untimed rehearsal started. Each timed pass still waits for a quiet box before its warm-up, and one that ran busy is flagged 'busy box' on its own. The load average 2.09 3.57 3.89 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: exact 14,597 searches in 30.0 s, default@1 19,365 searches in 30.0 s, default@8 75,982 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. exact: warm-up 15.0 s and 11,295 searches (0 failed) at 1 searcher; p50 of the last windows 1.281, 1.284, 1.280 ms (windows of at least 2 s and 100 searches); trial of 2,295 searches in 3.0 s at 1 searcher: p50 1.284 ms against the settled p50 1.281 ms, 0% apart (limit 10%); the timed pass p50 1.283 ms, 0% from the settled p50 1.281 ms (limit 10% or 0.1 ms; 0.003 ms); default@1: warm-up 15.0 s and 11,109 searches (0 failed) at 1 searcher; p50 of the last windows 1.327, 1.333, 1.324 ms (windows of at least 2 s and 100 searches); trial of 2,149 searches in 3.0 s at 1 searcher: p50 1.348 ms against the settled p50 1.327 ms, 2% apart (limit 10%); the timed pass p50 1.328 ms, 0% from the settled p50 1.327 ms (limit 10% or 0.1 ms; 0.001 ms); default@8: warm-up 15.0 s and 40,094 searches (0 failed) at 8 searchers; QPS of the last windows 2,667, 2,684, 2,676 (windows of at least 2 s and 100 searches); trial of 8,019 searches in 3.0 s at 8 searchers: 2,671 QPS against the settled 2,676 QPS, 0% apart (limit 10%); the timed pass 2,669 QPS, 0% from the settled 2,676 QPS (limit 10%). - Pass exact after rehearsal: 2026-10-06T15:25:27.301Z to 2026-10-06T15:26:27.301Z (60.0 s), 45,908 searches, 0 failed, p50 1.283 ms, mean 1.305 ms, p99 1.709 ms, 765.1 QPS (1000/QPS 1.307 ms). - Pass default@1 after exact: 2026-10-06T15:26:45.342Z to 2026-10-06T15:27:05.342Z (20.0 s), 14,788 searches, 0 failed, p50 1.328 ms, mean 1.350 ms, p99 1.758 ms, 739.4 QPS (1000/QPS 1.352 ms). - Pass default@8 after default@1: 2026-10-06T15:27:23.362Z to 2026-10-06T15:27:43.364Z (20.0 s), 53,384 searches, 0 failed, p50 2.861 ms, mean 2.995 ms, p99 5.443 ms, 2668.9 QPS (1000/QPS 0.375 ms). - Passes in the order run: exact, default@1, default@8; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 184,905 sent, 0 failed (warmupErrors). Index after the searches: NOT ready, 0 of 524 indexed (NO HNSW GRAPH, searches scan all 524 vectors: 1 segment(s), 524 vectors, total_vex_size_bytes 0 (_stats dense_vector), index_options hnsw m=16 ef_construction=128; Lucene builds no graph for a segment this small (measured at 1,024 dimensions: 1,042 vectors none, 1,043 vectors a graph), so default search equals exact search). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: container address on its Docker network, not the published port (no docker-proxy): container gvb-elasticsearch (13c7378dc5ef) on network engines_default at container-ip:9200, not through docker-proxy localhost:9200. Open after the passes: container-ip:9200 x8 (this target). - CPU pinning: cpuset from the container's creation when run-all started it on the engine CPUs, else docker update --cpuset-cpus (each container's change line says which), CPUs 2-3,6-7. Changed: container gvb-elasticsearch: already on CPUs 2-3,6-7 (its cpuset since creation), not changed. Read back: container gvb-elasticsearch: 74 thread(s) on CPUs 2-3,6-7 (every thread of the container). - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): exact 3492/3492 MHz, average 3491.7/3492.4, performance, outside load 0.23 before the warm-up, 0.22 during, client CPU 0.454 ms per search; default@1 3492/3492 MHz, average 3491.1/3492.4, performance, outside load 0.22 before the warm-up, 0.20 during, client CPU 0.456 ms per search; default@8 3492/3492 MHz, average 3492.0/3492.1, performance, outside load 0.22 before the warm-up, 0.24 during, client CPU 0.488 ms per search. ### sql-diskann - Engine: Microsoft SQL Server 2025 (RTM-CU9) (KB5122048) - 17.0.5005.3 (X64), Enterprise Developer Edition (64-bit), in container gvb-mssql (image mcr.microsoft.com/mssql/server:2025-CU9-ubuntu-24.04@sha256:2b5b581621126574f3d1f75e78d3eebe8d05aedb59ad0cfdf9aa42cb0634d726; cpuset 2-3,6-7 from its creation; SQL Server counts 4 CPU(s) and runs 4 visible scheduler(s), affinity AUTO) (compose) - Index: DiskANN (preview) via VECTOR_SEARCH, cosine, build {"StartId":"306", "L":"48", "M":"8", "R":"48"}; the exact mode scans the same table - Search settings: L=48, M=8, R=48, StartId=306 - Load: 524 rows in batches of 1000, 0.9 s of upserts (611 rows/s); count matched 0.0 s after the last upsert; index step 3.9 s (copied 524 rows to dbo.gvb_gvbbench_eshoponweb_ann (INT key) and built DiskANN, parameters {"StartId":"306", "L":"48", "M":"8", "R":"48"}, graph covers 524 rows) - Search: 5267 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 1.03 ms, 8 searchers 1.11 ms - Exact mode: 15,478 searches, p50 3.75 ms, p95 4.21 ms, recall 1.000 - RAM: 529.9 MiB (docker stats); disk: 4.52 MiB (table and its indexes, reserved pages) - Started by run-all for this measurement with every container created on CPUs 2-3,6-7 (started mssql.compose.yaml with every container created on CPUs 2-3,6-7: gvb-mssql cpuset 2-3,6-7 (read back from docker inspect)); stopped afterwards. - Engine files (SHA-256, first 12 hex digits): [mssql.compose.yaml b7359663cc15]. This run created the engine from these files. - Index after the load: ready, 524 of 524 indexed (DiskANN index built and used. sys.vector_indexes: vix_gvb_gvbbench_eshoponweb on dbo.gvb_gvbbench_eshoponweb_ann in