Repository navigation

Record prediction runs and monitor score drift - #64

Open

Vedaant-J wants to merge 394 commits into

vedaant/mlops-mlflow-trackingfrom

vedaant/mlops-score-sketch

Open

Record prediction runs and monitor score drift#64
Vedaant-J wants to merge 394 commits into
vedaant/mlops-mlflow-trackingfrom
vedaant/mlops-score-sketch

This pull request is big! We're only showing the most recent 250 commits

Commits on Oct 30, 2026
- Use the renamed drift metrics in the MLflow reference of mlops.md
  
  Vedaant-J
  and
  claude
  committed
  7fb2664
- Name a PSI's shift by the rule of thumb
  
  Vedaant-J
  and
  claude
  committed
  a9452f3
- Record where a run's artifacts are as a table of full locations
  
  Vedaant-J
  and
  claude
  committed
  39ccfd7
- Count a numerical target's entities per split, beside its mean
  
  Vedaant-J
  and
  claude
  committed
  441becc
- Name the Databricks workspace URL a tracking target knows
  
  Vedaant-J
  and
  claude
  committed
  456c064
- Sample system metrics every second for the whole run
  
  Vedaant-J
  and
  claude
  committed
  65c5424
- Log system metrics once a run has two readings
  
  Vedaant-J
  and
  claude
  committed
  83974c2
- Count each numerical column exactly on fixed value bins in col_stats.json
  
  Vedaant-J
  and
  claude
  committed
  0f6fedc
- Define psi_shift once in score_drift
  
  Vedaant-J
  and
  claude
  committed
  f5423bc
- Log input drift as one table per source table, with a psi for numbers
  
  Vedaant-J
  and
  claude
  committed
  55fddd8
- Record when a snapshot or training table was built and its size on disk
  
  Vedaant-J
  and
  claude
  committed
  14d218b
- Show the data preparation time and size on fit and prediction cards
  
  Vedaant-J
  and
  claude
  committed
  62d7de9
- Log data preparation time and size as Snowflake run params
  
  Vedaant-J
  and
  claude
  committed
  a361095
- Log a series marked for a chart as MLflow metrics over its steps
  
  Vedaant-J
  and
  claude
  committed
  04e6531
- Chart prediction counts per score hundredth and per item rank
  
  Vedaant-J
  and
  claude
  committed
  97fd327
- Leave charted series out of Snowflake runs, whose tables carry their counts
  
  Vedaant-J
  and
  claude
  committed
  841a1db
- Name runs kind, model, version, date
  
  Vedaant-J
  and
  claude
  committed
  ec862d3
- Log a fit's holdout metrics as test/<metric>
  
  Vedaant-J
  and
  claude
  committed
  cd352ce
- Log the model's holdout metrics against the training table
  
  Vedaant-J
  and
  claude
  committed
  df8d45c
- Give a fit's source tables the dataset context training
  
  Vedaant-J
  and
  claude
  committed
  270e626
- Sort an experiment's predictions by their as-of date in its note
  
  Vedaant-J
  and
  claude
  committed
  07d6dd6
- Document the run naming, test metrics and dataset roles
  
  Vedaant-J
  and
  claude
  committed
  811c9d8
- Document drift definitions, input drift tables, chart series, system metrics and warehouse monitors
  
  Vedaant-J
  and
  claude
  committed
  79cd99c
- Count value bins in a small open-addressing table instead of a BTreeMap
  
  Vedaant-J
  and
  claude
  committed
  bba2c50
- Split input size changes into a table-rows and a link-edges table
  
  Vedaant-J
  and
  claude
  committed
  33b9fdf
- Show the SageMaker MLflow UI hint once per tracking target
  
  Vedaant-J
  and
  claude
  committed
  fab62cf
- Tag a fit with the epoch its kept model is from
  
  Vedaant-J
  and
  claude
  committed
  cace1d8
- Show the binary holdout at its threshold as a 2x2 confusion matrix
  
  Vedaant-J
  and
  claude
  committed
  e8c5fab
- Write the run card as Markdown: a heading, a sentence and tables
  
  Vedaant-J
  and
  claude
  committed
  ca9567a
- Document the run card as written: heading, sentence and tables
  
