pingcap / tiflash

The analytical engine for TiDB and TiDB Cloud. Try free: https://tidbcloud.com/free-trial
https://docs.pingcap.com/tidb/stable/tiflash-overview
Apache License 2.0
946 stars 409 forks source link

Vector search lower recall when there is multiple ColumnFile in delta layer with vector index #9599

Closed Lloyd-Pottiger closed 58 minutes ago

Lloyd-Pottiger commented 2 days ago

Bug Report

Please answer these questions before submitting your issue. Thanks!

1. Minimal reproduce step (Required)

  1. create a table with vector type column
  2. insert more than 10K rows
  3. vector search

2. What did you expect to see? (Required)

recall > 90%

3. What did you see instead (Required)

recall < 20%

4. What is your TiFlash version? (Required)

master

Lloyd-Pottiger commented 2 days ago

Introduce by https://github.com/pingcap/tiflash/pull/9547