Benchmark regression - 2026-05-03 #227

Activity

1. github-actions

   added

   performanceBenchmark performance regressionsBenchmark performance regressions

   on May 3, 2026
2. adityamukho commented on May 13, 2026

   

   adityamukho

   on May 13, 2026

   Collaborator

   More actions

   I checked the referenced run against adjacent nightly benchmark runs. This does not look like a one-off runner spike.

   Evidence from concurrent_btree_scan median times:

   • 2026-05-02 (be3d598): 100k_2t 151.22 ms, 100k_4t 241.72 ms, 100k_8t 477.09 ms
   • 2026-05-03 (a51abf7, this alert): 100k_2t 164.44 ms, 100k_4t 271.40 ms, 100k_8t 574.52 ms
   • 2026-05-04 (8a71b97): 100k_2t 177.11 ms, 100k_4t 283.17 ms, 100k_8t 592.66 ms
   • 2026-05-13 (643fb68): 100k_2t 171.19 ms, 100k_4t 273.42 ms, 100k_8t 575.37 ms

   The 10k variants also moved, but the 100k variants show the clearest persistent step-change. The current run is still near the alerted level, so I am leaving this open.

   Likely starting point: changes between be3d598..a51abf7, especially the storage/durability work in #225. concurrent_btree_scan builds a checkpointed file DB and runs concurrent EAVT point/range reads, so the regression may be in the committed file/B+tree read path or in per-query work introduced around file locking/checksum/durability changes rather than benchmark runner noise.
3. adityamukho

   closed this as completedin #236on May 14, 2026