  Vedaant-J
  and
  claude
  committed
  309dda1
- Compare predictions within float tolerance in the tracking test
  
  Vedaant-J
  and
  claude
  committed
  0c6b40c
- Count every forecast frame when judging label maturity
  
  Vedaant-J
  and
  claude
  committed
  5783c99
- Report the Snowflake run name on the tracking ref
  
  Vedaant-J
  and
  claude
  committed
  0910a9a
- Document the Databricks monitor recipe as verified: appended rows, baseline, refresh
  
  Vedaant-J
  and
  claude
  committed
  6b33ab1
- Name Snowflake registry versions V1, V2, ... in order
  
  Vedaant-J
  and
  claude
  committed
  93215f0
- Fix the Snowflake monitor recipe's table names and say which task types the recipes cover
  
  Vedaant-J
  and
  claude
  committed
  27eb867
- Correct the SageMaker setup, Snowflake run naming, skipped-share metric and registry reuse claims
  
  Vedaant-J
  and
  claude
  committed
  497ad92
- Fix stale prediction-run text, the card order, categorical psi smoothing and the history sort in the docs
  
  Vedaant-J
  and
  claude
  committed
  3633d79
- Write the card's first line as a plain paragraph on one line
  
  Vedaant-J
  and
  claude
  committed
  2844b53
- Say on the card's last line that kumo rewrites what is above it
  
  Vedaant-J
  and
  claude
  committed
  528edcf
- Call a prediction's location its folder, not its run
  
  Vedaant-J
  and
  claude
  committed
  b8ea738
- Leave a scoring option's source out of a prediction's params when the option is unset
  
  Vedaant-J
  and
  claude
  committed
  f0ff7a2
- Show a link prediction fit's holdout metrics at its query's k only
  
  Vedaant-J
  and
  claude
  committed
  00f73ba
- Give no shift verdict for a link prediction's psi vs holdout in the card's first line
  
  Vedaant-J
  and
  claude
  committed
  f1a46c4
- Say the score PSI groups its bins at the holdout's deciles, and drop the rule of thumb for link prediction
  
  Vedaant-J
  and
  claude
  committed
  3d3be9d
- Put the drift blocks and drift notes right after the scores, the block's notes under a heading and its moved tables named
  
  Vedaant-J
  and
  claude
  committed
  da426e0
- Add a card note saying what a prediction chart's step is
  
  Vedaant-J
  and
  claude
  committed
  234103d
- Put the score drift note in its own section so the card shows it with the drift
  
  Vedaant-J
  and
  claude
  committed
  f26783e
- Place the card's drift blocks after the scores and pass its chart note and moved tables
  
  Vedaant-J
  and
  claude
  committed
  4d75730
- Log the plan's acc as accuracy, as binary fits do
  
  Vedaant-J
  and
  claude
  committed
  f0667c1
- Give input-drift columns short headers and mark a capped psi approximate
  
  Vedaant-J
  and
  claude
  committed
  368b2e4
- Drop the unshown Jensen-Shannon distance from column drift
  
  Vedaant-J
  and
  claude
  committed
  790a4b8
- Shorten the table-rows and link-edges headers
  
  Vedaant-J
  and
  claude
  committed
  81f641a
- Give link prediction's score drift vs holdout no shift verdict
  
  Vedaant-J
  and
  claude
  committed
  f7155a1
- Shorten the predicted-class and item-frequency headers
  
  Vedaant-J
  and
  claude
  committed
  4931554
- Carry exact data-preparation seconds and bytes; the card picks the size unit
  
  Vedaant-J
  and
  claude
  committed
  ab82e27
- Leave unshifted columns out of the most-shifted note
  
  Vedaant-J
  and
  claude
  committed
  645cc1c
- Sort two imports and drop unused test names, as CI's flake8 and isort ask
  
  Vedaant-J
  and
  claude
  committed
  b9a96b9
- Give every system-metrics reading its own increasing step
  
  Vedaant-J
  and
  claude
  committed
  55ccbc2
- Drop the snapshot-first dataset sort, which MLflow's store does not keep
  
