TECHNICAL EVIDENCE

Read benchmarks with their context.

Engine tests are not product guarantees. Data, hardware, versions, caching and query selection change the result.

EXTERNAL · VENDOR-PUBLISHED

E1 · AgentLogsBench

A VeloDB-published comparison at roughly 100M rows. AWS m6i.8xlarge (32 vCPU, 128 GiB, gp3), 20 queries; first run reported as cold and third as hot. Q13 hot: Doris 0.088s / Elasticsearch 1.094s. Q05 cold: Doris 11.3s / Elasticsearch 0.757s.

Open source

EXTERNAL · VENDOR-PUBLISHED

E2 · ASOF JOIN

A VeloDB-published 100M × 100M standard-NDV test. One ecs.g7.4xlarge (16 cores, 64 GB, 500 GB ESSD PL0); concurrency 1; fastest of three hot runs after a cold run. Doris 4.1.0: 0.97s; ClickHouse 26.3.3.20: 1.93s; DuckDB 1.5.1: 3.62s.

Open source

EXTERNAL · VENDOR-PUBLISHED

E3 · Advance.AI

A VeloDB customer story dated January 30, 2026 reports more than 50% cost improvement after a move from OpenSearch to Doris. This is not a UnifyLogs result or a savings forecast.

Open source

Limits

The exact E1 source revision is not pinned, and its cold method does not reset every Elasticsearch cache. E2 is advanced engine evidence—not a log-search or AI SLA—and ASOF support must be confirmed in the deployed Doris version. Re-test every result with your own data.

NEXT STEP

Validate it with your own logs.

Plan an evaluation