Migrate CAGRA merge() to caller-concatenated dataset + offsets contract - #2507
Open
HowardHuang1 wants to merge 5 commits into
Conversation
merge() previously allocated the merged dataset buffer and copied each input index's rows into it internally, hiding memory allocation and copies from the caller. This migrates it to the same explicit caller-owned-buffer contract already used by extend(): callers concatenate every input index's dataset (applying any row filter) into a single buffer themselves, and pass per-index offsets marking where each index's rows start. merge() now only builds/merges the graph and binds the returned index to that buffer. A new merged_dataset_offsets() helper is added for the bitset-filtered case, where the caller can't otherwise derive per-index surviving row counts; unfiltered callers can compute offsets directly (cumulative index sizes) without it. - C++: cagra::merge() and Fastener/rebuild internals updated; new cagra::merged_dataset_offsets(). - C API: cuvsCagraMerge/cuvsCagraMergeWithParams updated to match; new cuvsCagraMergedDatasetOffsets. - Java: CagraIndex.merge() public API updated to the same contract (breaking); real in-repo consumer (cuvs-lucene) and tests updated. - Rust, Python, Go: new merge wrappers added (none existed before), matching the C API's contract and modeled on each language's existing extend()/update_dataset() conventions. Also fixes an unrelated pre-existing build break in cpp/src/core/bloom_filter.cu (cuco::default_filter_policy renamed to cuco::bloom_filter_policy in the pinned cuCollections version), needed to get any build of libcuvs compiling in this environment. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
HowardHuang1
force-pushed
the
HH-migrate-merge-dataset-to-take-inputs-concat-buffer-and-offset-26_10
branch
from
August 26, 2026 04:33
46bed89 to
87a0075
Compare
Contributor
Author
|
/ok to test 87a0075 |
…taset-to-take-inputs-concat-buffer-and-offset-26_10
Contributor
Author
|
/ok to test 50dd606 |
…taset-to-take-inputs-concat-buffer-and-offset-26_10
Rust: regenerate cuvs-sys bindings.rs for the offsets-based merge contract, and switch the merge tests from PaddedDataset::new (owning copy, rejects already-aligned sources) to DatasetView::new, mirroring how Index::build picks a view. Go: add MakePaddedDatasetAuto, mirroring BuildIndex's existing padded-vs-standard branch, so merge_test.go can build its merged buffer without needing its own cgo import (which isn't allowed in _test.go files). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…taset-to-take-inputs-concat-buffer-and-offset-26_10
Contributor
Author
|
/ok to test a83394f |
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
1 participant
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.
Addresses issue #2401.
merge() previously allocated the merged dataset buffer and copied each input index's rows into it internally, hiding memory allocation and copies from the caller. This migrates it to the same explicit caller-owned-buffer contract already used by extend(): callers concatenate every input index's dataset (applying any row filter) into a single buffer themselves, and pass per-index offsets marking where each index's rows start. merge() now only builds/merges the graph and binds the returned index to that buffer.
A new merged_dataset_offsets() helper is added for the bitset-filtered case, where the caller can't otherwise derive per-index surviving row counts; unfiltered callers can compute offsets directly (cumulative index sizes) without it.
Also fixes an unrelated pre-existing build break in cpp/src/core/bloom_filter.cu (cuco::default_filter_policy renamed to cuco::bloom_filter_policy in the pinned cuCollections version), needed to get any build of libcuvs compiling in this environment.