Endpoint property projections leave node-record reads unbatched
perfloop-oss/helix-db · Under-Batching
https://perfloop.ai/t/oss/case_vqzxhf17zy
Verdict
VERIFIED · settled 2026-11-14 · pull request opened as HelixDB/helix-db#1300
What happened: The paired measurements met the required improvement.
Hypothesis
Native Values projections of endpoint properties batch the edge metadata but still await endpoint-node records one by one. When a batch contains many distinct endpoints, independent reads remain serialized. The local read pattern is confirmed, while elapsed-time savings, storage-cache conditions and production exposure are unknown.
project_values calls prefetch_rows, whose PropertySource::EdgeEndpoint branch collects only edge IDs. It does not collect the nodes referenced by those endpoints. RowValueResolver::element_property subsequently resolves each endpoint and calls element_row for the node; an uncached node reaches load_blob and property_bytes as an individual get. Ordinary current-element records are already batched. For a successfully completed batch of R edge rows selecting stored fields from both endpoints, this can leave as many as 2R individual node-record reads when every endpoint is distinct and not otherwise cached. Shared nodes reduce that count. Missing endpoints and identifier-only paths do not read node properties, and early errors can stop evaluation sooner.
The successful provisional probe in measurement receipt call_Oqq4P3MyJrFXyJzwibat59oY, output:1, ran `cargo test -p db --lib perfloop_projection_probe -- --nocapture --test-threads=1`. Values over 16 fixture-created edges with 32 distinct endpoint nodes and two endpoint-name selections produced 32 logical raw gets plus 16 multi-get keys. A comparator prefetched the known fixture node IDs through the existing resolver, returned the same values and produced zero raw gets plus 48 multi-get keys. Total requested keys were unchanged; the opportunity is grouping and overlap, not eliminating necessary data. These counters do not measure physical block reads or network round trips, and the comparator is not a complete repair.
Endpoint-property projection is documented in docs/database/helix-db/query-guides/projections.mdx:116-127. The landed batching implementation, https://github.com/HelixDB/helix-db/commit/ae6cdc2475c75fa824911db94c55fbd44e7ff68a, batches current-element records and edge endpoints but leaves the dependent node-record phase on demand. This is a distinct extension of that implementation. The bounded open-PR and landed-history checks did not identify active ownership of this phase. The README's Discord community was not accessible through the attempted public invite read, so off-host discussion remains a named evidence gap.
The first useful Case check is a provisional two-stage batch lookup driven by actual endpoint metadata, followed by native Values queries with distinct, shared, absent and corrupt endpoints. Compare logical read grouping and elapsed time, including warm-cache overhead. Error ordering, deadlines and memory-budget behavior need explicit comparison because eager node loading can change when failures surface. A failure to remove individual reads, unacceptable semantic changes or no useful end-to-end saving can reject the proposed move.
Change to test: Once edge endpoints are known, fetch the endpoint-node records required by selected property fields in bounded, deduplicated batches rather than individual reads. Preserve identifier-only and missing-endpoint skips and establish compatibility with the existing error, deadline and resource-accounting contracts.