GvbBenchDiskAnn, DiskANN, COSINE, enabled, build {"StartId":"306", "L":"48", "M":"8", "R":"48"}; graph table rows 524 of 524 table rows. Real VECTOR_SEARCH plan: Vector Index Seek on index vix_gvb_gvbbench_eshoponweb of GvbBenchDiskAnn.gvb_gvbbench_eshoponweb_ann (IndexKind DiskANN), 10 rows returned, 10 hits. The exact mode searches the SAME table (dbo.gvb_gvbbench_eshoponweb_ann, database GvbBenchDiskAnn) and its real plan is: Clustered Index Scan of GvbBenchDiskAnn.gvb_gvbbench_eshoponweb_ann through PK_gvb_gvbbench_eshoponweb_ann (IndexKind Clustered), 524 rows read, no vector index operator, 10 hits). Durability: A commit returns after its transaction-log records are written to disk (SQL Server write-ahead logging; delayed durability is DISABLED); a database created here copies the model database: recovery model FULL, page_verify CHECKSUM; no global trace flags are enabled; mssql.conf of container gvb-mssql (/var/opt/mssql/mssql.conf, on the host at ~/gvb-data/engines/mssql/mssql.conf) does not exist, so it sets: nothing; it has no [control] or [traceflag] entry, so SQL Server's own Linux defaults for flushing writes apply (Microsoft's Linux performance guide names trace flag 3982 as that default; read from the guide, not tested here). Not tested by cutting power; whether the disk's own write cache reaches the media was not checked. Container settings from its environment (names only): MSSQL_AGENT_ENABLED, MSSQL_MEMORY_LIMIT_MB, MSSQL_PID, MSSQL_RPC_PORT, MSSQL_SA_PASSWORD.. - Outside load when searching began: 0.25 CPUs on average over the last 6 s and 0.26 over the last few seconds (limit 0.3), a quiet box. The load average 4.08 3.52 3.77 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: exact 7,712 searches in 30.0 s, default@8 20,023 searches in 30.0 s, default@1 7,924 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. exact: warm-up 15.0 s and 3,842 searches (0 failed) at 1 searcher; p50 of the last windows 3.748, 3.732, 3.741 ms (windows of at least 2 s and 100 searches); trial of 786 searches in 3.0 s at 1 searcher: p50 3.754 ms against the settled p50 3.741 ms, 0% apart (limit 10%); the timed pass p50 3.749 ms, 0% from the settled p50 3.741 ms (limit 10% or 0.1 ms; 0.008 ms); default@8: warm-up 15.0 s and 10,201 searches (0 failed) at 8 searchers; QPS of the last windows 672, 694, 695 (windows of at least 2 s and 100 searches); trial of 2,084 searches in 3.0 s at 8 searchers: 693 QPS against the settled 694 QPS, 0% apart (limit 10%); the timed pass 682 QPS, 2% from the settled 694 QPS (limit 10%); default@1: warm-up 15.0 s and 3,896 searches (0 failed) at 1 searcher; p50 of the last windows 3.708, 3.721, 3.696 ms (windows of at least 2 s and 100 searches); trial of 810 searches in 3.0 s at 1 searcher: p50 3.676 ms against the settled p50 3.708 ms, 1% apart (limit 10%); the timed pass p50 3.694 ms, 0% from the settled p50 3.708 ms (limit 10% or 0.1 ms; 0.014 ms). - Pass exact after rehearsal: 2026-10-06T15:29:48.840Z to 2026-10-06T15:30:48.842Z (60.0 s), 15,478 searches, 0 failed, p50 3.749 ms, mean 3.873 ms, p99 4.758 ms, 258.0 QPS (1000/QPS 3.877 ms). - Pass default@8 after exact: 2026-10-06T15:31:06.869Z to 2026-10-06T15:31:26.877Z (20.0 s), 13,652 searches, 0 failed, p50 11.438 ms, mean 11.719 ms, p99 20.594 ms, 682.3 QPS (1000/QPS 1.466 ms). - Pass default@1 after default@8: 2026-10-06T15:31:44.883Z to 2026-10-06T15:32:04.885Z (20.0 s), 5,267 searches, 0 failed, p50 3.694 ms, mean 3.795 ms, p99 5.568 ms, 263.3 QPS (1000/QPS 3.798 ms). - Passes in the order run: exact, default@8, default@1; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 57,278 sent, 0 failed (warmupErrors). Index after the searches: ready, 524 of 524 indexed (DiskANN index built and used. sys.vector_indexes: vix_gvb_gvbbench_eshoponweb on dbo.gvb_gvbbench_eshoponweb_ann in GvbBenchDiskAnn, DiskANN, COSINE, enabled, build {"StartId":"306", "L":"48", "M":"8", "R":"48"}; graph table rows 524 of 524 table rows. Real VECTOR_SEARCH plan: Vector Index Seek on index vix_gvb_gvbbench_eshoponweb of GvbBenchDiskAnn.gvb_gvbbench_eshoponweb_ann (IndexKind DiskANN), 10 rows returned, 10 hits. The exact mode searches the SAME table (dbo.gvb_gvbbench_eshoponweb_ann, database GvbBenchDiskAnn) and its real plan is: Clustered Index Scan of GvbBenchDiskAnn.gvb_gvbbench_eshoponweb_ann through PK_gvb_gvbbench_eshoponweb_ann (IndexKind Clustered), 524 rows read, no vector index operator, 10 hits). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: container address on its Docker network, not the published port (no docker-proxy): container gvb-mssql (b41cee3eb1d8) on network gvb-mssql_default at container-ip:1433, not through docker-proxy localhost:14330. Open after the passes: container-ip:1433 x9 (this target). - CPU pinning: cpuset from the container's creation when run-all started it on the engine CPUs, else docker update --cpuset-cpus (each container's change line says which), CPUs 2-3,6-7. Changed: container gvb-mssql: already on CPUs 2-3,6-7 (its cpuset since creation), not changed. Read back: container gvb-mssql: 85 thread(s) on CPUs 2-3,6-7 (every thread of the container). - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): exact 3492/3492 MHz, average 3492.4/3492.5, performance, outside load 0.19 before the warm-up, 0.22 during, client CPU 1.165 ms per search; default@8 3492/3492 MHz, average 3492.0/3492.3, performance, outside load 0.22 before the warm-up, 0.20 during, client CPU 1.108 ms per