perf(gfql): whole-column aggregates trail a columnar scan at 1M+ nodes (Pokec mgBench)

Aperta
  • Qualcosa non va?
#1,875 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
4/5
Tempo stimato
3-5 giorni
Idoneità per principianti
35/100
Tipo di issue
Bug
Chiarezza
Abbastanza chiara
Stato di attività
Ferma
Stack tecnologico
python
Ambito
performance

Direzione di ricerca

Start with the GFQL Cypher-to-aggregate lowering and profile the four named Pokec mgBench global-aggregate queries at 1.63M nodes, comparing the results with the receipted Kuzu lane. Follow the standard probe → profile pass and record performance receipts to pyg-bench per measurement-rules; done means identifying and addressing the aggregate-path gap, with measurements showing the effect.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

Receipted gap

The four Pokec mgBench global-aggregate queries (not GraphBench qN — these are mgBench's own names/text, run verbatim): aggregate (MATCH (n:User) RETURN n.age, COUNT(*)), aggregate_with_distinct (COUNT(DISTINCT n.age)), aggregate_with_filter (WHERE n.age >= 18 + group), min_max_avg (min/max/avg(n.age)). GFQL best-of-4-engines vs embedded Kuzu, receipted lanes (pyg-bench pokec-gfql-relane-20260813 + pokec-kuzu-20260812):

  • 100k nodes: ~1 ms margins either way (aggregate 0.9x, min_max_avg 0.7x, with_filter TIE, with_distinct GFQL 1.3x)
  • 1.63M nodes: all four cells LOSE 0.2–0.3x — Kuzu's columnar scan stays ~3 ms while GFQL's best arm grows to 7–13 ms

Same lanes: GFQL wins every seeded traversal cell 1.8–22x at both scales, so this is isolated to the whole-column aggregate path.

Hypothesis (unverified — probe before fixing)

A polars/cudf group-by over one int column at 1.6M rows is well under 5 ms native, so the gap smells like per-query overhead in the Cypher→aggregate lowering (row-pipeline materialization or projection width), not kernel speed. Needs the standard probe → profile pass; #1755 (Cypher-surface tax) and #1658 (selective traversal) are the adjacent named levers.

Status

Not release-blocking (losses are honest cells on the scoreboard with this issue as the named lever). Perf receipts to pyg-bench per measurement-rules.

🤖 Generated with Claude Code

Lingua principale
Python
Stelle
2.6k
Fork
229
Merge medio
1g 5h
PR unite (30g)
99

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di graphistry/pygraphistry

Tutte le issue di graphistry/pygraphistry

Issue simili

Altre issue su Python

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.