[weireweire]

Summary

Make timeslice export linear in the number of stored list-metric values. Each slice now reads only its own inter_chunk_latency values from an index built once per export, instead of scanning every stored value. Exported results and timeslices are byte-for-byte unchanged.

Cause

• export_results (src/aiperf/metrics/accumulator.py:767) computes timeslices whenever slice_duration is set. The agentx preset sets it to 1 s.
• _compute_timeslices (accumulator.py:798) loops over every slice (:859) and calls _compute_results for each one (:875).
• _compute_results -> _collect_scalars_and_arrays (:429) -> _collect_one_list_column (:508) -> RaggedSeries.get_values_for_mask (:543).
• get_values_for_mask (src/aiperf/metrics/ragged_series.py:81) builds a boolean mask over all V stored values and gathers from it.
• Total cost is O(T x V) for T slices. Both T and V grow with run length, so the cost grows with its square. Before perf: skip unused live metric computations #53 the same pass also ran on every live stats tick.

Fix

• _build_slice_ragged_index assigns each value its record's slice, stable-sorts once per ragged column, and records each slice's range with searchsorted.
• _collect_one_list_column takes a slice's values as values[order[start:end]]. The stable sort keeps storage order within a slice, so the array matches get_values_for_mask(full_mask) element for element, and sums and percentiles are bitwise equal.
• All other paths, including the overall summary and the t-digest backend, still use the mask path.

Measurements

Saved records from a 3,600 s, concurrency-4096 AgentX run were replayed into MetricsAccumulator: 549,835 records, 161.5 M inter_chunk_latency values, 3,596 one-second slices, one CPU core.

	stock	this PR
export_results (3,596 slices)	1,717 s	141 s
export_results, first 20 K records (171 slices)	3.8 s	2.2 s

The exported results and all timeslices were compared through model_dump(). They are identical at 20 K records and at the full set. In the same run the client spent more than an hour between the end of profiling and the final export.

Tests

• New TestTimesliceRaggedIndex (2 cases: unscoped, and phase-scoped with warmup records). It ingests 400 records with shuffled start times, empty lists, and starts on slice edges. It checks that the timeslice path no longer calls get_values_for_mask and that the export is byte-identical to the per-slice mask path.
• tests/unit: 17,659 passed and 2 failed. Both failures also fail on master (test_completed_excludes_analyzer_metrics, test_every_numeric_field_has_bounds).
• pre-commit run on the changed files passes, including check-ergonomics and check-ruff-baselined.

This complements #55: if timeslices become opt-in, runs that opt in still get a linear export.

---

Note

Medium Risk
Changes hot-path timeslice aggregation math for list metrics; behavior is intended to be identical but regressions would affect exported percentiles and sums on long AgentX runs.

Overview
Timeslice export for replay-capable ragged metrics (e.g. inter_chunk_latency) no longer scans all stored list values on every slice. Once per _compute_timeslices run, _build_slice_ragged_index buckets each value by its record’s time bin (stable sort + searchsorted ranges). While each slice is computed, _collect_one_list_column gathers that slice via values[order[start:end]] instead of RaggedSeries.get_values_for_mask, with transient _slice_ragged_index / _active_slice_bin cleared in a finally block. The per-slice loop moves into _compute_timeslice_bins; overall summary and non-replay backends still use the mask path.

Tests add TestTimesliceRaggedIndex: randomized ingest checks that timeslice export barely calls get_values_for_mask and matches byte-identical JSON to the legacy mask path (with and without phase scoping).

^{Reviewed by Cursor Bugbot for commit e2569e3. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

Try out this PR

Quick install:

pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@e2569e3d128479944d49b7c454d7fcb6dfd671b9

Recommended with virtual environment (using uv):

uv venv --python 3.12 && source .venv/bin/activate
uv pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@e2569e3d128479944d49b7c454d7fcb6dfd671b9

---

Last updated for commit: e2569e3 • Browse code

[Oseltamivir]

Thanks for the contribution!