Skip to content

Performance regression: cached scalar search benchmarks sharply slower at d4da624 #8290

Description

@westonpace

Lance benchmark run: https://github.com/lancedb/lance-bench/actions/runs/31006283603\nLance SHA: d4da624\nSuite: Python ci_benchmarks\n\nConfirmed affected benchmarks:\n- test_basic_btree_search[cache-None-none]: 12.8 ms at 20c7378 -> 375.8 ms at d4da624; regression report +630.72%, p<0.000001.\n- test_basic_btree_search[cache-small_strings-none]: 48.3 ms -> 114.1 ms; +41.69%, p=0.000005.\n- test_basic_bitmap_search[cache-None-none]: the last-four window is +28.58%, p=0.000111.\n\nThe sharp BTree change appeared between 20c7378 (2026-07-31) and d4da624 (2026-08-05). b0b631e ("support register and switch pluggable cache backend") landed immediately after the prior measured commit and is a plausible starting point, though the full range should be bisected.\n\nPlease reproduce with python/ci_benchmarks/benchmarks/test_search.py using the cache variants above.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions