Repository navigation

Add kumo.job() for scheduled job scripts - #70

Closed

Vedaant-J wants to merge 123 commits into

vedaant/mlops-mlflow-modelfrom

vedaant/mlops-cli

Closed

Add kumo.job() for scheduled job scripts#70
Vedaant-J wants to merge 123 commits into
vedaant/mlops-mlflow-modelfrom
vedaant/mlops-cli

Commits on Oct 22, 2026
- Read the holdout of the published model, not the last-evaluated one
  
  Vedaant-J
  and
  claude
  committed
  5069de2
- Expose published_fit_mode on the artifact test's fake trainer
  
  Vedaant-J
  and
  claude
  committed
  ce86b46
- Record a snapshot manifest next to every graph index
  
  Vedaant-J
  and
  claude
  committed
  c1586bd
- Address snapshot manifest review feedback
  
  Vedaant-J
  and
  claude
  committed
  7d6ecbe
- Match the backfill diff when no build commit is recorded
  
  Vedaant-J
  and
  claude
  committed
  2e2b01d
- Record each source table's full location in the snapshot manifest
  
  Vedaant-J
  and
  claude
  committed
  244b5fe
- Drop graph_id and data_version from the snapshot manifest; keep the diff private
  
  Vedaant-J
  and
  claude
  committed
  9eff66d
- Tighten manifest schema checks and scheduler context
  
  Vedaant-J
  and
  claude
  committed
  aeb0b22
- Drop snapshot-manifest tests of private helpers
  
  Vedaant-J
  and
  claude
  committed
  16b844f
- Trim snapshot manifest PR to essentials
  
  Vedaant-J
  and
  claude
  committed
  048473f
- Reuse existing helpers in the snapshot manifest
  
  Vedaant-J
  and
  claude
  committed
  0ecb465
- Record DiskGraph's version and build commit in manifests
  
  Vedaant-J
  and
  claude
  committed
  2552bc1
- Type source versions and allow location-free snapshot manifests
  
  Vedaant-J
  and
  claude
  committed
  8b16cd1
- Pick the source-location lookup by connector type
  
  Vedaant-J
  and
  claude
  committed
  482b1b2
- Record Custom SQL text by default; one record_sources switch
  
  Vedaant-J
  and
  claude
  committed
  0d73b76
- Stamp the engine build commit into wheels; type directory versions
  
  Vedaant-J
  and
  claude
  committed
  d2cd1c0
- Carry the record_sources choice on GraphSnapshot
  
  Vedaant-J
  and
  claude
  committed
  72e2932
- Record Custom SQL's namespace and normalized fingerprint
  
  Vedaant-J
  and
  claude
  committed
  60e1ab6
- Make the snapshot manifest a frozen dataclass
  
  Vedaant-J
  and
  claude
  committed
  915e10c
- Record training-table lineage across export and reload
  
  Vedaant-J
  and
  claude
  committed
  bd39c7c
- Digest exported parts at the source for training-table lineage
  
  Vedaant-J
  and
  claude
  committed
  e9778ae
- Record a custom table's source instead of re-reading its export
  
  Vedaant-J
  and
  claude
  committed
  0a49692
- Record a custom table's source in the snapshot manifest's shape
  
  Vedaant-J
  and
  claude
  committed
  cf7fea2
- Keep sidecar writes best-effort; version custom tables from their read
  
  Vedaant-J
  and
  claude
  committed
  209e6ee
- Trim training-table lineage to essentials
  
  Vedaant-J
  and
  claude
  committed
  80807de
- Read row counts and source mode from DiskGraph's manifest
  
  Vedaant-J
  and
  claude
  committed
  fcde71e
- Record a custom table's source as the snapshot's manifest does
  
  Vedaant-J
  and
  claude
  committed
  bba451d
- Export a training table even when its manifest sidecar fails
  
  Vedaant-J
  and
  claude
  committed
  aeab584
- Follow the snapshot's record_sources choice
  
