Add additional support for autonomous ML experiments#521
Open
Saba9 wants to merge 25 commits into
Open
Conversation
…ers, run status, structured alerts
Adds agent-facing query APIs and tooling to make Trackio the definitive tracker
for autonomous ML experiments driven by AI coding agents.
New CLI commands:
- `trackio best` — rank runs by metric, return winner + leaderboard
- `trackio compare` — side-by-side run comparison across metrics
- `trackio summary` — full experiment overview with status/configs/metrics
New features:
- Run status tracking (running/finished/failed) with automatic lifecycle mgmt
- Structured alert data (`data={}` param on `trackio.alert()`)
- Metric watchers (`trackio.watch()`) for auto NaN/spike/stagnation detection
- `trackio.should_stop()` for training loop early stopping
- Python API: `run.metrics()`, `run.history()`, `run.summary`, `run.status`
Test infrastructure:
- Synthetic training simulator (no ML deps, runs in seconds)
- Agent test runner with 5 experiment types
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
Resolve merge conflicts integrating autonomous ML features with main branch changes (run_id support, RemoteClient, query command, OAuth, error handling). Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
Contributor
🪼 branch checks and previews
|
Contributor
🦄 change detectedThis Pull Request includes changes to the following packages.
|
🪼 branch checks and previews
Install Trackio from this PR (includes built frontend) pip install "https://huggingface.co/buckets/trackio/trackio-wheels/resolve/e58200b25c87922e98bbc1241009e5f2c26848eb/trackio-0.25.1-py3-none-any.whl" |
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
Fixes:
- Fix 1: Run status "failed" no longer overwritten by "finished" —
finish() now accepts a status parameter, _cleanup_current_run passes
status="failed" directly
- Fix 2: Watcher patience supports both min and max mode via new
mode parameter (was hardcoded to minimization only)
- Fix 3: Replace dead --minimize/--maximize flags with --direction
{min,max} on trackio best
- Fix 5: Watcher _values now uses deque(maxlen=window) to bound memory;
alert dedup via per-condition flags that reset on return to normal
- Fix 6: Warn if watchers exist when init() clears them; docstring
documents ordering requirement
- Fix 7: Drop Api.Run.summary cache — recompute on each access
- Fix 9: set_run_status/get_run_status now accept and use run_id,
INSERT handles run_id NOT NULL column in new schema
- Fix 12: Remove unused enumerate variable in agent_runner
Tests:
- test_watchers.py: 18 tests covering nan, spike, max/min threshold,
dedup, patience min/max mode, window bounds, manager propagation
- test_run_status.py: 6 tests covering running→finished, failed status,
idempotent finish, Api.Run.status, multi-run status
- test_cli_agent_commands.py: 10 tests covering best/compare/summary
in JSON and human-readable modes, error cases
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
Saba9
changed the title
…ests - Watchers: 18 → 9 tests by merging nan+inf, combining dedup with trigger tests, merging patience modes - Run status: 6 → 4 tests by dropping idempotent (covered by overwrite) - CLI: 10 → 4 tests by seeding project once via module fixture, dropping human-readable format tests, merging maximize/subset into main tests Total: 34 → 17 tests, 494 → 336 lines, same code path coverage. CLI tests 3x faster (7.5s vs 25s) from shared fixture. Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
Collaborator
|
I had Claude analyze it and write up some actionable feedback. At first I was going to go through and leave comments for each item in the source code, but I realized it's probably more efficient to just include the Markdown file here. Note that Claude considers two issues blockers: |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Short description
Apologies for the large PR. About half the additions are from tests and docs.
This PR adds features designed for use in AI-driven training loops where an agent or automated script needs to monitor metric health, compare runs, and query results
programmatically without human supervision.
Metric watchers (
trackio.watch()/trackio.should_stop()): Register rules upfront and have Trackio fire alerts automatically on everytrackio.log()call. Supportedconditions:
NaN/Infdetection, max/min threshold breachesNstepsfncallback:trackio.watch("loss", fn=my_check)Any condition that warrants stopping sets
trackio.should_stop()toTrue, which training loops can poll to exit early. Custom conditions can signal early stopping by including"stop": Truein a returned alert dict.Run status tracking: Runs now record a lifecycle status (
running→finished/failed) in SQLite. Status is set torunningoninit(),finishedonfinish(), andfailedon unexpected process exit. Thebest,compare, andsummarycommands default to finished runs only;--include-alloverrides this.New CLI commands:
trackio best --project X --metric Y [--direction min|max] [--mode last|min|max]— find the best run for a metrictrackio compare --project X [--runs ...] [--metrics ...]— side-by-side run comparisontrackio summary --project X [--runs ...]— per-run metric summary tablePython API additions on
Run:status,final_metrics,metrics(),history().AlertReasonconstants: Every watcher-generated alert includes adata["reason"]field matching one of thetrackio.AlertReasonconstants, allowing agents to identify alert types programmatically. Custom conditions that returnTrueuseAlertReason.CUSTOM.AI Disclosure
Type of Change
Related Issues
Closes:
Testing and linting
python -m pytest ruff check --fix --select I && ruff format