Repository navigation
[Parquet] Add BooleanArray based row selection - #6624
XiangpengHao wants to merge 6 commits into
Conversation
c597e4a to
42938bf
Compare
42938bf to
1f017ff
Compare
|
The benchmark can take a while to run, here're some of my results visualized by ChatGPT. Note that both x and y axies are in log scale. Selection ratio 1e-5 means 0.0001% of the rows are selected, 0.3 means 30% are selected, and so on.
|
|
Perhaps we could model the number of disjoint selections as opposed to raw selectivity, this more accurately models the sorts of selection one gets from evaluation against the page index (which necessarily has page level granularity) Edit: although the methods benchmarked are those used for late materialization, so perhaps it doesn't matter |
tustvold
left a comment
tustvold
left a comment
There was a problem hiding this comment.
I've not reviewed in fine detail, but the code looks good to me
|
I will try and find time to review this today or tomorrow as well -- thank you @XiangpengHao and @tustvold |
alamb
left a comment
alamb
left a comment
There was a problem hiding this comment.
Thanks @XiangpengHao -- I think this is an interesting direction to head (use bitmap for selection). However, I am not sure about this exact implementation)
I left some questions below.
One thing that seems like it might be more useful is to switch RowSelection into an enum so we can dynamically switch between the two representations based on what is actually selected
enum RowSelection {
Ranges(Vec<Selection>),
Btitmap(..)
}I realize one challenge with this is that BooleanArray is part of arrow and not available in Parquet directly.
| /// A selection of rows, similar to [`RowSelection`], but based on a boolean array | ||
| #[derive(Debug, Clone, PartialEq, Eq)] | ||
| pub struct BooleanRowSelection { | ||
| selector: BooleanBuffer, |
There was a problem hiding this comment.
One question I have while reviewing this PR is "why not just use BooleanArray directly?"
It seems like most of these methods would be generally useful on BooleanArray (or perhaps as kernels) -- and in fact most of them are already (like union is | and intersection is &).
Another thing that might be interesting to do if we really care about performance is to look into implementing in-place updates -- maybe something like https://docs.rs/arrow/latest/arrow/compute/fn.binary_mut.html for BooleanArray)
That should work, my initial thoughts was to only use bitmap-based selectors when evaluating predicates (i.e., when filter pushdown is enabled); and always do range-based selectors otherwise.
I agree that most of the functionality overlaps with |
|
Marking as draft as I believe we plan additonal work for this PR |
|
|
||
| #[cfg(target_arch = "x86_64")] | ||
| #[target_feature(enable = "bmi2")] | ||
| unsafe fn boolean_buffer_and_then_bmi2( |
There was a problem hiding this comment.
FWIW, if you ever encountered a situation where and_then show up in the flamegraph, here's the implementation that will speed up it by 5-10x. cc @alamb @zhuqi-lucas who are working on filter pushdown
I admit the code is much less readable than the fallback implementation, but it boils down to use the PDEP instructions available in almost all x86 machines. Rest of the code are just work to make sure alignment, handle edge cases, optimize for branching etc.
There was a problem hiding this comment.
I definitely struggled with this when thinking how to do fallback.
|
I started trying to incorporate this code in to the parquet reader: |
| unsafe { boolean_buffer_and_then_bmi2(left, right) } | ||
| } else { | ||
| // Fallback to the original implementation | ||
| boolean_buffer_and_then_fallback(left, right) |
There was a problem hiding this comment.
When I was trying to use this code in #8902, I could not find this fallback.
