arrow-csr: Memory efficient zero-copy path#626
Merged
Conversation
…ase 1) The DirectArrowResultCollector path already skipped ArrowArray materialization, but each per-batch CSRMetadata still built a dense global indptr (numSourceRows+1 entries, ~800MB each) and padded its tail to numSourceRows. With ~50 parallel batches that duplicated indptr was the entire ~45GB blow-up on a 0.75B-edge scan (indices/edgeIDs were already at theoretical size). Replace the dense per-batch indptr with a sparse (srcRows, counts) representation: one run per distinct source row the batch touched. This is valid because the rel scan emits edges in non-decreasing source order per thread (CSR storage + monotonic morsel acquisition) and each source node is scanned in exactly one morsel -> one thread, so per-batch srcRows sets are disjoint and sorted. kwayMergeCSRChunks now consumes the sparse chunks and reconstructs the single dense global indptr (800MB) once, on first consumer request: - single-chunk fast path moves indices/edgeIDs through without copying and synthesizes the indptr from (srcRows, counts) via prefix-sum; - multi-chunk path does a k-way union of the disjoint sorted runs via a min-heap (O(totalSrcRows + totalEdges log numChunks)) instead of the old O(maxSrcRow x numChunks) sweep, with a disjointness invariant check. mergeCSRMetadata (FIXED_ORDER pairwise path) is adapted to consume and produce the sparse shape. The merged CSRMetadata retains the dense indptr + flat indices/edgeIDs, so getCSRArrowArrays(), makeCSRArrowArray (which already aliases the merged vector), and the existing test API (metadata.indptr[src], metadata.indices[idx]) are unchanged. Steady-state memory: ~45GB -> ~7.6GB (no rel rowid) / ~13.6GB (with). Phase 2 (chunked zero-copy indices) will reach the ~6.8GB/12.8GB floor. Design: docs/design_arrow_csr_zero_copy.md
adsharma
changed the title
adsharma
changed the title
Phase 2 of the Arrow CSR zero-copy work. With Phase 1 the steady-state per-batch indptr bloat (45GB) was gone, but combineCSRChunks() still copied the full per-batch index/edgeID vectors (~6GB) into the merge argument and held all per-batch vectors + the merged copy for the duration of the k-way merge. Changes: - combineCSRChunks() now std::moves csrChunks into kwayMergeCSRChunks, avoiding the ~6GB argument copy and leaving csrChunks empty (hasCSRMetadata() stays true via the cached combinedCSR). - kwayMergeCSRChunks frees each per-batch chunk's vectors as soon as its source rows are exhausted by the min-heap, so the transient peak (per-batch + merged) stays low instead of holding both for the whole merge. - Single-chunk path and validation now handle the legacy dense-indptr form (empty srcRows/counts, pre-set indptr) used by test fixtures. - Replaced the brittle pointer-equality test with a value-equivalence check that verifies the exported Arrow buffers hold the merged CSR values without pinning the backing-pointer choice. Reaches the ~6.8GB floor for the 0.75B-edge sorted-by-source scan. True chunked zero-copy (no merged copy at all) requires a table SORTED BY (src) DDL, to be added separately; the merge already validates disjointness and falls back safely if the table is unsorted.
adsharma
force-pushed
the
skip_arrow_sort2
branch
from
7866dbc to
ce63471
Compare
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.
Cuts peak RSS from 45GB to around 16GB.
Details in the design doc.