xiangfu0 commented on PR #18103: URL: https://github.com/apache/pinot/pull/18103#issuecomment-4227506269
Follow-up on the split reproducer PR: I updated #18166 so it now includes a direct `3.2` vs `3.3` comparison, not just the `3.3`-only merge-order reproducer. What changed: - `TDigestVersionComparisonTest` loads `t-digest:3.2` and `t-digest:3.3` side-by-side in isolated classloaders and runs the same deterministic Pinot-like hierarchical merge dataset through both versions. - `TDigestMergeOrderReproducerTest` remains the pure `3.3` merge-order sensitivity check. So the answer to “is this a test that passes on 3.2 but fails on 3.3?” is: - the original merge-order test is `3.3`-only - the new comparison test is the direct A/B check between `3.2` and `3.3` Repro: ```bash ./mvnw -pl pinot-segment-local -Dtest=TDigestVersionComparisonTest,TDigestMergeOrderReproducerTest -Dsurefire.failIfNoSpecifiedTests=false test ``` The comparison test asserts that, on the minimized exact-quantile scenario: - `3.2 @ compression 100` stays below `0.0002` max normalized error - `3.3 @ compression 100` is above `0.004` max normalized error and more than `20x` worse than `3.2` - `3.3 @ compression 150` recovers below `0.0002` From the local exploration used to lock in that scenario, the observed numbers were roughly: - `3.2 @ 100`: `0.000074` max normalized error with `121` centroids - `3.3 @ 100`: `0.005094` max normalized error with `55` centroids - `3.3 @ 150`: `0.000049` max normalized error with `79` centroids I also reran the repro command in 10 fresh Surefire JVMs and it passed `10/10`. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
