justinli500 commented on code in PR #49855:
URL: https://github.com/apache/arrow/pull/49855#discussion_r3593749848
##########
cpp/src/parquet/arrow/reader.cc:
##########
@@ -1275,10 +1239,31 @@
FileReaderImpl::GetRecordBatchGenerator(std::shared_ptr<FileReader> reader,
reader_properties_.cache_options());
END_PARQUET_CATCH_EXCEPTIONS
}
+ // GH-39808: evict each row group's bytes as the decoded prefix advances, so
+ // memory stays bounded. Only this read-once path evicts, so PreBuffer()'s
+ // contract is unchanged for other callers.
+ std::shared_ptr<ReadCacheEvictionState> eviction_state;
+ if (reader_properties_.pre_buffer() && !column_indices.empty() &&
+ !row_group_indices.empty()) {
+ const int64_t kNoMoreRanges = std::numeric_limits<int64_t>::max();
+ std::vector<int64_t> evict_before(row_group_indices.size() + 1,
kNoMoreRanges);
+ for (int64_t i = static_cast<int64_t>(row_group_indices.size()) - 1; i >=
0; --i) {
Review Comment:
Just checked the descending order case - it seems safe, but eviction is
conservative and may retain data until the lower-offset groups finish. The
Dataset path targeted by this PR supplies ascending row groups, so I’ve kept
the simpler watermark approach here.
I think progressive eviction for arbitrary row-group order would require
broader range tracking, but I can extend the implementation if the maintainers
would like that in this PR.
--
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]
