XiaoHongbo-Hope opened a new pull request, #1478: URL: https://github.com/apache/paimon-rust/pull/1478
## Summary For OR predicates whose arms each contain one BTree-indexed equality on the same field plus unindexed conjuncts, evaluate the indexed equalities as one IN lookup. Keep the original predicate as the table reader's residual and mark the index result inexact. Other predicate shapes retain the existing OR path. ## Verification - Global index scanner suite: 80 passed, 1 ignored; `cargo fmt --all -- --check` and `git diff --check` passed. - On the customer image workload (one `content_key`, snapshot 33, full BLOB bytes, CRC enabled), all 224 reads across the 100-image and 16/1,000-image experiments returned the expected image SHA-256. - A separate 12-predicate check found identical serialized split hashes and all ten recorded scan-stat fields before and after this change. - With 100 images, median warm planning fell from 107.6 to 86.3 ms (paired improvement in 11/12 seeds). Warm full-read median was 203.4 vs 191.7 ms. With 1,000 images, warm planning improved by a paired median 22.8 ms (7/8 seeds); cold full-read latency was noisy and showed no reliable gain. - Disabling the global index remained faster for this broad random OR workload (100-image warm total 168.6 ms). This PR optimizes the indexed path; it does not propose an automatic index-skip threshold. Reproducible scripts, wheel/source hashes, environment, raw observations and paired analysis: [xrobot experiment archive](https://code.alibaba-inc.com/paimon/xrobot/-/tree/codex/xrobot-or-batch-20261121/experiments/customer-q3-real-images-20261119/single-lance-20261120/diagnostic-100/planning). -- 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]