  Vedaant-J
  and
  claude
  committed
  f09d2ef
- Refresh the experiment description kumo wrote when its text changes
  
  Vedaant-J
  and
  claude
  committed
  4d47458
- Point the experiment description at the sortable param anchors
  
  Vedaant-J
  and
  claude
  committed
  f265236
- Say that only MLflow and Unity Catalog reuse a registered bundle's version
  
  Vedaant-J
  and
  claude
  committed
  831f29b
- Put logloss right after the tuned metric in a version's headline
  
  Vedaant-J
  and
  claude
  committed
  bebab2f
- Show the model's exact holdout metrics on the fit card and its version
  
  Vedaant-J
  and
  claude
  committed
  e8c9637
- Tag fits with their entity and rebuild the fit sentence from tags
  
  Vedaant-J
  and
  claude
  committed
  fab8986
- Show a Snowflake fit's exact holdout metrics on its card, as MLflow does
  
  Vedaant-J
  and
  claude
  committed
  8697116
- Tag the Snowflake fit-card test's run with its entity, as fits now are
  
  Vedaant-J
  and
  claude
  committed
  3d49c2b
- Trim the job templates and their READMEs; name gs:// source files as Parquet
  
  Vedaant-J
  and
  claude
  committed
  424ee9a
- Read a skipped row's sample id from the chunk's first row group onward
  
  Vedaant-J
  and
  claude
  committed
  51d342f
- Stop registering after three Snowflake version names fail; drop unused _backfill
  
  Vedaant-J
  and
  claude
  committed
  04144c8
- Hold a retried trial's first epoch per attempt, not per trial
  
  Vedaant-J
  and
  claude
  committed
  6a632ae
Commits on Oct 31, 2026
- Read a card's metrics under one name for @k metrics, as Snowflake logs them raw
  
  Vedaant-J
  and
  claude
  committed
  e52818b
- Hide MLflow's run links on SageMaker also when the ARN is the global tracking URI
  
  Vedaant-J
  and
  claude
  committed
  1a6998a
- Give multilabel score drift vs holdout its shift; only link prediction has none
  
  Vedaant-J
  and
  claude
  committed
  ac56263
- Read a prediction run's scoring summaries inside its tracking call
  
  Vedaant-J
  and
  claude
  committed
  ece4515
- Give an export of predictions without an id no export marker
  
  Vedaant-J
  and
  claude
  committed
  7add2d6
- Skip the id link model's top-score scatter outside the test stage
  
  Vedaant-J
  and
  claude
  committed
  4fb0d68
- Move entity, number and the @ metric-name rule out of the card renderer
  
  Vedaant-J
  and
  claude
  committed
  c20f25e
- Emit build costs as prep.* tags for every tracker, not a Snowflake special case
  
  Vedaant-J
  and
  claude
  committed
  05d400a
- Route summaries once for every tracker; Snowflake keeps chart series as tables
  
  Vedaant-J
  and
  claude
  committed
  52963cb
- Share the one-run prediction and monitoring hooks across trackers
  
  Vedaant-J
  and
  claude
  committed
  5cf33e2
- Pick a monitored model's holdout metrics with logged_names, not a copy of its refit rule
  
  Vedaant-J
  and
  claude
  committed
  4bf3002
- Drop unused fit results: test_counts and write_provenance's task_type and holdout_path
  
  Vedaant-J
  and
  claude
  committed
  cc02dad
- Cut kumo-core's score-count and drift docstrings to one line
  
  Vedaant-J
  and
  claude
  committed
  78c3cd1
- Cut _monitoring's docstrings to one line; its fields are in docs/mlops.md
  
  Vedaant-J
  and
  claude
  committed
  e50e646
- Cut the card, summary-table, score-stats and prep-cost docstrings and comments
  
  Vedaant-J
  and
  claude
  committed
  741686d
- Cut _data_summaries' docstrings to one line
  
  Vedaant-J
  and
  claude
  committed
  2c1cc8d
- Cut _tracking's docstrings and comments to one line, outside those #60 trims
  
  Vedaant-J
  and
  claude
  committed
  a884d67
- Cut the Snowflake tracker's docstrings and comments
  
