Draft POC: Push batch with filter without copy #8103
Conversation
There was a problem hiding this comment.
Thank you @zhuqi-lucas -- this is very cool -- I left some ideas
| IterationStrategy::Slices(s) => FilterPlan::Slices(s), // moved directly | ||
| IterationStrategy::Indices(i) => FilterPlan::Indices(i), // moved directly | ||
| IterationStrategy::SlicesIterator => { | ||
| FilterPlan::Slices(SlicesIterator::new(&pred.filter).collect()) |
There was a problem hiding this comment.
avoiding this allocation will likely help
There was a problem hiding this comment.
Thank you @alamb , i tried now, but it not improved the regression. The compute_filter_plan almost cost nothing for the benchmark profile.
| }); | ||
|
|
||
| // For each contiguous slice, copy rows in chunks fitting target_batch_size | ||
| for (mut start, end) in slices { |
There was a problem hiding this comment.
I suspect to really make this fast it will need to have specialized implementations for the different array types (not use mutable array data)
I think we could yoink / reuse some of the existing code from the filter kernel:
arrow-rs/arrow-select/src/filter.rs
Line 322 in 04f217b
There was a problem hiding this comment.
Thank you @alamb for review and good suggestion.
I found the hot path for profile is, copy_rows:
Especially for the code for null handling:
// add nulls if necessary
if let Some(nulls) = s.nulls().as_ref() {
let nulls = nulls.slice(offset, len);
self.nulls.append_buffer(&nulls);
} else {
self.nulls.append_n_non_nulls(len);
};It may due to we will call more times for copy_rows here, but the original logic is just filter and concact to a batch, and then to push_batch, so it will be friendly for SIMD. Also it will make copy_rows SIMD friendly for bigger batch. So the original run is pretty faster for most cases.
May be we need to only change this logic for selective < 0.005 for example, but it's hard for me to decide.
Which issue does this PR close?
We generally require a GitHub issue to be filed for all bug fixes and enhancements and this helps us generate change logs for our releases. You can link an issue to this PR using the GitHub syntax.
Rationale for this change
Why are you proposing this change? If this is already explained clearly in the issue then this section is not needed.
Explaining clearly why changes are proposed helps reviewers understand your changes and offer better suggestions for fixes.
What changes are included in this PR?
There is no need to duplicate the description in the issue here but it is sometimes worth providing a summary of the individual changes in this PR.
Are these changes tested?
We typically require tests for all PRs in order to:
If tests are not included in your PR, please explain why (for example, are they covered by existing tests)?
Are there any user-facing changes?
If there are user-facing changes then we may require documentation to be updated before approving the PR.
If there are any breaking changes to public APIs, please call them out.