Where it lives
perfloop-oss/helix-db · crates/db/src/execution/interpreter/stream/projection/dispatch.rs
Evidence
Endpoint Values on property-path data · 17 sample pairs
| metric | baseline | candidate | paired median change | confidence range | required | result |
|---|---|---|---|---|---|---|
ns/query |
16100 |
11940 |
−24.9% (−4010) |
−5220 to −2060 |
gain > 0; gain > 0.05 × |baseline median|; arm medians separated |
PASSED |
allocated_bytes/query |
47610 |
41860 |
−12.1% (−5750) |
−5750 to −5750 |
loss ≤ 0.05 × |baseline median| |
PASSED |
process_cpu_ns |
380000000 |
300000000 |
−21.1% (−80000000) |
−130000000 to −40000000 |
loss ≤ 0.05 × |baseline median| |
PASSED |
peak_rss_bytes |
57917440 |
58044416 |
−184320 |
−1716224 to +847872 |
loss ≤ 0.05 × |baseline median| |
PASSED |
Endpoint Values with distinct nodes · 18 sample pairs
| metric | baseline | candidate | paired median change | confidence range | required | result |
|---|---|---|---|---|---|---|
ns/query |
3090500 |
1657000 |
−46.1% (−1424000) |
−1656000 to −1240000 |
gain > 0; gain > 0.05 × |baseline median|; arm medians separated |
PASSED |
allocated_bytes/query |
4788000 |
2057000 |
−57% (−2730000) |
−2731000 to −2730000 |
loss ≤ 0.05 × |baseline median| |
PASSED |
process_cpu_ns |
2600000000 |
1485000000 |
−43.1% (−1120000000) |
−1380000000 to −1000000000 |
loss ≤ 0.05 × |baseline median| |
PASSED |
peak_rss_bytes |
68997120 |
68859904 |
−145408 |
−2027520 to +1363968 |
loss ≤ 0.05 × |baseline median| |
PASSED |
Endpoint Values on ring data · 18 sample pairs
| metric | baseline | candidate | paired median change | confidence range | required | result |
|---|---|---|---|---|---|---|
ns/query |
38665 |
28995 |
−24.8% (−9605) |
−11630 to −6910 |
loss ≤ 0.05 × |baseline median| |
PASSED |
allocated_bytes/query |
96870 |
79790 |
−17.6% (−17080) |
−17080 to −17080 |
loss ≤ 0.05 × |baseline median| |
PASSED |
Endpoint Values with shared nodes · 54 sample pairs
| metric | baseline | candidate | paired median change | confidence range | required | result |
|---|---|---|---|---|---|---|
ns/query |
966650 |
967500 |
+5500 |
−20700 to +31200 |
loss ≤ 0.05 × |baseline median| |
PASSED |
allocated_bytes/query |
1610000 |
1609000 |
−0.1% (−1000) |
−1000 to −1000 |
loss ≤ 0.05 × |baseline median| |
PASSED |
process_cpu_ns |
900000000 |
910000000 |
+10000000 |
−30000000 to +40000000 |
loss ≤ 0.05 × |baseline median| |
PASSED |
peak_rss_bytes |
62969856 |
62887936 |
−202752 |
−1458176 to +765952 |
loss ≤ 0.05 × |baseline median| |
PASSED |
Single stored endpoint field · 18 sample pairs
| metric | baseline | candidate | paired median change | confidence range | required | result |
|---|---|---|---|---|---|---|
ns/query |
17575 |
15440 |
−12.3% (−2160) |
−3800 to −1470 |
loss ≤ 0.05 × |baseline median| |
PASSED |
allocated_bytes/query |
63900 |
63910 |
+10 |
+10 to +10 |
loss ≤ 0.05 × |baseline median| |
PASSED |
Single endpoint ID · 48 sample pairs
| metric | baseline | candidate | paired median change | confidence range | required | result |
|---|---|---|---|---|---|---|
ns/query |
13755 |
13710 |
−60 |
−790 to +590 |
loss ≤ 0.05 × |baseline median| |
PASSED |
allocated_bytes/query |
56830 |
56860 |
+0.1% (+30) |
+30 to +30 |
loss ≤ 0.05 × |baseline median| |
PASSED |
Single node field · 33 sample pairs
| metric | baseline | candidate | paired median change | confidence range | required | result |
|---|---|---|---|---|---|---|
ns/query |
13920 |
13940 |
+20 |
−1330 to +600 |
loss ≤ 0.05 × |baseline median| |
PASSED |
allocated_bytes/query |
57580 |
57610 |
+0.1% (+30) |
+30 to +30 |
loss ≤ 0.05 × |baseline median| |
PASSED |
Checks: 5 of 5 passed. Verification: no defect found.
Timeline
2026-11-14· Case opened2026-11-14· PR opened