  Vedaant-J
  and
  claude
  committed
  370351b
- Expect the listing kind for a reloaded export directory
  
  Vedaant-J
  and
  claude
  committed
  f266ec5
- Record a custom table's source only when the snapshot did
  
  Vedaant-J
  and
  claude
  committed
  23f57a2
- Make the training-table manifest a frozen dataclass
  
  Vedaant-J
  and
  claude
  committed
  f7c62c3
- Record bundle_id and fit provenance in every model bundle
  
  Vedaant-J
  and
  claude
  committed
  d867b57
- Address bundle provenance review feedback
  
  Vedaant-J
  and
  claude
  committed
  4ced958
- Read Trainer.bundle_id defensively
  
  Vedaant-J
  and
  claude
  committed
  51c767a
- Record the published model's fit mode in provenance.json
  
  Vedaant-J
  and
  claude
  committed
  8047610
- Mirror the training table's source in bundle provenance
  
  Vedaant-J
  and
  claude
  committed
  f03673d
- Trim provenance.json to facts no other bundle file holds
  
  Vedaant-J
  and
  claude
  committed
  5f3dbef
- Record fit_mode and source versions in provenance; expose it on Trainer
  
  Vedaant-J
  and
  claude
  committed
  777bdb4
- Trim bundle provenance to essentials
  
  Vedaant-J
  and
  claude
  committed
  6eab1d6
- Reuse shared helpers for bundle provenance
  
  Vedaant-J
  and
  claude
  committed
  1321f3c
- Compute a fresh bundle's id from its publication manifest
  
  Vedaant-J
  and
  claude
  committed
  ef4485c
- Copy a custom table's full source entry into provenance
  
  Vedaant-J
  and
  claude
  committed
  fd6e8d7
- Store the bundle id in provenance.json
  
  Vedaant-J
  and
  claude
  committed
  e0ad8ac
- Make bundle provenance a frozen dataclass
  
  Vedaant-J
  and
  claude
  committed
  4274048
- Add GraphSnapshot export and load through object storage
  
  Vedaant-J
  and
  claude
  committed
  160f8a6
- Ship the object-store sink in the public connector image
  
  Vedaant-J
  and
  claude
  committed
  62b034b
- Report source changes since a snapshot and warn on stale loads
  
  Vedaant-J
  and
  claude
  committed
  4cd042e
- Match source changes on the recorded connector and location
  
  Vedaant-J
  and
  claude
  committed
  dee6bb2
- Keep an existing export when a rebuilt snapshot is re-exported
  
  Vedaant-J
  and
  claude
  committed
  59d62fa
- Pass allow_delayed_open to S3 only when pyarrow supports it
  
  Vedaant-J
  and
  claude
  committed
  ef1a4ad
- Trim comments and implementation-detail tests
  
  Vedaant-J
  and
  claude
  committed
  a58e325
- Trim snapshot export PR to essentials
  
  Vedaant-J
  and
  claude
  committed
  4c9f341
- Move snapshot export checksums into DiskGraph and reuse its manifest
  
  Vedaant-J
  and
  claude
  committed
  6ca9017
- Check loaded snapshots with what DiskGraph already stamps
  
  Vedaant-J
  and
  claude
  committed
  f8563b4
- Derive GraphSnapshot.graph_key from the snapshot id
  
  Vedaant-J
  and
  claude
  committed
  a51952c
- Make snapshot export atomic and pass the snapshot manifest through
  
  Vedaant-J
  and
  claude
  committed
  660c41b
- Drop the S3 upload tuning from snapshot export
  
  Vedaant-J
  and
  claude
  committed
  b583a96
- Repair files a concurrent export wiped before committing
  
  Vedaant-J
  and
  claude
  committed
  85676da
- Compare directory listings in source_changes
  
  Vedaant-J
  and
  claude
  committed
  f37a3bd
- Drop the connector-image allowlist entry for the removed sink test
  
  Vedaant-J
  and
  claude
  committed
  c82556a
- Raise ConcurrentExportError when another export keeps replacing the destination
  
