KKcorps commented on PR #20106: URL: https://github.com/apache/pinot/pull/20106#issuecomment-6598319179
This comes from a model of upsert thread orders that StarTree runs on upsert PRs. **The new head still clears the known orders where a consumer starts before older segments of its partition finish loading, and now also clears them on full upsert tables with `dropOutOfOrderRecord` or `outOfOrderRecordColumn`. The model found nothing new.** On those tables, a replica that got the partition in a rebalance used to commit a late row as if it were in order, because the row it should lose to was still loading. Dedup tables no longer store a second row for a key in the same order. I replayed the model's known bad orders with and without the change, and none breaks only with it. A 62-minute search of 393 runs found nothing new. The merged `acquire` checks readiness only while the permit is free, so a load that lands between the check and the permit looks the same as one that lands right after it. Skipping the registration wait was safe in every order the model tried. The model covers one replica of one partition and leaves out the readiness snapshot cache, so a clean search only means this search found nothing. <!-- upsert-model-check head=24e449d42a5f2c6a421af91a3df7afe5591c6010 --> -- 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]
