Avoid excessive malloc/free #1291

Activity

1. adsharma

   added a parent issue on Oct 17, 2026

   • Fix per query constant overhead #1290
2. adsharma commented on Oct 17, 2026

   

   adsharma

   on Oct 17, 2026

   ContributorAuthor

   More actions

   @zahariash

   Ground truth (uprobes, warm @prrao87 LDBC Q1): main 3,021 → 2,288/exec with everything merged. The remaining ~2,300 break down roughly as:
   per-task operator clones (~150–250, spread across print infos/schemas/evaluators), ValueVector/scan-state setup+teardown (~100),
   txn machinery (~50), FactorizedTable result assembly, ArtKey::encode temps, hash-table builds.

   Attempted and reverted: per-task replica pooling. ProcessorTask::run clones its sink subtree per task (~6 pipelines × subtree).
   Pooling it crashed intermittently (native heap corruption in result strings) — pooled task operators inherit construction-time
   state (TableScanState positions, per-op vectors) that init* doesn't fully reset, and auditing every operator is a bigger, riskier
   project. Reverted cleanly; init hardening stays as its foundation. Details + rationale are on the PR so nobody re-tries it blind.

   👍React with 👍1Reacted by zahariash
3. adsharma

   changed the title [-]avoid excessive malloc/free[/-] [+]Avoid excessive malloc/free[/+] on Oct 17, 2026