Fix duration minimum for zero-task stages

[amahussein]

Fixes #2151.

A stage that produced no task records is held in memory with numTasks=0 and zeroed extrema, so TaskMetricsAccumRec.finalizeAggregation leaves durationMin=0. Two later folds read that synthetic zero as an observed task duration: the cached-record fold in TaskMetricsAccumRec.addRecord, which builds the job and SQL aggregates, and StageAggTaskMetricsProfileResult.aggregateStageProfileMetric, which merges stage attempts. On the stored gpu_oom_eventlog fixture this made SQL 24 report duration_min=0 while its task-bearing stages had minima of 3085, 4374, and 4760 ms.

Both folds now key on task count rather than value, because 0 remains a legitimate duration for a task that ran. The record fold skips a zero-task record, and the attempt merge takes the populated operand's minimum when exactly one side is empty.

A standalone zero-task stage row also publishes an empty duration_min instead of 0. QualX pools stage rows by SQL and reduces _min columns with pandas min; it recomputes duration_mean from pooled sums but cannot reconstruct a minimum, so retaining numeric 0 would continue feeding the model a false value. For QualX owners: the shipped XGBoost models use duration_min, so predictions can change for affected applications; an all-empty group now supplies a missing value instead of a synthetic 0.

duration_max and duration_avg deliberately keep their zero values on an empty row. Zero never wins a maximum, and the mean is recomputed downstream from pooled sums and counts, so the minimum is the only sentinel that propagates.

Tests cover the stored fixture through SQL aggregation and cover stage-attempt merging with populated, empty-left, empty-right, and both-empty operands. Full verification passed on Scala 2.12 and 2.13.

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge; no actionable correctness, security, or repository-rule issues were identified.

Summary

• Ignores zero-task records when folding cached stage metrics into job and SQL aggregates.
• Merges stage-attempt minima according to task presence while preserving legitimate zero-duration tasks.
• Serializes an undefined minimum as an empty value for zero-task stage rows.
• Adds fixture-level SQL aggregation coverage and stage-attempt merge coverage.

_{Reviews (1) · Last reviewed commit: "Fix duration minimum for zero-task stage..."}