Performance

Faster than Trino — now at SF10 too.

Every query differentially validated against Trino 481 — same StorageGRID Iceberg warehouse, both engines capped at 12 GB — with DuckDB's official dbgen/dsdgen as an independent data oracle.

7 / 8suites faster than Trino at SF1
6 / 7suites faster at SF10 too — SSB now at parity
8.9×top suite speedup (TPC-BB, SF1)
Where we stand

Faster at SF1 — and now ahead at SF10

Faster than Trino at SF1 across seven of eight industry suites — and, as of this run, ahead at SF10 too.

The SF10 crossover is closed. SSB — the one suite we trailed for months (0.53× only a few weeks ago) — is now at parity (20.4s vs 21.4s) after shipping build-side key sets to workers. TPC-H (1.7×) and TPC-DS (1.5×, and now 99/99 correct at SF1) lead at SF10. Two honest exceptions we show rather than hide: TPC-C's sub-second OLTP reads at SF1, where fixed dispatch overhead dominates, and the new bank suite at SF10, where spill still degrades under an 8 GB pool — both being worked.

Scale factor ≈ dataset size: SF1 ≈ 1 GB, SF10 ≈ 10 GB. Rig: a dedicated 8-core / 31 GB box, both engines capped at 12 GB against the same StorageGRID Iceberg warehouse, Trino 481.

log scale · drag to zoom · toggle scales in the legend

SQE run duration over time — pick a suite, zoom in. Updated on each sync from the raw benchmark results.

SQE distributed execution: coordinator and workers
Receipts

The eight suites — what each measures, and how we do

Suite & purpose SF1 SF10
SQETrino 481SpeedupPass SQETrino 481SpeedupPass
TPC-H (22) 22-query decision-support benchmark — the most widely benchmarked SQL workload and SQE's primary comparison target. 14.2s 40.8s 2.9x 22/22 66.5s 111.9s 1.7x 22/22
TPC-DS (99) 99-query decision-support benchmark — larger and harder than TPC-H; pushes joins, window functions, and grouping sets. 49.3s 114.9s 2.3x 99/99 216.6s 320.5s 1.5x 98/99
SSB (13) Star Schema Benchmark — 13 queries over one fact joined to four dimensions; the shape data warehouses are optimised for. The suite we trailed longest. 5.2s 7.3s 1.4x 13/13 20.4s 21.4s 1.0x 13/13crossover closed — parity
ClickBench (43) 43 queries over one wide table — log analytics: heavy string filters, point lookups, and counts. SQE's caching shines here. 3.1s 7.3s 2.4x 40/43 10.0s 31.6s 3.2x 40/433 pending a LIMIT-drop regression (q17)
TPC-E (11) Financial-trading workload: date-range filters, point lookups against trade-history, aggregations across many small dimensions. 3.8s 11.6s 3.0x 11/11 test-only 18/18SF10 diff needs ~14 GB for the 10M-row result — validated test-only
TPC-BB (10) BigBench mixed workload: structured analytics plus JSON and text operations over a retail warehouse schema. 53.7s 480.5s 8.9x 10/10 405.5s 1429.7s 3.5x 10/10
Bank (8) Purpose-built banking workload — customer / KYC / accounts with day-partitioned transactions; time-windowed "last 14 days out of N" scans that lean on partition pruning. 10.5s 19.8s 1.9x 8/8 370.5s 85.5s 0.23x 7/8spill stalls under an 8 GB pool — fix in progress
TPC-C (8 read) The read SELECTs from the TPC-C OLTP workload — point lookups against small Iceberg tables (SQE runs the read subset, not transactions). 5.6s 4.3s 0.77x 8/8sub-second reads — fixed dispatch overhead dominates 5.0s 18.0s 3.6x 8/8

Zero row differences between SQE and Trino across every run — each non-match is a case where both engines agree (0 rows). Two exceptions we show rather than hide: TPC-C's sub-second OLTP reads at SF1 (0.77×, fixed dispatch overhead dominates when a query finishes in milliseconds) and the new bank suite at SF10 (0.23×, spill stalls under an 8 GB pool — being fixed). Full per-query method: docs, or run it yourself via the benchmark quickstart.

The leaps

How it got fast — and still is

  1. 12 Jun 2026

    The benchmark that lied

    Differential validation with DuckDB as an independent oracle: caught 16 vacuous TPC-DS queries and a Trino rounding bug (q75).

    Read the story →
  2. 16 Apr 2026

    Our nemesis: TPC-DS q72

    Runtime-filter pushdown drops q72 from 10.7s to 0.77s.

    Read the story →
  3. 14 Apr 2026

    DataFusion 53 + vendored Iceberg fork

    ~40x faster query planning.

    Read the story →
  4. 12 Apr 2026

    Five layers of caching

    8.8x faster than Trino across the suite.

    Read the story →

Current as of 16 August 2026 · synced from the SQE source