  Vedaant-J
  and
  claude
  committed
  747183f
- Cut the MLflow tracker's docstrings and comments added for prediction runs
  
  Vedaant-J
  and
  claude
  committed
  8633890
- Cut the trainer, manifest and prediction-table docstrings and comments added for runs
  
  Vedaant-J
  and
  claude
  committed
  f71a647
- Cut the docstrings and comments added to kumo-core training, scoring and the Snowflake registry
  
  Vedaant-J
  and
  claude
  committed
  1a200d0
- Shorten the warehouse-monitor recipes' prose, keeping each step and caveat once
  
  Vedaant-J
  and
  claude
  committed
  3e3b0c3
- Keep a card table cell with a newline on its row
  
  Vedaant-J
  and
  claude
  committed
  714638c
- Decide in one place whether psi vs the holdout reads as a shift
  
  Vedaant-J
  and
  claude
  committed
  43f0f1a
- Document the prep.* names as they are logged, and on both trackers
  
  Vedaant-J
  and
  claude
  committed
  6036a4c
- Stop passing write_provenance the arguments it dropped in the stub-fit test
  
  Vedaant-J
  and
  claude
  committed
  d678bdd
- Merge #60's review fixes into #64
  
  Vedaant-J
  and
  claude
  committed
  03db190
- Check once whether a tracker logs models
  
  Vedaant-J
  and
  claude
  committed
  85eec74
- Merge #60 into #64
  
  Vedaant-J
  and
  claude
  committed
  f833f26
- Give predict test stubs the manifest a real TrainingResult has
  
  Vedaant-J
  and
  claude
  committed
  950b69e
- Give fit test stubs the published_counts the training result carries
  
  Vedaant-J
  and
  claude
  committed
  1362c1a
- Treat a target with no time range as static when naming its window
  
  Vedaant-J
  and
  claude
  committed
  ac61c80
- Merge #60's ADLS version kind into #64
  
  Vedaant-J
  and
  claude
  committed
  1d9846f
- Merge #60's public-image fix into #64
  
  Vedaant-J
  and
  claude
  committed
  0e0d9e7
- Compose run names in one module for every tracker
  
  Vedaant-J
  and
  claude
  committed
  0a0ac07
- Build the run card in _card for both trackers
  
  Vedaant-J
  and
  claude
  committed
  4e6e450
- Type evaluated rows for Snowflake Model Monitor in SnowflakeConnector
  
  Vedaant-J
  and
  claude
  committed
  6a46493
- Build each score drift reference with one function and table shape
  
  Vedaant-J
  and
  claude
  committed
  c327bed
- Split evaluation into named steps
  
  Vedaant-J
  and
  claude
  committed
  495c1c7
- Split Trainer._fit into named steps
  
  Vedaant-J
  and
  claude
  committed
  203156f
- Split Trainer._predict into named steps
  
  Vedaant-J
  and
  claude
  committed
  4951b20
- Drop the holdout record's never-set columns and label_positive_rate
  
  Vedaant-J
  and
  claude
  committed
  be94976
- Drop the prediction stats' never-set columns and predicted_positive_rate
  
  Vedaant-J
  and
  claude
  committed
  a7e5e2f
- Build prediction stats records directly
  
  Vedaant-J
  and
  claude
  committed
  4a49576
- Log the model's metrics without a dataset so versions share columns
  
  Vedaant-J
  and
  claude
  committed
  8f2ec20
- Merge r10/60 into r10/64a
  
  Vedaant-J
  and
  claude
  committed
  7f4c849
- Drop the Curve summary, which nothing produces
  
  Vedaant-J
  and
  claude
  committed
  7445f55
- Read the prediction, monitoring and run-name tracker members through TrackingBase
  
  Vedaant-J
  and
  claude
  committed
  2437306
- Drop TrackingBase's per-summary hooks, which the routing trackers replace
  
  Vedaant-J
  and
  claude
  committed
  8af2df2
Commits on Nov 1, 2026
- Drop the col-stats percentiles and category-limit options, which no caller sets
  
  Vedaant-J
  and
  claude
  committed
  ca89b0c
- Share the export read and write steps between predictions and evaluations
  
