MariaDB Foundation: Reproducing and Validating MDEV‑37992 with TAF

Thanks to Mark Callaghan for filing MDEV‑37992. This regression only appears under specific LRU and buffer‑pool conditions, and requires a memory‑resident workload to expose.

Using TAF, I reproduced the issue across 11.8, 12.3, and 13.1 using both parent commits and patched commits from MDEV‑33966. The fix restores expected throughput in all branches.

The workload used was POINTS‑COVERED‑PK with 10M rows × 8 tables, 40 threads, warmup enabled, and a 90G buffer pool. TAF’s warmup iteration, rest detection, and stable‑state perf capture made the regression immediately visible and ensured consistent results across branches.

The exact TAF properties file and driver script used for reproduction are committed here:

TAF reproduction properties + shell driver script commit:

https://github.com/MariaDB/TAF/commit/64d46f0c0da54d2c380654d49b8b881989e9154c

Below are the exact commit IDs and TPS results so others can reproduce the behavior.

11.8 Branch

Parent: 14362356d14

Patched: db94f391be4e

TPS (Parent → Patched): 17,925 → 66,653

12.3 Branch

Parent: 5981efcecd8

Patched: 4083fbef83c6

TPS (Parent → Patched): 18,739 → 68,646

13.1 Branch

Parent: d0661d9bb8c

Patched: 5ab51d6248bb6

TPS (Parent → Patched): 18,737 → 66,145

Across all three branches, the regression is consistent:

~18–19k TPS in parent builds → ~66–68k TPS in patched builds.

A clear 3.5× throughput improvement after applying MDEV‑33966.

Hardware Acknowledgment

Testing was conducted on a Hetzner Sponsor Host, provided by MariaDB Foundation sponsor Hetzner.com:

Hetzner Sponsor Host

• CPU: AMD EPYC 9454 48‑Core Processor
• CPU Count: 96
• Core Count: 48
• Socket Count: 1
• RAM: 124.98 GB
• OS: AlmaLinux 9.8 (Olive Jaguar)
• Kernel: 5.14.0‑687.29.1.el9_8.x86_64

This system provided the horsepower required to investigate this MDEV.

We appreciate the outstanding support and the quality of the systems Hetzner.com provides.

Thanks again to Mark Callaghan for filing the MDEV and providing the initial direction.

This was a clean example of how TAF helps expose engine‑level regressions and verify fixes across multiple branches with consistent, reproducible workloads.

Jonathan (Jeb) Miller has experience in military leadership, Fire/EMS leadership, computer operations leadership, teaching, and more than 27 years of database performance engineering (PervasiveSQL, MySQL, MariaDB).

TAF is the third benchmarking framework he has helped design and build.

#MariaDB #MySQL #PostgreSQL #Database #DatabasePerformance #Performance