Latest Results
fix(metrics): forward peak_state_bytes through distributed pipeline nodes
Several distributed pipeline nodes (Sort, IntoBatches, Limit,
StageCheckpointKeys) define their own RuntimeStats impls that
destructure StatSnapshot::Default for phase-aware row/byte accounting.
None of them forwarded the new peak_state_bytes field, so peak memory
reported by stateful local sinks (Sort, Aggregate, TopN, Dedup) was
silently dropped at the distributed boundary. On Flotilla this
surfaced as a "-" in the dashboard's Peak Memory column even for the
distributed Sort node whose final-sort phase wraps the local Sort sink.
Adds a forward_default_snapshot_peak helper on BaseCounters that the
four affected handlers call after their own bookkeeping. Distributed
nodes that use the catch-all DefaultRuntimeStats (Aggregate, TopN,
Distinct, etc.) were already covered by the previous commit.
Extends the existing distributed Sort regression test to assert
peak_state_bytes propagates to the exported snapshot.
https://claude.ai/code/session_012tEp7R8QPzkK16NmBbD62dclaude/review-memory-instrumentation-mV5Sb Latest Branches
+10%
chris/fix-it-iceberg-setup +11%
0%
claude/review-memory-instrumentation-mV5Sb © 2026 CodSpeed Technology