search; default@1 3492/3492 MHz, average 3492.5/3492.4, performance, outside load 0.20 before the warm-up, 0.25 during, client CPU 1.031 ms per search. ### chroma - Engine: Chroma 1.4.4 (single node, REST API v2) (compose) - Index: HNSW M=16 ef_construction=128, ef_search=100 (Chroma default), cosine; approximate only (no exact mode) - Search settings: M=16, ef_construction=128, ef_search=100 - Load: 524 rows in batches of 1000, 0.6 s of upserts (820 rows/s); count matched 0.0 s after the last upsert; index step 0.0 s (Chroma has no index build to wait for and exposes no index status; checked in 0.0 s: Chroma exposes no index status (indexing_status endpoint: HTTP 500 {"error":"InternalError","message":"Method scout_logs is not implemented"}), so indexed vectors are not reported and ready only means the two counts agree to within 1 percent: count endpoint 524, a vector search asking for 524 results returned 524) - Search: 9379 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 0.47 ms, 8 searchers 0.47 ms - RAM: 44.93 MiB (docker stats); disk: 414.16 KiB added by this load (2.01 GiB (whole engine data folder), 2.01 GiB before) - Started by run-all for this measurement with every container created on CPUs 2-3,6-7 (started chroma.compose.yaml with every container created on CPUs 2-3,6-7: gvb-chroma cpuset 2-3,6-7 (read back from docker inspect)); stopped afterwards. - Engine files (SHA-256, first 12 hex digits): [chroma.compose.yaml f7ed9d3c8ff7]. This run created the engine from these files. - Index after the load: ready, ? of 524 indexed (Chroma exposes no index status (indexing_status endpoint: HTTP 500 {"error":"InternalError","message":"Method scout_logs is not implemented"}), so indexed vectors are not reported and ready only means the two counts agree to within 1 percent: count endpoint 524, a vector search asking for 524 results returned 524). Durability: SQLite rollback journal with its default synchronous=FULL under the data directory (chroma.compose.yaml sets IS_PERSISTENT=1 and PERSIST_DIRECTORY=/data and no sync setting): every write is committed to chroma.sqlite3 with fsync of the journal, the directory and the database file before the call returns (strace: 15 database fsyncs and 38 journal fsyncs for one create, four 500-vector upserts and one delete), and the HNSW files are written every sync_threshold=1000 vectors and rebuilt from the SQLite log after a crash; measured with kill -9 and a restart: all 50, 300 and 2,000 vectors were still stored and searchable. - Outside load when searching began: 0.25 CPUs on average over the last 1 s and 0.25 over the last few seconds (limit 0.3), a quiet box. The load average 2.72 3.13 3.54 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: default@1 14,115 searches in 30.0 s, default@8 23,300 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. default@1: warm-up 15.0 s and 7,041 searches (0 failed) at 1 searcher; p50 of the last windows 2.100, 2.106, 2.107 ms (windows of at least 2 s and 100 searches); trial of 1,405 searches in 3.0 s at 1 searcher: p50 2.109 ms against the settled p50 2.106 ms, 0% apart (limit 10%); the timed pass p50 2.105 ms, 0% from the settled p50 2.106 ms (limit 10% or 0.1 ms; 0.001 ms); default@8: warm-up 15.0 s and 11,647 searches (0 failed) at 8 searchers; QPS of the last windows 780, 772, 778 (windows of at least 2 s and 100 searches); trial of 2,335 searches in 3.0 s at 8 searchers: 776 QPS against the settled 778 QPS, 0% apart (limit 10%); the timed pass 777 QPS, 0% from the settled 778 QPS (limit 10%). - Pass default@1 after rehearsal: 2026-10-06T15:33:31.761Z to 2026-10-06T15:33:51.761Z (20.0 s), 9,379 searches, 0 failed, p50 2.105 ms, mean 2.131 ms, p99 2.628 ms, 468.9 QPS (1000/QPS 2.133 ms). - Pass default@8 after default@1: 2026-10-06T15:34:09.785Z to 2026-10-06T15:34:29.793Z (20.0 s), 15,551 searches, 0 failed, p50 10.217 ms, mean 10.289 ms, p99 12.378 ms, 777.2 QPS (1000/QPS 1.287 ms). - Passes in the order run: default@1, default@8; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 59,843 sent, 0 failed (warmupErrors). Index after the searches: ready, ? of 524 indexed (Chroma exposes no index status (indexing_status endpoint: HTTP 500 {"error":"InternalError","message":"Method scout_logs is not implemented"}), so indexed vectors are not reported and ready only means the two counts agree to within 1 percent: count endpoint 524, a vector search asking for 524 results returned 524). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: container address on its Docker network, not the published port (no docker-proxy): container gvb-chroma (653fefb9c7e5) on network gvb-chroma_default at container-ip:8000, not through docker-proxy localhost:8000. Open after the passes: container-ip:8000 x8 (this target). - CPU pinning: cpuset from the container's creation when run-all started it on the engine CPUs, else docker update --cpuset-cpus (each container's change line says which), CPUs 2-3,6-7. Changed: container gvb-chroma: already on CPUs 2-3,6-7 (its cpuset since creation), not changed. Read back: container gvb-chroma: 9 thread(s) on CPUs 2-3,6-7 (every thread of the container). - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): default@1 3492/3492 MHz, average 3492.0/3492.5, performance, outside load 0.17 before the warm-up, 0.15 during, client CPU 0.466 ms per search; default@8 3492/3492 MHz, average 3492.0/3492.3, performance, outside load 0.16 before the warm-up, 0.14 during, client