feat: add Spark-compatible xxhash64 function#21967
Draft
andygrove wants to merge 3 commits intoapache:mainfrom
Draft
feat: add Spark-compatible xxhash64 function#21967andygrove wants to merge 3 commits intoapache:mainfrom
andygrove wants to merge 3 commits intoapache:mainfrom
Conversation
…ning Replace the flattened single-macro dispatch with the modular macro structure from Apache DataFusion Comet. The new macros add: - Optimized list-of-primitives walker (`hash_list_with_primitive_elements!` + `hash_list_primitive!`) that walks the underlying values buffer directly, avoiding a `Vec<u64>` allocation per non-null list element and the recursive dispatch overhead. - Monomorphized fast paths for common `Map` (key, value) type combinations. - Specialized `hash_array_*` macros per category (boolean, primitive, primitive_float, small_decimal, decimal) so the dispatch is shorter and each path is independently optimizable. - Per-row scalar return path in `xxhash64` that skips the `Int64Array` allocation when `num_rows == 1`. Fix a bug in the previous list-hashing path: it seeded every element with the current row hash and then overwrote the row hash with each element's result, so only the last element actually contributed. Spark chains the hashes across list elements (`hash(elem_n, hash(elem_n-1, ... seed ...))`). The SLT expected value for `xxhash64(make_array(1, 2, 3))` is updated accordingly to match Spark. Add `Utf8View`, `BinaryView`, and `Null` arms in `create_hashes_internal!` so the new dispatch covers everything the previous implementation did.
- Reinterpret `Vec<u64>` as `ScalarBuffer<i64>` via `Buffer::from_vec` so
the result `Int64Array` reuses the hashes allocation instead of walking
it through `into_iter().map(|h| h as i64).collect()`.
- Drop the unused `'a` lifetime and slice return on `create_xxhash64_hashes`;
every caller discards the slice.
- Use `key.as_usize()` (matching `datafusion_common::hash_utils`) instead of
the fallible `key.to_usize().ok_or_else(...)` block — dictionary key types
always fit in `usize`.
- Drop the `args.args.is_empty()` guard; `Signature::variadic_any` already
rejects zero-arg calls during type coercion.
- Replace the `test_xxhash64_f32` / `_f64` tests (which hashed
`if -0.0 == 0.0 && (-0.0).is_sign_negative() { ... }` — a constant-true
conditional) with tests that exercise the `hash_array_primitive_float!`
normalization path through `create_xxhash64_hashes`.
Member
Author
Contributor
|
Can this help with apache/datafusion-comet#4131 ? |
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
Which issue does this PR close?
Rationale for this change
Donate the
xxhash64hash function from Comet so that other projects can benefit from it.The function was initially implemented in Comet by @advancedxy.
This is a continuation of #19627, which has gone stale. To keep the change focused
and easier to review, this PR adds only
xxhash64. Murmur3 will follow in a separate PR.The first commit is the same as the version of #19627 that was previously approved.
The second commit implements optmizations and bug fixes from the latest Comet version.
The third commit is cleanup.
What changes are included in this PR?
xxhash64(expr1, expr2, ...)todatafusion-spark.-0.0normalization, dictionaries (with and without nulls),
FixedSizeBinary,Struct, andList.datafusion/sqllogictest/test_files/spark/hash/xxhash64.sltwith values verified against Spark.
Are these changes tested?
Yes, both Rust unit tests and sqllogictest are included.
Are there any user-facing changes?
A new
xxhash64scalar function is available indatafusion-spark.