  Vedaant-J
  and
  claude
  committed
  3058757
- Derive the score axis and top-score flag from each task's score name
  
  Vedaant-J
  and
  claude
  committed
  6ed8bc6
- Drop score_counts.merged, which only tests used
  
  Vedaant-J
  and
  claude
  committed
  6476e13
- Describe a registered version through one MlflowTracking helper
  
  Vedaant-J
  and
  claude
  committed
  d37b211
- Log log_model's holdout metrics through MlflowTracking._log_metrics
  
  Vedaant-J
  and
  claude
  committed
  414fae5
- Search the registry for a bundle's version once per registration
  
  Vedaant-J
  and
  claude
  committed
  1cf5783
- Group series and histograms inside route, without the SummaryTables class
  
  Vedaant-J
  and
  claude
  committed
  a1e46da
- Drop the unused _tracking._text helper
  
  Vedaant-J
  and
  claude
  committed
  0400656
- Import metric_key in the MLflow integration instead of redefining it
  
  Vedaant-J
  and
  claude
  committed
  3e18715
- Drop private docstrings that repeat their function's name
  
  Vedaant-J
  and
  claude
  committed
  e7dd204
- Shorten the public docstrings this change adds
  
  Vedaant-J
  and
  claude
  committed
  7081dd4
- Shorten the comments and test docstrings this change adds
  
  Vedaant-J
  and
  claude
  committed
  5ed6ee0
- Make the graph-mismatch, frozen-card and immature-label messages one sentence
  
  Vedaant-J
  and
  claude
  committed
  aa2216a
- Merge #60's round-10 cleanups into #64
  
  Vedaant-J
  and
  claude
  committed
  cbaab86
- Drop the test assertion on the removed holdout columns field
  
  Vedaant-J
  and
  claude
  committed
  8a89605
- Name the registry metrics helper apart from the run one and keep tracker helpers private
  
  Vedaant-J
  and
  claude
  committed
  217c655
- Drop the evaluation's unread _exported field
  
  Vedaant-J
  and
  claude
  committed
  b51b187
- Import Note once at the top of _tracking
  
  Vedaant-J
  and
  claude
  committed
  cc968db
- Read the scoring snapshot directly in the prediction record
  
  Vedaant-J
  and
  claude
  committed
  1eac5d1
- Catch OSError alone when an export has no sidecar
  
  Vedaant-J
  and
  claude
  committed
  13c2923
- Stop scanning skipped-row samples once enough ids are read
  
  Vedaant-J
  and
  claude
  committed
  285c1d2
- Read the published evaluation's counts without getattr
  
  Vedaant-J
  and
  claude
  committed
  88b8e86
- Write the merged prediction counts with save_counts
  
  Vedaant-J
  and
  claude
  committed
  8c42c48
- Share TOP_CLASSES and kept_horizons between the count modules
  
  Vedaant-J
  and
  claude
  committed
  6cf7dcc
- Merge main into #64 through #60
  
  Vedaant-J
  and
  claude
  committed
  9fb8961
- Call column stats with default options through one test helper
  
  Vedaant-J
  and
  claude
  committed
  d86eacd
- Name grouped evaluation tasks once and pick the class type in one branch
  
  Vedaant-J
  and
  claude
  committed
  e323929
- Return the native prediction build directly
  
  Vedaant-J
  and
  claude
  committed
  da0d13c
- Document compute_col_stats_for_parquet's fixed_edges
  
  Vedaant-J
  and
  claude
  committed
  d8071e3
- Merge the updated #64 head into the round-11 cleanups
  
  Vedaant-J
  and
  claude
  committed
  ce04cc0
- Merge #60's re-review commits into the round-11 cleanups
  Vedaant-J
  committed
  efcf568
- Correct the MLOps guide for the holdout, input-drift and Models tab changes
  
  Vedaant-J
  and
  claude
  committed
  044e86c
- Merge #60's re-review cleanups into #64
  
  Vedaant-J
  and
  claude
  committed
  755f06b
- Merge #60's churn moves into #64
  
  Vedaant-J
  and
  claude
  committed
  c7e6f90
- Drop an unneeded mut in the column-stats options test
  
