[weireweire]

Summary

Keep the AgentX copies of #58 and #49 identical to their upstream form in ai-dynamo/aiperf, so later upstream syncs apply cleanly. This PR tracks ai-dynamo#1592 and will be updated whenever that PR changes.

Changes

1. Timeslice ragged index: skip records ingested after the export snapshot (follow-up to #58)

Same commit as the review fix on ai-dynamo#1592.

• export_results runs in a worker thread while ingestion can continue. _build_slice_ragged_index indexed record_bin (sized at the snapshot) with every current record index, so a ragged value appended for a later record raised IndexError and aborted the export.
• Record indices beyond the snapshot now map to no slice, as get_values_for_mask already does. Value positions are unchanged, so exported timeslices are unchanged.
• Drops the redundant _compute_timeslice_bins docstring.

2. Timeslice ragged index: no shared per-computation state (follow-up to #58)

Same commit as the second review fix on ai-dynamo#1592.

• perf(metrics): gather timeslice ragged values from a per-slice index #58 stored the slice index and the active slice on the accumulator while the export's slice loop ran. export_results runs in a worker thread and the realtime summarize runs on the event loop with no lock, so a summary that overlapped the slice loop gathered inter_chunk_latency from the export's current slice instead of its own mask.
• The index is now a local of _compute_timeslices, and each slice's value positions are passed to _compute_results as an argument. The _compute_timeslice_bins split from perf(metrics): gather timeslice ragged values from a per-slice index #58 is undone; the slice loop is back to its original form plus that argument.
• Index trim: int32 slice ids and positions, and slice bounds from bincount + cumsum. At 20 M values: 422 MB peak and 80 MB retained, down from 504 MB and 160 MB.
• The only difference from upstream is in the new overlap test, which calls _summarize_for_export_context(..., include_timeslices=False) as AgentX's realtime summarize does (perf: skip unused live metric computations #53); upstream has no include_timeslices argument.

3. Warmup baseline tagging: align #49 with upstream

Upstream fixes the same bug with phase_kind="warmup" on the synthesized agentic warmup alone (ai-dynamo#1325). #49 also added a PhaseBaselineRequestMessage.phase fallback for boundaries without a kind. That fallback cannot trigger here: validated PhaseConfigs always carry a kind, and the agentic warmup now sets one.

• Keeps phase_kind="warmup" and its test assertion.
• Removes the phase message field, the publisher wiring, the manager fallback, their tests, and the added comments. The scrape tagging is now identical to upstream: phase_kind == "warmup" -> warmup, otherwise profiling.

Tests

• New test_summary_overlapping_timeslice_export_keeps_its_own_values: a phase-scoped summary run in the middle of the export's slice loop matches its standalone result, and so does the export. On the commit before this fix the summary's inter_chunk_latency has 54 values (one warmup slice) instead of 6,849.
• New test_ragged_values_ingested_after_snapshot_are_excluded: appends ragged values for a record past the snapshot while the index is built; timeslices match the export without them. It raises IndexError without the fix.
• tests/unit/{post_processors,metrics,exporters,server_metrics,timing,records}: 3817 passed, 10 failed. The same 10 tests fail on master (test_completed_excludes_analyzer_metrics, test_auto_offset_affects_execution_duration x6, test_global_idle_cap_rechecks_after_barrier_defers_near_timer, test_offset_child_uses_shared_scheduler_for_idle_cap, test_produces_intermediate_values).
• ruff check passes on the changed files.

[github-actions]

Try out this PR

Quick install:

pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@a75c9f3c83e8795a4d1bbbf0cf2ee05c4abbf182

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@a75c9f3c83e8795a4d1bbbf0cf2ee05c4abbf182

---

Last updated for commit: a75c9f3 • Browse code