CPU 0.469 ms per search. ### typesense - Engine: Typesense 30.2 (vector search) (compose) - Index: HNSW float32 (hnswlib), m=16, ef_construction=128, cosine; search k=top, ef=100; exact mode = filter ordinal:>=0 with flat_search_cutoff; index held in memory - Search settings: ef=100, ef_construction=128, k=top, m=16 - Load: 524 rows in batches of 1000, 0.5 s of upserts (991 rows/s); count matched 0.0 s after the last upsert; index step 0.0 s (nothing to build, Typesense inserts into the HNSW graph while writing; ready after 0.0 s: HNSW graph returns 524 of 524 vectors, no writes queued: num_documents 524, pending_write_batches 0, health ok, unfiltered vector query with k above the count returned 524; the graph is built in memory during each write, so there is no later build step) - Search: 4691 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 0.53 ms, 8 searchers 0.59 ms - Exact mode: 10,428 searches, p50 5.68 ms, p95 6.05 ms, recall 1.000 - RAM: 117.5 MiB (docker stats); disk: 0 B added by this load (272.46 MiB (whole engine data folder), 280.58 MiB before) - Started by run-all for this measurement with every container created on CPUs 2-3,6-7 (started typesense.compose.yaml with every container created on CPUs 2-3,6-7: gvb-typesense cpuset 2-3,6-7 (read back from docker inspect)); stopped afterwards. - Engine files (SHA-256, first 12 hex digits): [typesense.compose.yaml 96a165f62566]. This run created the engine from these files. - Index after the load: ready, 524 of 524 indexed (HNSW graph returns 524 of 524 vectors, no writes queued: num_documents 524, pending_write_batches 0, health ok, unfiltered vector query with k above the count returned 524; the graph is built in memory during each write, so there is no later build step). Durability: Every acknowledged write is appended to Typesense's raft log and fsynced before the answer: braft raft_sync=true with raft_sync_policy=0 (sync immediately), read from the running container's brpc /flags page on 2026-10-04. A restart replays the log from the last snapshot (typesense.compose.yaml sets TYPESENSE_SNAPSHOT_INTERVAL_SECONDS=300) and rebuilds the in-memory HNSW graph, so by those settings a crash loses no acknowledged write, but the restart is slow after a big load. This is read from the settings, not shown by pulling power. One node and no replicas, so a lost disk loses the data.. - Outside load when searching began: 0.24 CPUs on average over the last 6 s and 0.25 over the last few seconds (limit 0.3), a quiet box. The load average 2.80 2.84 3.36 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: default@8 17,271 searches in 30.0 s, exact 5,225 searches in 30.0 s, default@1 7,031 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. default@8: warm-up 15.0 s and 8,595 searches (0 failed) at 8 searchers; QPS of the last windows 572, 576, 574 (windows of at least 2 s and 100 searches); trial of 1,728 searches in 3.0 s at 8 searchers: 574 QPS against the settled 574 QPS, 0% apart (limit 10%); the timed pass 574 QPS, 0% from the settled 574 QPS (limit 10%); exact: warm-up 15.0 s and 2,617 searches (0 failed) at 1 searcher; p50 of the last windows 5.692, 5.686, 5.679 ms (windows of at least 2 s and 100 searches); trial of 521 searches in 3.0 s at 1 searcher: p50 5.685 ms against the settled p50 5.686 ms, 0% apart (limit 10%); the timed pass p50 5.682 ms, 0% from the settled p50 5.686 ms (limit 10% or 0.1 ms; 0.004 ms); default@1: warm-up 15.0 s and 3,515 searches (0 failed) at 1 searcher; p50 of the last windows 4.225, 4.211, 4.201 ms (windows of at least 2 s and 100 searches); trial of 707 searches in 3.0 s at 1 searcher: p50 4.197 ms against the settled p50 4.211 ms, 0% apart (limit 10%); the timed pass p50 4.203 ms, 0% from the settled p50 4.211 ms (limit 10% or 0.1 ms; 0.008 ms). - Pass default@8 after rehearsal: 2026-10-06T15:36:31.300Z to 2026-10-06T15:36:51.310Z (20.0 s), 11,486 searches, 0 failed, p50 13.387 ms, mean 13.932 ms, p99 26.073 ms, 574.0 QPS (1000/QPS 1.742 ms). - Pass exact after default@8: 2026-10-06T15:37:09.318Z to 2026-10-06T15:38:09.319Z (60.0 s), 10,428 searches, 0 failed, p50 5.682 ms, mean 5.752 ms, p99 7.490 ms, 173.8 QPS (1000/QPS 5.754 ms). - Pass default@1 after exact: 2026-10-06T15:38:27.331Z to 2026-10-06T15:38:47.332Z (20.0 s), 4,691 searches, 0 failed, p50 4.203 ms, mean 4.262 ms, p99 5.085 ms, 234.5 QPS (1000/QPS 4.264 ms). - Passes in the order run: default@8, exact, default@1; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 47,210 sent, 0 failed (warmupErrors). Index after the searches: ready, 524 of 524 indexed (HNSW graph returns 524 of 524 vectors, no writes queued: num_documents 524, pending_write_batches 0, health ok, unfiltered vector query with k above the count returned 524; the graph is built in memory during each write, so there is no later build step). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: container address on its Docker network, not the published port (no docker-proxy): container gvb-typesense (0942e4b1b27f) on network engines_default at container-ip:8108, not through docker-proxy localhost:8108. Open after the passes: container-ip:8108 x1 (this target). - CPU pinning: cpuset from the container's creation when run-all started it on the engine CPUs, else docker update --cpuset-cpus (each container's change line says which), CPUs 2-3,6-7. Changed: container gvb-typesense: already on CPUs 2-3,6-7 (its cpuset since creation), not changed. Read back: container gvb-typesense: 303 thread(s) on CPUs 2-3,6-7 (every thread of the container). - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): default@8 3492/3492 MHz, average 3492.0/3492.5, performance, outside load 0.21 before the warm-up, 0.07 during, client CPU 0.591 ms per search; exact 3492/3492 MHz, average 3492.4/3492.9, performance, outside load 0.11 before the warm-up, 0.21 during, client CPU 0.544 ms per search; default@1 3492/3492 MHz, average 3492.3/3492.7, performance, outside load 0.21 before the warm-up, 0.21 during, client CPU 0.528 ms per search. ### mongodb - Engine: MongoDB 8.0.32 Atlas Local (mongod + mongot Vector Search) (compose) - Index: vectorSearch index, HNSW maxEdges=16 numEdgeCandidates=128, float32 binData, cosine, numCandidates=20x hits (min 100); exact mode = $vectorSearch exact:true - Search settings: maxEdges=16, numCandidates=20x, numEdgeCandidates=128 - Load: 524 rows in batches of 1000, 0.2 s of upserts (2,896 rows/s); count matched 1.2 s after the last upsert; index step 10.2 s (index status READY, queryable True; mongot holds 524 of 524 documents in 2 segment(s) (278 + 246 documents), 2 searched through the HNSW graph (Approximate); segment layout unchanged for 10.2 s over 11 readings; wait took 10.2 s) - Search: 13868 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 0.31 ms, 8 searchers 0.28 ms - Exact mode: 46,307 searches, p50 1.26 ms, p95 1.50 ms, recall 1.000 - RAM: 646.8 MiB (docker stats); disk: 148.12 KiB added by this load (1.37 GiB (whole engine data folder), 1.37 GiB before) - Started by run-all for this measurement with every container created on CPUs 2-3,6-7 (started mongodb.compose.yaml with every container created on CPUs 2-3,6-7: gvb-mongodb cpuset 2-3,6-7 (read back from docker inspect)); stopped afterwards. - Engine files (SHA-256, first 12 hex digits): [mongodb.compose.yaml 4ee6f6411f1f]. This run created the engine from these files. - Index after the load: ready, 524 of 524 indexed (index status READY, queryable True; mongot holds 524 of 524 documents in 2 segment(s) (278 + 246 documents), 2 searched through the HNSW graph (Approximate)). Durability: Acknowledged writes are journaled before the acknowledgement. The sink sets no write concern, so the server default applies: getDefaultRWConcern gives w majority and the one-member replica set (--replSet gvbmongo in the image) has writeConcernMajorityJournalDefault true, with WiredTiger journaling on (journalCommitInterval 100 ms is only the interval for unacknowledged work). Measured 2026-10-04: 30 acknowledged single-document inserts raised WiredTiger 'log sync operations' by 30 and strace showed fdatasync on /data/db/journal/WiredTigerLog files. A crash loses no acknowledged write. mongot's search index is not part of that promise: it follows the collection asynchronously and is rebuilt from it.. - Outside load when searching began: 0.23 CPUs on average over the last 13 s and 0.22 over the last few seconds (limit 0.3), a quiet box. The load average 1.46 2.74 3.26 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: default@1 17,380 searches in 30.0 s, exact 22,428 searches in 30.0 s, default@8 59,788 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. default@1: warm-up 15.0 s and 10,396 searches (0 failed) at 1 searcher; p50 of the last windows 1.407, 1.409, 1.396 ms (windows of at least 2 s and 100 searches); trial of 2,058 searches in 3.0 s at 1 searcher: p50 1.416 ms against the settled p50 1.407 ms, 1% apart (limit 10%); the timed pass p50 1.408 ms, 0% from the settled p50 1.407 ms (limit 10% or 0.1 ms; 0.000 ms); exact: warm-up 15.0 s and 11,607 searches (0 failed) at 1 searcher; p50 of the last windows 1.264, 1.270, 1.265 ms (windows of at least 2 s and 100 searches); trial of 2,318 searches in 3.0 s at 1 searcher: p50 1.265 ms against the settled p50 1.265 ms, 0% apart (limit 10%); the timed pass p50 1.263 ms, 0% from the settled p50 1.265 ms (limit 10% or 0.1 ms; 0.002 ms); default@8: warm-up 15.0 s and 31,530 searches (0 failed) at 8 searchers; QPS of the last windows 2,124, 2,079, 2,134 (windows of at least 2 s and 100 searches); trial of 6,402 searches in 3.0 s at 8 searchers: 2,132 QPS against the settled 2,124 QPS, 0% apart (limit 10%); the timed pass 2,115 QPS, 0% from the settled 2,124 QPS (limit 10%). - Pass default@1 after rehearsal: 2026-10-06T15:40:57.585Z to 2026-10-06T15:41:17.586Z (20.0 s), 13,868 searches, 0 failed, p50 1.408 ms, mean 1.440 ms, p99 1.987 ms, 693.4 QPS (1000/QPS 1.442 ms). - Pass exact after default@1: 2026-10-06T15:41:35.602Z to 2026-10-06T15:42:35.602Z (60.0 s), 46,307 searches, 0 failed, p50 1.263 ms, mean 1.293 ms, p99 1.826 ms, 771.8 QPS (1000/QPS 1.296 ms). - Pass default@8 after exact: 2026-10-06T15:43:04.652Z to 2026-10-06T15:43:24.656Z (20.0 s), 42,314 searches, 0 failed, p50 3.655 ms, mean 3.779 ms, p99 6.709 ms, 2115.3 QPS (1000/QPS 0.473 ms). - Passes in the order run: default@1, exact, default@8; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 163,907 sent, 0 failed (warmupErrors). Index after the searches: ready, 524 of 524 indexed (index status READY, queryable True; mongot holds 524 of 524 documents in 2 segment(s) (278 + 246 documents), 2 searched through the HNSW graph (Approximate)). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: container address on its Docker network, not the published port (no docker-proxy): container gvb-mongodb (7fdbe1aa067f) on network gvb-mongodb_default at container-ip:27017, not through docker-proxy localhost:27017. Open after the passes: container-ip:27017 x10 (this target). - CPU pinning: cpuset from the container's creation when run-all started it on the engine CPUs, else docker update --cpuset-cpus (each container's change line says which), CPUs 2-3,6-7. Changed: container gvb-mongodb: already on CPUs 2-3,6-7 (its cpuset since creation), not changed. Read back: container gvb-mongodb: 211 thread(s) on CPUs 2-3,6-7 (every thread of the container). - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): default@1 3492/3492 MHz, average 3492.2/3492.4, performance, outside load 0.27 before the warm-up, 0.30 during, client CPU 0.311 ms per search; exact 3492/3492 MHz, average 3492.1/3492.2, performance, outside load 0.28 before the warm-up, 0.32 during, BUSY BOX, client CPU 0.305 ms per search; default@8 3492/3492 MHz, average 3492.0/3492.1, performance, outside load 0.30 before the warm-up after waiting 11 s, 0.24 during, client CPU 0.284 ms per search. - WARNING: busy box during exact: processes outside the benchmark used 0.32 CPUs on average during the pass, above the limit of 0.3. ### sqlitevec - Engine: SQLite 3.53.3 + sqlite-vec 0.1.7-alpha.2.1 (vec0, embedded, in process) (embedded) - Index: vec0 brute-force scan, no ANN index (exact), float32, cosine distance, default chunk_size=1024; score = 1 - cosine distance; searches run concurrently, one WAL reader connection per searcher (opened as searchers arrive, at most 32), writes run one at a time and may overlap searches - Search settings: chunk_size=1024 - Load: 524 rows in batches of 1000, 0.1 s of upserts (4,338 rows/s); count matched 0.0 s after the last upsert; index step 0.0 s (nothing to build, every search scans all vectors; checked in 0.00 s: no index, exact scan by design: gvb_gvbbench_eshoponweb is a vec0 virtual table holding 524 vectors and every search compares all of them (EXPLAIN QUERY PLAN: SCAN gvb_gvbbench_eshoponweb VIRTUAL TABLE INDEX 0:3{___}___ | USE TEMP B-TREE FOR ORDER BY)) - Search: 10286 latency samples, target held 524 rows, 0 errors - Client CPU per search: 1 searcher 2.04 ms, 8 searchers 3.51 ms - Exact mode: 30,746 searches, p50 1.91 ms, p95 2.16 ms, recall 1.000 - RAM: in this process, not measured; disk: 9.72 MiB added by this load (186.34 MiB (engine data folder), 176.61 MiB before) - Index after the load: ready, ? of 524 indexed (no index, exact scan by design: gvb_gvbbench_eshoponweb is a vec0 virtual table holding 524 vectors and every search compares all of them (EXPLAIN QUERY PLAN: SCAN gvb_gvbbench_eshoponweb VIRTUAL TABLE INDEX 0:3{___}___ | USE TEMP B-TREE FOR ORDER BY)). Durability: PRAGMA journal_mode=WAL and PRAGMA synchronous=NORMAL (set by this sink when it opens the file): each commit is appended to the -wal file and handed to the operating system, but the WAL is fsynced only when SQLite checkpoints it (measured with strace: 40 single-record commits caused 3 WAL fsyncs), so a crash of the process loses nothing (measured with kill -9 and a reopen: all 50, 300 and 2,000 rows were there), while an operating-system crash or power cut can lose the newest commits since the last checkpoint and leaves the database consistent. - Outside load when searching began: not known yet (too little sampled history). Load average 2.30 2.86 3.17 (1/5/15 min; it also counts this benchmark's own client and engines, so it is recorded, not judged). - Rehearsal before any timed pass, untimed, every pass type at its own concurrency for at least 30 s through the same code the timed passes use: default@1 15,462 searches in 30.0 s, exact 15,486 searches in 30.0 s, default@8 34,082 searches in 30.0 s. - Settle: settled before every timed pass, each by its own warm-up at its own concurrency and a trial of the same pass within 10% of the warm-up's settled figure, and the timed pass itself within 10% of it. default@1: warm-up 15.0 s and 7,704 searches (0 failed) at 1 searcher; p50 of the last windows 1.890, 1.903, 1.892 ms (windows of at least 2 s and 100 searches); trial of 1,559 searches in 3.0 s at 1 searcher: p50 1.891 ms against the settled p50 1.892 ms, 0% apart (limit 10%); the timed pass p50 1.910 ms, 1% from the settled p50 1.892 ms (limit 10% or 0.1 ms; 0.018 ms); exact: warm-up 15.0 s and 7,701 searches (0 failed) at 1 searcher; p50 of the last windows 1.912, 1.908, 1.904 ms (windows of at least 2 s and 100 searches); trial of 1,541 searches in 3.0 s at 1 searcher: p50 1.913 ms against the settled p50 1.908 ms, 0% apart (limit 10%); the timed pass p50 1.914 ms, 0% from the settled p50 1.908 ms (limit 10% or 0.1 ms; 0.006 ms); default@8: warm-up 15.0 s and 17,001 searches (0 failed) at 8 searchers; QPS of the last windows 1,134, 1,131, 1,138 (windows of at least 2 s and 100 searches); trial of 3,406 searches in 3.0 s at 8 searchers: 1,134 QPS against the settled 1,134 QPS, 0% apart (limit 10%); the timed pass 1,136 QPS, 0% from the settled 1,134 QPS (limit 10%). - Pass default@1 after rehearsal: 2026-10-06T15:46:14.463Z to 2026-10-06T15:46:34.464Z (20.0 s), 10,286 searches, 0 failed, p50 1.910 ms, mean 1.942 ms, p99 2.357 ms, 514.3 QPS (1000/QPS 1.944 ms). - Pass exact after default@1: 2026-10-06T15:46:52.477Z to 2026-10-06T15:47:52.479Z (60.0 s), 30,746 searches, 0 failed, p50 1.914 ms, mean 1.949 ms, p99 2.366 ms, 512.4 QPS (1000/QPS 1.952 ms). - Pass default@8 after exact: 2026-10-06T15:48:12.513Z to 2026-10-06T15:48:32.516Z (20.0 s), 22,732 searches, 0 failed, p50 