  Vedaant-J
  and
  claude
  committed
  1b2d873
Commits on Nov 2, 2026
- Merge #114 through #60 into #64
  
  Vedaant-J
  and
  claude
  committed
  9d3438d
- Refuse an occupied volume directory under APPEND, since OVERWRITE empties it
  
  Vedaant-J
  and
  claude
  committed
  e2c420e
- Name the per-hundredth score counts pred/score_histogram, per class for multi-class
  
  Vedaant-J
  and
  claude
  committed
  5d03569
- Write each bar chart summary as an SVG image beside the run's tables, on MLflow and Snowflake
  
  Vedaant-J
  and
  claude
  committed
  6a7ce47
- Chart a prediction run's scores against the holdout's as charts/pred/score_histogram.svg
  
  Vedaant-J
  and
  claude
  committed
  85dd03a
- Shorten the experiment note to its two run filters, with no angle-bracket placeholders
  
  Vedaant-J
  and
  claude
  committed
  da508b5
- Draw one equal-width bar per bin and label the bin edges, so uneven bins chart honestly
  
  Vedaant-J
  and
  claude
  committed
  b65fe1c
- Skip summary types a tracker does not know instead of raising
  
  Vedaant-J
  and
  claude
  committed
  f833aa0
- Chart a regression run's predicted values against the holdout's as charts/pred/prediction_histogram.svg
  
  Vedaant-J
  and
  claude
  committed
  0478f5f
- Merge #60's unknown-summary skip into #64
  Vedaant-J
  committed
  3093a6c
- Merge main through #60 into #64
  
  Vedaant-J
  and
  claude
  committed
  cb0032c
- Name the per-rank recommendation counts pred/item_popularity
  
  Vedaant-J
  and
  claude
  committed
  defa7fc
- Shrink a chart title's font when the title is too long to fit
  
  Vedaant-J
  and
  claude
  committed
  659414e
- Chart a recommendation run's top 100 items as charts/pred/item_popularity.svg
  
  Vedaant-J
  and
  claude
  committed
  63ee39b
- Merge #112's chunked-writer fixes and #50's MLOps scope note into #64
  Vedaant-J
  committed
  6ef7147
- Separate the durable predict path from the platform DataFrame adapters in the guide
  
  Vedaant-J
  and
  claude
  committed
  6ede208
- Merge #60's pyfunc adapter wording into #64
  
  Vedaant-J
  and
  claude
  committed
  2bc03fb
- Keep explicit anchor timestamps from snapping a split's boundaries
  
  Vedaant-J
  and
  claude
  committed
  b3d3fa7
- Read evaluate labels from every generated row, not only the split stages
  
  Vedaant-J
  and
  claude
  committed
  e1dd004
- Share the holdout's error-by-step table
  
  Vedaant-J
  and
  claude
  committed
  0738518
- Evaluate forecasts per step, multilabel classes and ranking at the fit's k in DiskGraph
  
  Vedaant-J
  and
  claude
  committed
  a2d8d85
- Evaluate every task type the trainer fits
  
  Vedaant-J
  and
  claude
  committed
  e20d808
- Document evaluate's metrics for every task type
  
  Vedaant-J
  and
  claude
  committed
  f51ce77
- Count holdout scores in bounded chunks, ranking top-k without a full sort
  
  Vedaant-J
  and
  claude
  committed
  68fd976
- Count link top scores on one sorted copy
  
  Vedaant-J
  and
  claude
  committed
  dc0523b
- Remove EvaluationResult.frame; the evaluated rows stay in actuals
  
  Vedaant-J
  and
  claude
  committed
  beef553
- Log realized error by step as a tracked table, not an EvaluationResult field
  
  Vedaant-J
  and
  claude
  committed
  12d1694
- Test that evaluate's labels are the fit table's rows for every task type
  
  Vedaant-J
  and
  claude
  committed
  b68c404
- Merge evaluate for every task type into #64
  
  Vedaant-J
  and
  claude
  committed
  9d311cb
- Merge #60's population removal and test fixes into #64
  
