LoginSign Up
← Back to all sparks
A

Aim

ANALYTICS
Velocity0.0

Open-source experiment tracker for AI and ML training runs

aimstack.io ↗

An experiment tracker grinding on storage performance — and quiet for over a year.

experiment-trackingmlopsstorage-performanceopen-sourcestale-cadence
◆Current state
Aim is an open-source ML experiment tracker whose 3.2x line reads almost entirely as storage and indexing work: constant indexing of in-progress runs, reading from a single unified database, fallbacks when the index is missing, stalled-run detection. The user-facing additions in this window are narrow — a read-only UI mode, report creation, self-signed SSL support, PytorchLightning logger contexts. The most recent entry here is from May 2025, making this feed over a year stale.
◆Where it's heading
The direction across these releases is toward making the local storage layer trustworthy at scale rather than expanding what the tracker does. Repeated fixes around index corruption, empty index.db handling, false-positive metric checks, and session refresh point at users hitting durability problems on long-running or high-volume tracking. Integration surface grows only where contributors push it — S3 client config, Lightning contexts, remote mass updates all arrive as outside contributions rather than a planned roadmap.
◆Prediction
With no release visible in over a year, the honest read is that cadence has stopped rather than shifted; the entries give no signal of a 4.x line or a direction change. If work resumes, the pattern suggests more storage-correctness fixes before any new capability.

◆Recent moves

  1. 1y ago

    🚀 v3.29.1 - Improved query performance by reading from single unified database and constant data indexing, fixes in min/max calculation in UI and jupiter/colab integration

    Constant indexing of in-progress runs and reads from a single unified database are the payoff of the storage thread running through this whole line — live runs become queryable without waiting for a finalize step. Stalled-run detection and a polling file observer patch two ways the tracker previously lost track of reality.

    View source ↗
  2. 1y ago

    🚀 v3.28.0 - Improved performance by removing redundant checks and bypassing runs known to yield false results, new callback for hugging face distributed runs, fixes in Tag duplicates handling, remote tracking exception handling and more, code style improvements.

    Performance work by subtraction: skipping metric checks on runs already known to yield nothing and cutting redundant client keep-alive time. The py.typed marker is small but real for anyone using Aim as a library, exposing existing annotations to type checkers.

    View source ↗
  3. 1y ago

    🚀 v3.27.0 - Enhancements for PytorchLightning logger and S3ArtifactsStorage, fixes for RunStatusReporter, metric aggregations and tag creation from parallel runs

    Integration surface widens at the edges — custom boto3 client parameters for S3 artifact storage and custom contexts in the PytorchLightning logger — both contributed externally. The RunStatusReporter and tag-collision fixes address failures that surface specifically under parallel runs.

    View source ↗
  4. 1y ago

    🚀 v3.24.0 - Support for mass updates in remote tracking, fixes in database error handling and bookmarks page scroll

    A read-only mode for the UI and remote mass updates are the two additions here that change what teams can do with a shared Aim instance. Note the version ordering in this feed is unreliable — 3.24.0 is stamped later than 3.25.x despite preceding it.

    View source ↗
  5. 1y ago

    🚀 v3.25.1 - Fixes in empty index.db handling and python 3.12 builds

    Two fixes: a false corruption marking on empty indexes and Python 3.12 build failures. Both are the kind of durability and packaging papercut that recurs constantly in this line.

    View source ↗
  6. 1y ago

    🚀 v3.25.0 - Reports support, ability to use self-signed SSL certificates

    Report creation is the rare addition here that expands what Aim presents rather than how it stores, and self-signed SSL support unblocks internal-network deployments. Two features in a release is unusually feature-heavy for this line.

    View source ↗