Skip decode steps in Parquet reader when nullable columns have no nulls #15332
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@abellina when you have a chance I'd be curious to see how this works with your benchmarks. cc @nvdbaranec |
github-actions
bot
added
Python
github-actions
bot
added
Python
|
@etseidl I'll take a look at this and we will try to run with our benchmarks, but I think it might be early next week. I am sorry about the delay in responding, I have to fix some communication settings on my end. I would not block this PR as we will detect regressions quickly in our benchmark runs. |
mhaseeb123
approved these changes
May 3, 2024
|
/ok to test |
|
The idea and changes look good to me. I would like to see improvement results from benchmarks as well. As always, thanks Ed for a great effort! 🙂 |
vuule
reviewed
May 3, 2024
Co-authored-by: Vukasin Milovanovic <vmilovanovic@nvidia.com>
|
/ok to test |
|
Apologies for leaving a trailing whitespace in one of the comments. As they say, "this little maneuver will cost us 51 years" |
vuule
approved these changes
May 3, 2024
… to developer negligence
|
/ok to test |
😬 I'm fine with waiting to see how this works in the spark environment. It's easier to not commit than to have to revert 😄 |
|
I should have numbers for this today @etseidl sorry for the delay. |
|
Thanks for testing this @abellina 🙏 |
|
@etseidl I am not seeing a difference in our tests. I am having trouble getting baselines to be as fast as normal in this environment, so take this with a grain of salt, but if I compare last night's nightly against this branch it's just noise, some are higher some are lower, and the queries I expect to see impact from scan (query9, query88) don't really see a change. |
vuule
added
5 - Ready to Merge
|
Thanks @abellina. Looks like there are no regressions, this is good to merge. |
|
/merge |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
Closes #15266.
Some block and warp operations and some atomics can be avoided during Parquet page decoding when nullable columns are known to not contain any null values. One issue, however, is that since the columns are nullable, the null mask has to be set. My approach in this PR was to initialize nullable column buffers with an
ALL_VALIDmask rather thanALL_NULL. This will work when nulls are present becausestore_validity()sets the bitmask to all zeros before ORing in the passedvalid_mask.Benchmarks modified to not emit nulls showed a good improvement in decoding time for fixed-width data types.
Checklist