BENCHMARKS · TPC-H SF1
The only Postgres-compatible engine to finish all 22.
On this run only EzraDB and StarRocks complete every TPC-H query. EzraDB places second overall — behind a specialized OLAP engine running its tuned dialect — and lands 3.2–5.5× ahead of the general-purpose engines on the shared query set.
Honest by design: the full five-engine table is below, StarRocks in the lead. The methodology is disclosed in full, and the harness ships with the public release so every number can be rerun on your own hardware.
THE RACE · REPLAYED IN REAL TIME
Watch the run. No compression, no staging.
Derived from TPC-H, SF1 · 2026-05-23 · single-tenant exclusive host · 10 warmup / 30 timed
TPC-H SF1 replay: StarRocks finishes first; EzraDB is second and is the only Postgres-compatible engine to complete all 22 queries. The general-purpose engines run 3–5× longer and abort several queries.
Replayed in real time — every lane runs exactly as long as its measured Σ p50. Lower is better; lane length is proportional to time.
StarRocks runs its tuned dialect; EzraDB runs canonical SQL. Multipliers on the shared query set. SF10/SF100 pending. StarRocks crosses first — it won this run. The device worth watching is completion: only two lanes light all 22 pips, and one of them speaks the Postgres wire protocol.
Five engines, one exclusive host.
Derived from TPC-H, SF1 · 2026-05-23 · single-tenant exclusive host · 10 warmup / 30 timed
| Engine | Class | Σ p50 | Completed |
|---|---|---|---|
StarRocks 3.3.5Leader | Specialized OLAP | 1,838 ms | 22/22 |
EzraDB2nd | HTAP | 2,383 ms | 22/22 |
ClickHouse 24.12 | General-purpose | 7,616 ms | 15/22 |
PostgreSQL 18 | General-purpose | 12,564 ms | 16/22 |
TimescaleDB 2.17.2 | General-purpose | 13,184 ms | 17/22 |
Lower Σ p50 is better. Only the two engines that complete all 22 queries are directly comparable end-to-end; the general-purpose engines abort a handful of queries, so their totals cover fewer results.
ON THE SHARED QUERY SET
3.2–5.5× over general-purpose engines.
Speed-ups are computed only across queries every engine in the pair completed — no credit for queries a competitor never finished.
METHODOLOGY
Disclosed in full — because you can rerun it.
Exclusive host
Single-tenant, exclusive host per run — no noisy neighbours, one engine measured at a time.
10 warmup / 30 timed
Each query gets 10 warmup executions, then 30 timed runs; we report the p50 and sum across the query set.
Canonical vs tuned SQL
EzraDB runs canonical TPC-H SQL. StarRocks runs its own tuned dialect — an advantage we disclose rather than hide.
Pinned images & versions
Every competitor is pinned to the exact image and version shown in the table, captured on the run date.
SCALE LADDER
SF1 today. SF10 and SF100 are pending.
We publish only what we have measured. Larger scale factors are scaffolded in the harness but have not been run — so there are no numbers to show yet.
The only publishable run today.
Harness scaffolded; not yet run.
Harness scaffolded; not yet run.
REPRODUCIBLE BY DESIGN
One harness, on your own host.
The full run — all five engines, warmups, timings, and result digests — is driven by a single exclusive-host harness. It ships alongside the public release so every number on this page can be reproduced on your own hardware.
CAVEATS
What this run does and does not prove.
StarRocks runs its tuned dialect; EzraDB runs canonical SQL. Multipliers on the shared query set. SF10/SF100 pending.
Results are derived from the TPC-H Benchmark and are not comparable to published TPC-H Benchmark results. This is not an audited TPC result and reports no official TPC metrics. TPC and TPC-H are trademarks of the Transaction Processing Performance Council.
Cross-engine digest diffs
Four TPC-H queries produce result digests that differ across engines — driven by ordering of ties and NUMERIC/rounding conventions rather than a wrong answer. We flag them instead of quietly dropping them:
This is a TPC-H analytics benchmark at SF1 on a single exclusive host. It is not a measure of concurrent transactional throughput, ingest, or multi-node scale — those runs are not published here.