@XiangpengHao did it in liquid cache:
|
Thank you for your contribution. Unfortunately, this pull request is stale because it has been open 60 days with no activity. Please remove the stale label or comment or this will be closed in 7 days. |
) # 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. --> - Closes #10140 - closes #8844 - closes #6624 from @XiangpengHao - Past attempts: #8902 # Rationale for this change `RowSelection` currently stores selections as `Vec<RowSelector>` (16 bytes per selector). This is compact for long runs, but expensive for scattered matches. With ~35% isolated single-row hits, it uses about 11.2 bytes per input row. A `BooleanBuffer` uses 1 bit per input row, about 90x less memory. The reader can also choose the `Mask` strategy, which converts selectors back into a bitmap. When the caller already had a bitmap, this conversion round-trip is unnecessary. **This PR lets `RowSelection` preserve a caller-provided bitmap and pass it directly to mask execution.** This is not intended to claim broad DataFusion / TPC-DS / ClickBench speedups. Current common DataFusion SQL paths generally do not naturally produce bitmap-backed `RowSelection`s. The practical benefit is for integrations that already have a row-level bitmap and need Parquet to consume it without materializing a large selector list. # What changes are included in this PR? `RowSelection` can now be backed by either `Vec<RowSelector>` or `BooleanBuffer`. New public construction: ```rust pub fn RowSelection::from_boolean_buffer(mask: BooleanBuffer) -> Self; impl From<BooleanBuffer> for RowSelection; ``` Methods that can work directly on the bitmap now do so: - `iter()` still returns `&RowSelector` (non-breaking); mask-backed selections lazily materialize a selector cache on first call, while internal hot paths bypass it and use the `BooleanBuffer` / `MaskRunIter` directly - `row_count` / `skipped_row_count` use a cached popcount - `selects_any` uses `set_indices().next()` - `trim` preserves mask backing via `BooleanBuffer::slice` - `intersection` / `union` on `Mask`+`Mask` use `BitAnd` / `BitOr` - `split_off` on a mask uses `BooleanBuffer::slice` (`O(1)`, both halves stay mask-backed) - `limit` slices at the selected-row boundary via `find_nth_set_bit_position`, staying mask-backed - `offset` finds the first selected row to keep via `find_nth_set_bit_position` and rebuilds only the mask buffer, avoiding selector materialization - `and_then` applies the inner selection over the mask's set positions, returning a mask-backed result - `FromIterator<RowSelection>` concatenates `BooleanBuffer`s when every input is mask-backed Mixed inputs, and existing selector-backed inputs, still use the existing selector helpers. Existing callers keep the same behavior. **The reader (`ReadPlanBuilder::build`) passes a mask-backed selection straight to `RowSelectionCursor::new_mask_from_buffer`, so it skips rebuilding the bitmap from selectors.** `Auto` resolution works directly on the bitmap (early-exit run counting), without converting the backing. Integrates with #10288: both mask cursor constructors carry `LoadedRowRanges`, so a caller-provided bitmap stays within loaded pages when page pruning skips pages. Also adds `MaskRunIter` + `RowSelection::as_mask` for zero-allocation RLE iteration over a mask, the `row_selector_boolean_buffer` benchmark, and `read_auto` / mask-backed input modes in `row_selection_cursor`. # Are these changes tested? Yes. This PR extends the existing `RowSelection` unit tests with coverage for: - constructing from `BooleanBuffer`, including empty and all-unset masks - `From<BooleanBuffer>` - preserving mask backing across clone, `split_off`, `limit`, `offset`, `and_then`, and all-mask `FromIterator<RowSelection>` - falling back to selector backing for mixed-backed concatenation - equality between equivalent selector-backed and mask-backed selections - mask-backed `intersection` / `union`, including uneven-length inputs - fuzz-style equivalence between mask-backed selections and the existing `from_filters` selector path - sparse-page regression tests from a non-byte-aligned `BooleanBuffer::slice(...)`, under both `Mask` and `Auto` policies (#10288 integration seam) - mask chunks stop at `LoadedRowRanges` boundaries - fuzz tests for `boolean_mask_from_selectors` and `trim_mask` (including non-zero offsets) - remaining branches checked with `cargo llvm-cov` # Are there any user-facing changes? No breaking API changes. New public APIs: `RowSelection::from_boolean_buffer`, `From<BooleanBuffer>`, `RowSelection::as_mask`, `MaskRunIter`. --------- Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org>
Which issue does this PR close?
Rationale for this change
Implement the row selection based on BooleanArray.
The implementation only covers the core operations like
union,intersection,and_then, and is already quite large. Once we agree with the generally direction, I'll add remaining implementaions in follow up PRs.What changes are included in this PR?
I added more benchmark to cover different selection ratios, as suggested by @tustvold and @alamb
Are there any user-facing changes?