  Vedaant-J
  and
  claude
  committed
  92d35b2
- Bin a forecasting target in the split stats' histogram
  
  Vedaant-J
  and
  claude
  committed
  2e70093
- Take holdout target bins only from the split statistics, noting when there are none
  
  Vedaant-J
  and
  claude
  committed
  35718b0
- Merge #50's streamed training-table store into #64
  
  Vedaant-J
  and
  claude
  committed
  b74a2d7
- Compute exact NumPy quantiles of a column in the parquet col-stats pass
  
  Vedaant-J
  and
  claude
  committed
  10d1857
- Spill holdout top scores to Parquet and count them with DiskGraph
  
  Vedaant-J
  and
  claude
  committed
  91b6f69
- Remove holdout_evaluation and its holdout baseline rows
  
  Vedaant-J
  and
  claude
  committed
  9b69a83
- Document the warehouse monitors without a holdout baseline
  
  Vedaant-J
  and
  claude
  committed
  9c4f71e
- Count multilabel and forecasting holdouts in bounded chunks
  
  Vedaant-J
  and
  claude
  committed
  d7cd940
- Tally recommended items in place and summarise them in chunks
  
  Vedaant-J
  and
  claude
  committed
  6756093
- Return an all-skipped prediction source with its row failures
  
  Vedaant-J
  and
  claude
  committed
  a175d69
- Refuse prediction rows that none can be scored
  
  Vedaant-J
  and
  claude
  committed
  a78d8aa
- Say that a predict failing before scoring records no prediction run
  
  Vedaant-J
  and
  claude
  committed
  3981526
- Merge #64's head into the holdout and items counting work
  
  Vedaant-J
  and
  claude
  committed
  57ab0c4
- Merge #60's dataset roles into #64
  
  Vedaant-J
  and
  claude
  committed
  dd9c34d
- Label a prediction's scoring snapshot and its tables by role
  
  Vedaant-J
  and
  claude
  committed
  2a43ba8
- Name top-score spills by rank and remove them once counted
  
  Vedaant-J
  and
  claude
  committed
  af7a45e
- List every skipped prediction row in skipped/ Parquet parts
  
  Vedaant-J
  and
  claude
  committed
  7032567
- Keep, store and track a prediction run's skipped rows
  
  Vedaant-J
  and
  claude
  committed
  83d7022
- Name skipped ids in the run card and document skipped/
  
  Vedaant-J
  and
  claude
  committed
  d534dc2
- Test that every skipped prediction row is stored and tracked
  
  Vedaant-J
  and
  claude
  committed
  929e554
- Count a run's skipped rows from the skipped/ part footers
  
  Vedaant-J
  and
  claude
  committed
  3ebd238
- Say skipped rows were found while building the prediction table
  
  Vedaant-J
  and
  claude
  committed
  046bb27
- Find the training table by its new name in the logged-model test
  
  Vedaant-J
  and
  claude
  committed
  7f2eb0f
- Show how a run's data depend on each other in its card's Lineage block
  
  Vedaant-J
  and
  claude
  committed
  3c7f50f
- Merge the bounded holdout and items counting work into #64
  
  Vedaant-J
  and
  claude
  committed
  3924443
- Merge the all-skipped fix and the full skipped-rows list into #64
  
  Vedaant-J
  and
  claude
  committed
  7830ffd
- Say predict raises only when no requested entity can be scored
  
  Vedaant-J
  and
  claude
  committed
  d9209b9
- Match the skipped-entities warning's new wording in the pyfunc test
  
  Vedaant-J
  and
  claude
  committed
  e0e8cc1
- Delete trainer.py's second copies of _warm_start_plan, _validate_model_plan, _explain and _preflight
  
  Vedaant-J
  and
  claude
  committed
  58ead71
- Drop the docstrings of #64's private helpers
  
  Vedaant-J
  and
  claude
  committed
  65223b8
- Name the path-based connectors once, in _manifest
  
  Vedaant-J
  and
  claude
  committed
  8f575e1
- Share one untracked-run warning, inline record_failure and read the trained query once in _tracking
  
  Vedaant-J
  and
  claude
  committed
  f785539
Commits on Nov 3, 2026
- Build the holdout top-score spill in one Trainer method
  