6.773 ms, mean 7.029 ms, p99 13.588 ms, 1136.4 QPS (1000/QPS 0.880 ms). - Passes in the order run: default@1, exact, default@8; each after its own warm-up. Untimed searches (rehearsals, per-pass warm-ups, settle checks and extensions): 103,942 sent, 0 failed (warmupErrors). Index after the searches: ready, ? of 524 indexed (no index, exact scan by design: gvb_gvbbench_eshoponweb is a vec0 virtual table holding 524 vectors and every search compares all of them (EXPLAIN QUERY PLAN: SCAN gvb_gvbbench_eshoponweb VIRTUAL TABLE INDEX 0:3{___}___ | USE TEMP B-TREE FOR ORDER BY)). - Benchmark copy gvbbench_eshoponweb dropped afterwards. - Connection: in this process (embedded), no network: no network address. No open connection seen (no open TCP connection was found). - CPU pinning: none: an embedded engine runs inside the benchmark client process, so it shares the client CPUs, CPUs 0-1,4-5. - Clock per pass (median MHz of engine CPUs / client CPUs, their average MHz, governor, CPUs busy outside the benchmark, client CPU per search; the clock was pinned at 3500 MHz (turbo off), and a pass whose average on either group is more than 1% off it is flagged): default@1 3492/3492 MHz, average 3493.8/3493.5, performance, outside load 0.23 before the warm-up, 0.21 during, client CPU 2.038 ms per search; exact 3492/3492 MHz, average 3493.6/3493.2, performance, outside load 0.22 before the warm-up, 0.23 during, client CPU 2.045 ms per search; default@8 3492/3492 MHz, average 3492.3/3492.0, performance, outside load 0.23 before the warm-up after waiting 2 s, 0.23 during, client CPU 3.512 ms per search. ## Notes - Order: targets ran one at a time in a random order from seed 702 (runSeed; --seed 702 repeats it, targetOrder lists it). Inside each target the timed passes also ran in a random order from the same seed and the target's name (passOrder): default@1 is one searcher for 20 s (every search's latency gives p50/p95/p99, the completed searches give QPS@1, its first answer to each query gives recall and nDCG), default@N is N searchers for 20 s (QPS@N), exact is the engine's exact mode, one searcher for 60 s cycling the queries. - Preparation, untimed, before any timed pass: a rehearsal of every pass type at its own concurrency for 30 s each, through the same code the passes use. - Warm-up and settle check, untimed, right before every timed pass: the pass's own search at the pass's own number of searchers for at least 15 s and at least 20 searches (at most 120 s), read in windows of at least 2 s and 100 searches; then a 3 s trial of the same pass. The trial's figure (p50 with one searcher, QPS with several) must lie within 10% of the warm-up's settled figure (the median of its last 3 windows, which must agree within 5%); if not, the warm-up is extended once (at least 30 s, until its windows agree, at most 120 s), where an extension's windows agree when its level test passes: the older and the newer half of its latest windows, 5 to 10 windows a half, each half read as the pass reads it (the p50 of all its searches with one searcher, its searches per second with several), agree within 5%, judged on the windows that stopped it (an extension whose cap runs out with fewer than 10 windows is judged by its last 3 instead); then a second trial is taken, which must lie inside the range of the newer half's windows or within 10% of the newer half's figure, and the warm-up runs again before the pass. After the pass its own figure is held against the settled figure, within 10%. In the level test, the second trial and the hold, two one-searcher p50s no more than 0.1 ms apart agree whatever their percentage (below what this box resolves at one searcher). Each target's notes give every check, and a pass that still disagrees, or whose own figure did not hold, flags its target as unsettled. With machine control on, the check for a quiet box is made when the warm-up is announced, before it starts (a wait between the warm-up and the timed pass let the engine go cold), so by the time the clock opens that check is as old as the warm-up and the trial (about 18 s, more after an extension); each pass's conditions record that lead (quietCheckLeadSeconds). Failed untimed searches are counted per target (warmupErrors) and are not in the timed error counts. - Load rows/s counts only time inside each target's upsert calls: one writer, batches of 1000, rows already in memory, the collection dropped and created fresh first. Engines that build or finish their index after the writes do it in a separate timed index step (the load's index seconds), and the run waits for it before searching. - Index proof: each engine's own report of its index (indexState) is read after the load and again after the last pass. A target whose index was not ready after the load was still measured and carries a WARNING; an engine that reports nothing counts as not ready. - Search settings: each searched target records the settings its own index description states (searchSettings: the build parameters and the effort per query, such as m, ef_construction and ef_search), read after the load; consolidating never averages runs whose settings differ. - Durability: each target's crash-safety setting as configured here (durability). Engines that do not force writes to disk on every commit load faster for that reason. - Latency is client-side wall time around each search (network and driver included), every search of the default@1 window, one searcher, the