  Vedaant-J
  and
  claude
  committed
  48f5796
- Read snapshot manifests as dataclasses in export and source_changes
  
  Vedaant-J
  and
  claude
  committed
  b6394e2
- Add PredictiveQuery.from_snapshot for graph-free scoring
  
  Vedaant-J
  and
  claude
  committed
  d4272b5
- Expose Trainer.training_query and keep the graph-bound snapshot check
  
  Vedaant-J
  and
  claude
  committed
  eb0f0db
- Trim comments and implementation-detail tests
  
  Vedaant-J
  and
  claude
  committed
  568ce32
- Show where a graph-free job gets its snapshot
  
  Vedaant-J
  and
  claude
  committed
  3df8621
- Read the native graph and graph id from GraphSnapshot
  
  Vedaant-J
  and
  claude
  committed
  8d23970
- Add Trainer.predict_snapshot for one-call graph-free scoring
  
  Vedaant-J
  and
  claude
  committed
  975d144
- Distinguish graph-free scoring in the model reload docs
  
  Vedaant-J
  and
  claude
  committed
  2c619ea
- Test graph-free scoring of an exported snapshot end to end
  
  Vedaant-J
  and
  claude
  committed
  2cc4e5c
- Warn when predict_snapshot scores another graph's snapshot
  
  Vedaant-J
  and
  claude
  committed
  49bae78
- Report fit progress events to callbacks and JSON-lines stdout
  
  Vedaant-J
  and
  claude
  committed
  6bdd795
- Bind the trial error name before the progress-event closure
  
  Vedaant-J
  and
  claude
  committed
  4c09c57
- Drop the library JSON progress option; callbacks cover it
  
  Vedaant-J
  and
  claude
  committed
  bcc7503
- Report trial_ended only once the trial is recorded as succeeded
  
  Vedaant-J
  and
  claude
  committed
  936f289
- Carry the error type and a short message on failure events
  
  Vedaant-J
  and
  claude
  committed
  eebc54a
- Trim internal docstrings and implementation-detail test checks
  
  Vedaant-J
  and
  claude
  committed
  4e7502b
- Trim fit progress events to essentials
  
  Vedaant-J
  and
  claude
  committed
  c1b3e24
- Keep fit_started best-effort when the inputs are malformed
  
  Vedaant-J
  and
  claude
  committed
  18e4d3f
- Reuse the progress logger, searcher and instrumentation for fit events
  
  Vedaant-J
  and
  claude
  committed
  8197f42
- Never fail a fit when an epoch event cannot be built
  
  Vedaant-J
  and
  claude
  committed
  6b5f8eb
- Record Trainer.fit in MLflow with opt-in fit(tracking=)
  
  Vedaant-J
  and
  claude
  committed
  6dea3f7
- Move tracking names into core run records; thin MLflow adapter
  
  Vedaant-J
  and
  claude
  committed
  0dfa5d6
- Keep record building best-effort; declare MLflow's Lightning step axis
  
  Vedaant-J
  and
  claude
  committed
  8473c1d
- Tag model links in core; log per-epoch steps and finish tags in MLflow
  
  Vedaant-J
  and
  claude
  committed
  24bc0ef
- Drop the removed progress setting from the MLflow test fixture
  
  Vedaant-J
  and
  claude
  committed
  e61dc62
- Record run kind and failures as tags; keep ids out of params
  
  Vedaant-J
  and
  claude
  committed
  ef0ed18
- Move the tracking adapter API to kumo.engine.tracking
  
  Vedaant-J
  and
  claude
  committed
  a8317fb
- Stop calling a tracker for the rest of the fit after it raises
  
  Vedaant-J
  and
  claude
  committed
  00987ae
- Trim tracking docstrings and tests of internals
  
  Vedaant-J
  and
  claude
  committed
  9972e15
- Trim MLflow tracking to essentials
  
  Vedaant-J
  and
  claude
  committed
  18a973d
- Build tracking records once in the session; adapters get records
  
