andygrove opened a new issue, #7499: URL: https://github.com/apache/datafusion-comet/issues/7499
Part of #7267. ## Problem Two things make the native Parquet scan, and the native Iceberg scan that uses the same arrow-rs reader, overflow a string or binary column where Spark's reader does not. **Flat dictionary-encoded columns are expanded per batch.** Comet asks arrow-rs for plain `Utf8`/`Binary`, so the reader copies a dictionary value into the batch once per row that references it (`extend_from_dictionary`, parquet 59.3.2 `src/arrow/buffer/offset_buffer.rs:92-118`). A column with a few wide distinct values is stored as a small dictionary page plus tiny keys, and can still expand past 2 GiB in one 8192-row batch: one 300 KB value repeated 8192 times is 2.4 GB. The scan fails with `index overflow decoding byte array`. Spark's vectorized reader keeps the dictionary ids and decodes a value only when a row reads it, so it succeeds. #7254 covers the nested case; flat columns overflow inside arrow-rs instead. **Twice the rows per batch.** Comet reads `spark.comet.batchSize` rows per batch, 8192 by default. Spark's vectorized Parquet reader reads `spark.sql.parquet.columnarReaderBatchSize` rows (4096), and Iceberg's Java reader 5000. With values of 256-512 KiB in row groups large enough to fill a batch (large row-group sizes or highly compressible values), Comet's batch passes 2 GiB and Spark's does not. The DELTA_LENGTH_BYTE_ARRAY decoder panics instead of returning that error (parquet 59.3.2 `src/arrow/array_reader/byte_array.rs:514`). The epic lists the scan as honoring the batch size. It does, but in both cases the batch size alone doesn't bound the bytes. ## Workaround Lower `spark.comet.batchSize`. ## Proposed fix - Keep dictionary-encoded byte columns as dictionaries out of the reader and unpack them after the scan in row ranges bounded by bytes, or cap scan batches by bytes. - At least, default the native scan's batch size to Spark's reader batch size. - Upstream: return an error from the DELTA_LENGTH_BYTE_ARRAY decoder instead of panicking. ## Tests A Parquet file with a flat dictionary-encoded string column, one 300 KB value repeated 8192 times, stays tiny on disk. Reading it with the native scan and comparing with Spark shows the problem, but needs about 2.5 GB of memory, so test the byte-bounded batching with a small budget injected for the test. -- 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]