queries cycled; a window with fewer than 200 searches, a p50 above 1.25 x its mean, or a mean above its p99 is flagged. - QPS: N searchers back to back for 20 s per level; completed searches divided by the window's elapsed time. - Recall@10: share of the exact top 10 (brute force in memory) that the engine returned. A hit whose exact similarity ties the 10th best (within 1e-5) also counts, because duplicate rows embed to identical vectors. - Routes: a container engine (the benchmark's own SQL Server and Qdrant containers included) is reached at its container's own address on its Docker network, never through the published localhost port (docker-proxy); the native comparison targets (sql-native, qdrant-native) are native services reached directly over loopback. Each target's addresses and the connections the client held open after its passes are in its notes and in conditions.connections. - RAM of compose engines, the benchmark's own SQL Server and Qdrant containers (sql, sql-diskann, qdrant, qdrant-hnsw) included, is docker stats of the engine's containers; for the native comparison targets (sql-native, qdrant-native) it is the whole native process, including every other database or collection it serves. Disk is the table's reserved pages (SQL), the collection folder (Qdrant), or for other engines what the load added to the engine's data folder (engines that keep data in memory until a snapshot show almost nothing). - Engines: run-all starts an engine that is down and stops it afterwards only if it was not running when the run began; an engine that was already running is left running. - WARNING: the engine did not report a ready index after the load for: elasticsearch. They were measured anyway; their numbers may come from a scan or a half-built index (see each target's index state). - Machine control on: governor performance on every CPU during the run (before: schedutil; at the end: performance; after putting it back: schedutil). CPU clock pinned for the run: turbo off (intel_pstate/no_turbo 0 -> 1), so every CPU's clock is held at its ceiling of 3500 MHz whatever the engine runs; the uncore (L3 and memory) clock is held at 3000 MHz (min 3000 MHz, max 3000 MHz (MSR 0x620 = 0x1e1e); was min 1200 MHz, max 3000 MHz (MSR 0x620 = 0xc1e)) (before: turbo on (intel_pstate/no_turbo 0), ceiling 3600 MHz on CPUs 0-7, floor 1200 MHz on CPUs 0-7; while pinned: turbo off (intel_pstate/no_turbo 1), ceiling 3500 MHz on CPUs 0-7, floor 1200 MHz on CPUs 0-7; at the end: turbo off (intel_pstate/no_turbo 1), ceiling 3500 MHz on CPUs 0-7, floor 1200 MHz on CPUs 0-7; after putting it back: turbo on (intel_pstate/no_turbo 0), ceiling 3600 MHz on CPUs 0-7, floor 1200 MHz on CPUs 0-7). Uncore limit at the end: min 3000 MHz, max 3000 MHz (MSR 0x620 = 0x1e1e); after putting it back: min 1200 MHz, max 3000 MHz (MSR 0x620 = 0xc1e). Each target's clock note gives every pass's average MHz on the engine CPUs and on the client CPUs; a pass whose average on either is more than 1% off the pinned 3500 MHz is flagged. Figures from runs made with turbo on are not comparable with these in absolute terms. engine CPUs 2-3,6-7 (cores 2,6 and 3,7), client CPUs 0-1,4-5 (cores 0,4 and 1,5); the client process was pinned; each engine was pinned to the engine CPUs for its turn and put back after (conditions.engines); an engine run-all started (and stopped) was asked to be created on them, and its notes say whether the host did so or it was moved there after the start. Busy box: right before each pass's warm-up the run waits, up to 10 min, while processes outside the benchmark (everything but this client and the engine under test's cgroups) use more than 0.3 CPUs on average over the last 60 s or the last 5 s, neither window reaching back past the start of the target or of its engine (so the engine's own start-up is not outside work); if the box does not clear the pass runs anyway, flagged 'busy box', as is a pass whose own outside load is above the limit. Outside load counts busy = user + nice + system + irq + softirq + steal (guest time is already in user); CONFIG_IRQ_TIME_ACCOUNTING is not set, so task and cgroup run time include the interrupt and softirq time that hit them and it is added; CONFIG_PARAVIRT_TIME_ACCOUNTING is not set, so steal is added (/boot/config-6.8.0-142-generic) (conditions.cpuAccounting); each pass also records this client's own CPU time per search (conditions.passes[].clientCpuMsPerSearch). CPU clocks were sampled every 250 ms; each pass's min/median/max per CPU is in conditions.passes. CPU idle states, recorded and left as found: driver intel_idle, governor menu, intel_idle max_cstate 9; POLL on, C1 on, C1E on, C3 OFF on CPUs 0-7 (default disabled), C6 on (conditions.cpuIdle). How each target was reached: conditions.connections. Client build Release, .NET 10.0.12. - duckdb is embedded: it ran inside the client process on the client CPUs 0-1,4-5, sharing them with the client. - sqlitevec is embedded: it ran inside the client process on the client CPUs 0-1,4-5, sharing them with the client. - WARNING: busy box during mongodb exact: processes outside the benchmark used 0.32 CPUs on average during the pass, above the limit of 0.3. - WARNING: clock off its pinned value during oracle default@8: the engine CPUs averaged 3129 MHz, -10.6% against the pinned 3500 MHz; the client CPUs averaged 3230 MHz, -7.7% against the pinned 3500 MHz (limit 1%); this pass is not comparable with passes at the pinned clock.