  Vedaant-J
  and
  claude
  committed
  5d1c1a5
- Keep experiment tracking off the training thread
  
  Vedaant-J
  and
  claude
  committed
  8d829a7
- Drop the training table id from fit runs
  
  Vedaant-J
  and
  claude
  committed
  469742d
- Record pointers, not lineage, on tracked fit runs
  
  Vedaant-J
  and
  claude
  committed
  529e09c
- Log source tables as native MLflow inputs; cut unused tracking surface
  
  Vedaant-J
  and
  claude
  committed
  9892fc2
- Log manifest source tables as MLflow inputs whatever the connector case
  
  Vedaant-J
  and
  claude
  committed
  2686f5c
- Require MLflow 3.13 in the mlflow extra
  
  Vedaant-J
  and
  claude
  committed
  c9ff49f
- Expose the fit run's id as MlflowTracking.run_id
  
  Vedaant-J
  and
  claude
  committed
  5a78d42
- Read fit inputs through the manifest dataclasses
  
  Vedaant-J
  and
  claude
  committed
  6c1943c
- Record Trainer.fit in Snowflake Experiment Tracking
  
  Vedaant-J
  and
  claude
  committed
  b05926f
- Expose the fit run's name as SnowflakeTracking.run_id
  
  Vedaant-J
  and
  claude
  committed
  f5eb73f
- Carry scoring defaults in a derived model bundle
  
  Vedaant-J
  and
  claude
  committed
  b5b405e
- Normalise scoring defaults and allow a multilabel threshold default
  
  Vedaant-J
  and
  claude
  committed
  75378e4
- Trim scoring defaults to essentials
  
  Vedaant-J
  and
  claude
  committed
  f670d50
- Reuse bundle_task_type and write_manifest for scoring defaults
  
  Vedaant-J
  and
  claude
  committed
  fe4ec2a
- Record a derived bundle's own id in its provenance.json
  
  Vedaant-J
  and
  claude
  committed
  b8e8d9e
- Record bundle derivation on the provenance dataclass
  
  Vedaant-J
  and
  claude
  committed
  248091b
- Accept trailing-slash directories when fetching Databricks models
  
  Vedaant-J
  and
  claude
  committed
  c606def
- Add model aliases and a bundle id pointer to model registries
  
  Vedaant-J
  and
  claude
  committed
  0a07a2d
- Roll back failed Snowflake alias moves; validate aliases before fit
  
  Vedaant-J
  and
  claude
  committed
  18e3b42
- Add prediction_output_columns and connector_for_root helpers
  
  Vedaant-J
  and
  claude
  committed
  f08aeee
- Log, register and serve fitted models through MLflow pyfunc
  
  Vedaant-J
  and
  claude
  committed
  0b693a1
- Score each pyfunc predict in its own scratch dir; state MLflow 3.13
  
  Vedaant-J
  and
  claude
  committed
  136d779
- Bound the model-logging wait; give only the upload a longer timeout
  
  Vedaant-J
  and
  claude
  committed
  bb34a54
- Read provenance and scheduler context as dataclasses for MLflow
  
  Vedaant-J
  and
  claude
  committed
  5255ad9
- Add kumo.job() for scheduled job scripts
  
  Vedaant-J
  and
  claude
  committed
  4e852aa
- Print the job summary in finally; exit 0 or non-zero
  
  Vedaant-J
  and
  claude
  committed
  0a788b2
- Keep at most 20 distinct, truncated warnings in the job summary
  
  Vedaant-J
  and
  claude
  committed
  a792a18
- Leave deprecation warnings out of the job summary
  
  Vedaant-J
  and
  claude
  committed
  235a677
- Do not recurse when another patcher wraps kumo.job()'s showwarning
  
  Vedaant-J
  and
  claude
  committed
  b5d56e0
- Write the job summary's scheduler from SchedulerContext
  
  Vedaant-J
  and
  claude
  committed
  2c18051

Commits

Commits on Oct 22, 2026