XiaoHongbo-Hope opened a new pull request, #1246: URL: https://github.com/apache/paimon-rust/pull/1246
## What Honor explicit MAP value ROW projections in the Parquet reader. Keep all map keys and read only requested value leaves. Preserve file field IDs, names and order during schema reconciliation. Non-Parquet formats and selected-key synthetic ROW projections retain existing behavior. Predicates referencing the map still include all its leaves. No new configuration, file rewrite or Arrow dependency changes. ## Tests - Projection mask and predicate widening. - ZSTD Parquet roundtrip covering null/empty maps, duplicate keys, null value structs and null children. - File-field ID/name/order preservation and unchanged non-Parquet behavior. - Candidate before rebasing: 101 Parquet tests and 44 data-file-reader tests passed. - Real Python-to-native reads: 232 full/projected JPEG queries matched; 232 logical map comparisons passed. The PR is isolated on current main; tests above ran on the experiment base, not this rebased head. CI validation is pending. ## Performance scope 64-worker mixed-version workload, same candidate binary, 180s warmup + 180s measurement per arm: full/projected/full = 250.31 / 268.80 / 263.82 samples/s, no missing images. This is not a pure-main baseline: all arms include separate known-file-size and process-front-cache experiments. Only 1.9% improvement over the trailing full arm; no stable throughput gain claimed. Applications must explicitly request the narrower read type to benefit. -- 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]