  Vedaant-J
  and
  claude
  committed
  88f3dfb
- Pass the holdout scoring options to both test calls from one dict
  
  Vedaant-J
  and
  claude
  committed
  2de0030
- Compute a score column's mean, min and max in one helper
  
  Vedaant-J
  and
  claude
  committed
  4302b68
- Read the card's tuned metric in one helper
  
  Vedaant-J
  and
  claude
  committed
  d5e9689
- Inline _smoothed into _category_psi
  
  Vedaant-J
  and
  claude
  committed
  f809175
- Build histogram cumulative shares with accumulate
  
  Vedaant-J
  and
  claude
  committed
  2c6a363
- Say once in mlops.md that the warehouse monitor recipes are untested, and point the card's metric meanings at Metrics
  
  Vedaant-J
  and
  claude
  committed
  4da226d
- Call the counts dict counts in _labels_mature
  
  Vedaant-J
  and
  claude
  committed
  a9557ff
- Drop or shorten field comments the field names already say
  
  Vedaant-J
  and
  claude
  committed
  cb6efca
- Parametrize the card heading tests
  
  Vedaant-J
  and
  claude
  committed
  15337f4
- Merge the one-epoch note tests and drop duplicate imports in the tracking contract
  
  Vedaant-J
  and
  claude
  committed
  270919d
- Import shared names once in the MLflow tracking tests
  
  Vedaant-J
  and
  claude
  committed
  64211f8
- Share graph building, one evaluation and the tracker failure cases in the monitoring tests
  
  Vedaant-J
  and
  claude
  committed
  e7ac208
- Run the prediction manifest tests on the shared fitted model and one plain prediction
  
  Vedaant-J
  and
  claude
  committed
  a5928d0
- Share the brute-force histogram and parametrize the at-k cases in the count tests
  
  Vedaant-J
  and
  claude
  committed
  8a8d31d
- Give the metric-list trainer stub a global rank, as the top-score spill reads it
  
  Vedaant-J
  and
  claude
  committed
  e74882d
- Merge remote-tracking branch 'origin/vedaant/mlops-score-sketch' into q64-src
  Vedaant-J
  committed
  4430736
- Simplify the col-stats options code and share its Arrow helpers
  
  Vedaant-J
  and
  claude
  committed
  5b91312
- Reuse the shared Arrow helpers in evaluation and split out the actuals batch
  
  Vedaant-J
  and
  claude
  committed
  eb648fc
- Compact the col-stats options and evaluation tests
  
  Vedaant-J
  and
  claude
  committed
  54fb267
- Shorten the col-stats options and evaluation docstrings
  
  Vedaant-J
  and
  claude
  committed
  2772737
- Trim comments on value bins, skipped rows and forecast histograms
  
  Vedaant-J
  and
  claude
  committed
  855783b
- Simplify the skipped-rows tests
  
  Vedaant-J
  and
  claude
  committed
  2e7ff5e
- Merge #64 head into the Rust quality pass
  Vedaant-J
  committed
  4cb1bea
- Merge remote-tracking branch 'origin/vedaant/mlops-score-sketch' into q64-src
  Vedaant-J
  committed
  c4788f9
- Name the tracker run in score_drift's and evaluate's tracking refs
  
  Vedaant-J
  and
  claude
  committed
  c176507
- Merge #60's quality pass into #64
  
  Vedaant-J
  and
  claude
  committed
  98ca5e2
- Merge mlops-mlflow-tracking into mlops-score-sketch
  
  Vedaant-J
  and
  claude
  committed
  3ada1be
- Compare timestamp drift across index clocks in seconds
  
  Vedaant-J
  and
  claude
  committed
  c96d1e5
- Merge mlops-mlflow-tracking into mlops-score-sketch
  
  Vedaant-J
  and
  claude
  committed
  8a31c5e
- Merge vedaant/mlops-mlflow-tracking into vedaant/mlops-score-sketch
  
  Vedaant-J
  and
  claude
  committed
  454a0bd

Commits

Commits on Oct 30, 2026

Commits on Oct 31, 2026

Commits on Nov 1, 2026