Skip to content

SYSTEM Cited by 1 source

HammerDB

What it is

HammerDB is an open-source database benchmarking tool at hammerdb.com. It provides workload-simulation harnesses for several databases (Postgres, MySQL, Oracle, SQL Server, MariaDB, Db2, Redis, etc.) and the primary OLTP workload TPROC-C — a TPC-C-derived warehouse ordering benchmark that reports New Orders Per Minute (NOPM).

Why it shows up on this wiki

Databricks used HammerDB TPROC-C to measure the impact of disabling compute-side Full Page Write and moving image generation to the storage tier (per the 2026-05-07 Lakebase post). Canonical measurement:

Compute size Before (NOPM) After (NOPM) Gain
4 vCPU 78,876 94,891 +20%
16 vCPU 95,832 269,189 2.8×
32 vCPU 95,686 439,300 4.5×+

The pre-change 16v→32v flat scaling (95,832 → 95,686) is the canonical signal the FPW bottleneck was capping throughput before compute did. Post-change linear scaling confirms the bottleneck was removed. (Source: sources/2026-05-07-databricks-how-lakebase-architecture-delivers-5x-faster-postgres-writes)

TPROC-C framing

HammerDB documents TPROC-C as "a derivative of the TPC-C specification, used for internal benchmarking only" — the results are not official TPC-C results because HammerDB does not run the full audited TPC benchmark protocol. The workload shape (warehouses, districts, customers, new-orders, payments, order-status, delivery, stock-level) matches TPC-C and produces comparable NOPM / TPM-C figures for within-deployment comparison.

Not to be confused with:

  • pgbench — Postgres's built-in benchmark tool; simpler workload, different scale-factor semantics.
  • systems/sysbench — general-purpose benchmark with OLTP + custom-Lua workloads; often used for comparative MySQL / Postgres write-throughput measurements.
  • YCSB — Yahoo Cloud Serving Benchmark; KV-shape workload, not OLTP.

Representativeness caveats

TPC-C (and its derivatives) is a write-heavy OLTP workload with a specific mix of transactional patterns. Every database workload is different; a 5× improvement on TPROC-C does not imply a 5× improvement on any particular customer's production workload. Per AGENTS.md concepts/benchmark-representativeness, vendor-published benchmark numbers should be read as directionally correct rather than as production predictions.

Lakebase's 2026-05-07 post mitigates this by pairing the synthetic-benchmark numbers with production customer measurements (56 vCPU workload: WAL 30 MB/s → 1 MB/s; Synced Tables customer: 17k → 62k rows/sec), which provides cross-validation on real workload shapes.

Seen in

Last updated · 542 distilled / 1,571 read