Benchmarks / Elasticsearch (Rally)
Elasticsearch's own benchmark, pointed at WarpSearch.
Rally is the tool Elastic uses to benchmark Elasticsearch every night. We run five of its official tracks against WarpSearch's Elasticsearch-compatible API and set the results next to Elastic's published nightly numbers. No client changes, no special tuning: stock esrally and the standard Elasticsearch API.
Head to head
Faster where it counts
Median latency, shorter is better, except indexing, where longer is better. Elasticsearch figures are Elastic's own nightly results, single node, 30-day median.
documents per second, 11.2M Stack Overflow questions with nested answers
Scoreboard
Track by track
| Track | What it exercises | Documents | Result |
|---|---|---|---|
| nested | Nested documents, nested queries, inner hits, nested aggregations | 11.2M | Faster than Elasticsearch on every nested task at p50, p90 and p99; date histogram 6.8x, indexing 2.7x |
| geopoint | Polygon, bounding box and distance queries, geo grids, ES|QL | 60.8M | Filters in 1 to 2 ms; up to 80x faster than Elasticsearch |
| geonames | Structured search, filters, sorts, terms aggregations, scroll | 11.4M | Term, phrase and sorts in 1 ms; aggregations 2.4x faster |
| pmc | Long scientific articles, full-text, date histograms, scroll | 574K | Term, phrase, date histogram and scroll faster than Elasticsearch |
| geopointshape | Shape queries, grid aggregations, ES|QL spatial | 60.8M | 60.8M shapes indexed with 0 errors; polygon, bounding box and ES|QL spatial queries |
| percolator | 2M stored queries matched against documents | 2M | 2.7x to 5.9x faster than Elasticsearch, on WarpSearch's percolate API |
Next: http_logs, nyc_taxis, big5 and the vector tracks (so_vector, openai_vector).
Nested documents
nested: every nested task faster
11.2M Stack Overflow questions, each with its answers as nested documents. Rally service time, p50 / p90 / p99.
| Task | WarpSearch | Elasticsearch | At p50 |
|---|---|---|---|
| Nested date histogram | 145.7 / 155.0 / 159.0 ms | 984 / 1,037 / 1,089 ms | 6.8x faster |
| Randomized nested queries | 9.4 / 15.9 / 22.1 ms | 26.9 / 40.4 / 49.7 ms | 2.9x faster |
| Nested, inner hits | 12.4 / 19.5 / 26.4 ms | 26.4 / 39.8 / 48.1 ms | 2.1x faster |
| Sorted term queries | 23.1 / 41.6 / 49.6 ms | 43.4 / 86.7 / 102.8 ms | 1.9x faster |
| Nested, inner hits, large size | 18.9 / 25.2 / 29.6 ms | 34.5 / 47.4 / 57.2 ms | 1.8x faster |
| Bulk indexing | 164,000 docs/s | 61,630 docs/s | 2.7x faster |
WarpSearch: Serverless, Elasticsearch API, runs 2026-11-15 to 2026-11-17, 0 errors; indexing is 11.2M documents in 68 seconds. Elasticsearch: Elastic nightly benchmarks, track nested, single node, 30-day median.
Geo
geopoint: 60.8M points, answered in milliseconds
OpenStreetMap points with polygon, bounding box, distance and grid queries. WarpSearch server time against Elasticsearch's nightly latency, p50.
| Task | WarpSearch | Elasticsearch | At p50 |
|---|---|---|---|
| Distance range | 28 ms | 2,254 ms | 80x faster |
| Bounding box | 1 ms | 43.4 ms | 43x faster |
| Geohex grid | 1 ms | 27.3 ms | 27x faster |
| Distance filter | 1 ms | 23.9 ms | 24x faster |
| Polygon intersects count, ES|QL | 1 ms | 20.5 ms | 21x faster |
| Polygon intersects | 1 ms | 12.8 ms | 13x faster |
| Polygon within | 1 ms | 11.3 ms | 11x faster |
| Polygon intersects, ES|QL | 1 ms | 11.1 ms | 11x faster |
| Polygon disjoint | 2 ms | 12.1 ms | 6x faster |
| Geotile grid | 1 ms | 5.8 ms | 5.8x faster |
| Geohash grid aggregation | 2 ms | 7.2 ms | 3.6x faster |
| Geotile grid aggregation | 2 ms | 7.0 ms | 3.5x faster |
| Centroid | 176 ms | 599 ms | 3.4x faster |
| Polygon centroid | 49 ms | 144 ms | 2.9x faster |
| Geohex grid aggregation | 6 ms | 15.6 ms | 2.6x faster |
WarpSearch: Serverless, server time, run 2026-11-14. Elasticsearch: Elastic nightly benchmarks, track geopoint, single node, 30-day median to 2026-11-13.
Full-text and structured search
geonames and pmc
Structured search over 11.4M place names and full-text search over 574K scientific articles. WarpSearch server time against Elasticsearch's nightly latency, p50 / p99.
| Track / task | WarpSearch | Elasticsearch | At p50 |
|---|---|---|---|
| geonames sort by geonameid, descending | 1 / 2 ms | 20.4 / 25.5 ms | 20x faster |
| geonames sort by geonameid, descending, with search_after | 2 / 3 ms | 18.8 / 23.6 ms | 9.4x faster |
| geonames sort by geonameid, ascending | 1 / 1 ms | 6.7 / 8.7 ms | 6.7x faster |
| geonames sort by population, descending | 1 / 2 ms | 6.5 / 8.5 ms | 6.5x faster |
| geonames match all | 1 / 1 ms | 3.9 / 6.0 ms | 3.9x faster |
| geonames term | 1 / 1 ms | 2.9 / 4.2 ms | 2.9x faster |
| geonames phrase | 1 / 1 ms | 2.8 / 4.1 ms | 2.8x faster |
| geonames country terms aggregation | 53 / 54 ms | 127.4 / 146.8 ms | 2.4x faster |
| geonames geo decay function score | 344 / 406 ms | 491.4 / 532.8 ms | 1.4x faster |
| pmc term | 2 / 4 ms | 6.3 / 8.1 ms | 3.2x faster |
| pmc phrase | 3 / 4 ms | 6.2 / 8.3 ms | 2.1x faster |
| pmc monthly date histogram | 5 / 8 ms | 12.6 / 16.5 ms | 2.5x faster |
| pmc match all | 4 / 6 ms | 6.3 / 8.3 ms | 1.6x faster |
| pmc scroll, whole pass | 247 / 256 ms | 405.8 / 457.8 ms | 1.6x faster |
WarpSearch: Serverless, server time (took), geonames on v0.4.0 (2026-11-14), pmc on v0.4.1 (2026-11-16; scroll pass 2026-11-14). Elasticsearch: Elastic nightly benchmarks, single node, 30-day median of Rally latency, measured with the client next to the node.
Reproduce
Stock esrally 2.13.0 and the official rally-tracks against the Elasticsearch-compatible endpoint, benchmark-only pipeline. Replace the key with your own.
esrally race --track=nested --pipeline=benchmark-only \ --target-hosts=api.warpsearch.io:443 \ --client-options="use_ssl:true,verify_certs:true,api_key:'<your key>'"
Run reports: nested 20261116T210359Z (queries) and 20261117T064411Z (indexing); geopoint 20261114T083347Z; geonames 20261114T061319Z; pmc 20261116T162706Z and 20261114T092546Z.