jrgemignani opened a new pull request, #2424:
URL: https://github.com/apache/age/pull/2424
Targets the agtype value-tree allocation cost that dominates LDBC IC4 and
other sort/comparator-bound queries. Two complementary optimizations (A1 arena,
A2 fast-id) plus the sanitizer issues uncovered while verifying the change.
A1 — per-tuple agtype arena (Pattern B applied):
Add agt_arena_begin / AGT_ARENA_BEGIN_SHARED / agt_arena_reset /
agt_arena_end helpers (agtype_util.c, agtype.h). Two usage patterns:
disposable per-call arenas for one-shot tree builds, and a long-lived
shared arena (file-static, parented under CacheMemoryContext) for hot
inner loops that would otherwise pay AllocSetCreate cost per call.
Apply to compare_agtype_scalar_containers: when the comparator
deserializes VERTEX/EDGE/PATH composites via fill_agtype_value_no_copy,
route those allocations into a shared arena and reset (O(1)) at the
end of the comparison instead of recursively pfree'ing the tree (O(N)).
Master IC4 spent 17.64% of total CPU in pfree_agtype_value_content from
this comparator; the arena collapses that walk to one MemoryContextReset.
A2 — binary-direct id extraction:
When compare_agtype_scalar_containers compares two VERTEX or two EDGE
values, only the graphid 'id' determines order
(compare_agtype_scalar_values
ignores all other fields). Master IC4 spent 73.48% inclusive in
ag_deserialize_composite called from this comparator, building the entire
agtype_value tree (label, properties, nested values) just to read one
int64.
Add extract_composite_id_fast: walks the binary VERTEX/EDGE container
directly and reads the int64 id at field index 0 (always 'id' due to
length-sorted key ordering established by uniqueify_agtype_object — the
same invariant PR #2302 leveraged for direct array indexing).
Wired into compare_agtype_scalar_containers as a fast path before the
existing arena+slow path. Falls through on cross-type comparisons
(VERTEX vs EDGE), PATH, and malformed extended-type entries.
ASAN/UBSan fixes (10 issues found while verifying A1+A2 under
-fsanitize=address -fsanitize=undefined; 1 introduced by A2, 9 pre-existing).
The 11th finding (uninitialized parent_cpstate fields in cypher_analyze.c) was
independently fixed upstream by PR #2423; this branch is rebased on top of that
fix and no longer carries a duplicate change.
- HIGH: heap-buffer-overflow in agtype_raw.c:write_container — VARSIZE
included the 4-byte varlena header but copy started past it, reading VARHDRSZ
bytes past source allocation. Subtract VARHDRSZ. ASan flagged 4 times in
installcheck.
- 7 unaligned int64/float8 reads/writes (AGT_HEADER is 4 bytes, leaving
payload 4-byte- but not 8-byte-aligned). UB under strict-alignment rules.
Replaced typed loads/stores with memcpy in agtype_ext.c (4 sites),
agtype_util.c (3 sites incl. extract_composite_id_fast), agtype_raw.c
(write_graphid).
- agtype.c:agtype_hash_cmp: reading r->val.array.raw_scalar at
WAGT_END_ARRAY where the iterator does not populate *val. Replaced with
explicit raw_scalar bool stack tracked across BEGIN/END pairs. Hash output
unchanged.
- agtype_util.c:copy_to_buffer: memcpy(NULL, NULL, 0) is UB. Guard with if
(len > 0).
Performance (SF3 LDBC, paired same-session vs master):
IC sum: 201,930 -> 157,129 ms = -22.19%
IC4: 50,570 -> 7,615 ms = -84.94%
IC5: 21,779 -> 20,646 ms = -5.20%
IS, IU: within noise (<= +/-3% on sub-second queries)
Tests:
34/34 installcheck on production PG18.3
34/34 installcheck on ASAN+UBSan PG18.3 (zero unique runtime errors,
zero heap-buffer-overflows; the original audit found 10 unique
runtime issues plus 1 HBO with 4 instances, all addressed here
or upstream)
Co-authored-by: Claude <[email protected]>
modified: src/backend/utils/adt/agtype.c
modified: src/backend/utils/adt/agtype_ext.c
modified: src/backend/utils/adt/agtype_raw.c
modified: src/backend/utils/adt/agtype_util.c
modified: src/include/utils/agtype.h
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
