andygrove opened a new issue, #4994: URL: https://github.com/apache/datafusion-comet/issues/4994
## What is the problem the feature request solves? Comet's native Parquet scan caches each file's footer and page index in DataFusion's `FileMetadataCache`, which it obtains from the task's `RuntimeEnv` (`native/core/src/parquet/parquet_exec.rs`). Comet builds a fresh `RuntimeEnv` for every Spark task, so that cache is discarded when the task ends. Two tasks on the same executor therefore re-fetch and re-parse the same footer, once each. This recurs in two common shapes: - a large Parquet file split into many Spark partitions, each read by a different task; - the same table scanned again by a later stage or query in the same application. Each miss costs a round trip to object storage plus a Thrift parse. Spark's own reader has the same per-task behavior, so this is parity with Spark, but Velox caches file handles and footers process-wide, and this is one of the gaps identified as item P5 in #4842. Sharing the cache also reduces memory: today each concurrent task gets its own 50MB budget. ## Describe the potential solution Hold the metadata caches in a process-wide registry keyed by object store URL, and use it at the scan's single call site, falling back to the per-task cache when sharing is disabled. This requires the cache validity check to actually work first. DataFusion validates cached metadata against `(size, last_modified)`, but Comet builds its `PartitionedFile` with `PartitionedFile::new_with_range`, which hardcodes `last_modified` to epoch 0, and `SparkPartitionedFile` carries no modification time. Validation is therefore size-only today. That is harmless for a cache that lives for one task and a wrong-results hazard for one that lives for the executor, so Spark's `PartitionedFile.modificationTime` needs to reach `object_meta.last_modified`. ## Additional context Item P5 of #4842. -- 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]
