-
Type:
Task
-
Resolution: Duplicate
-
Priority:
Major - P3
-
None
-
Affects Version/s: None
-
Component/s: None
-
None
-
Storage Execution
-
None
-
None
-
None
-
None
-
None
-
None
-
None
Summary
After the implementation sub-tasks land, run a sys-perf patch comparing the knob off/on (the index-build write path is hot code, so a sys-perf patch runs before merge regardless), then flip primaryDrivenIndexBuildIncrementalSorterSpilling to default on.
Success criteria
- Scan-phase flow-control metrics match the POC results: ~−98% writes queued by flow control, ~−40% standby lag, reduced isLagged engagement
- Peak indexBulkBuilder.memUsage stays within the memory budget (the POC ran +46% over; the double buffer should bound it)
- No foreground p50 regression beyond noise (the POC showed +10%, unexplained; this run re-checks it)
References
- POC results table and method: design doc (POC section)
- Existing perf rig: build_index_with_side_writes_locust on disagg-m8g-perf-11-node.arm.aws, 3 runs per arm; scripts and raw data in /home/ubuntu/scratch/server132056_perf/