Reduce memory use when writing tables with very short columns to ORC #18136
Conversation
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
vuule
changed the title
There was a problem hiding this comment.
The logic looks right to me, although I don't know the stripe stream bits well enough to understand why we start off with a 2d span with a dimension that is always of size one. However, this does seem to be reducing the memory usage enough to get CI to pass on the test we've been having problems with, so I'm approving this in the hopes that we can get more data from nightlies over the weekend. Thanks @vuule!
|
I think we still need this to unblock nightly CI? I will update this with the latest |
This test is failing in multiple places right now, such as [this run](https://github.com/rapidsai/cudf/actions/runs/13595690128/job/38014725800) on #18133 and [this run](https://github.com/rapidsai/cudf/actions/runs/13636334843/job/38118996773?pr=18136) on #18136. Let's skip it until we can debug why so that we unblock other CI. --------- Co-authored-by: Peter Andreas Entschev <[email protected]>
Description
Closes #18059
To avoid estimating the maximum compressed size for each actual block in the file, ORC writer uses the estimate for the (uncompressed) block size limit, which defaults to 256KB. However, when we write many small blocks, this compressed block size estimate is much larger than what is needed, leading to high memory use for wide/short tables.
This PR adds logic to take the actual block size into account, and to use the size of the actual largest block in the file, not the largest possible block. This changes the memory usage by orders of magnitude in some tests.
Checklist