Skip to main content

Model operations in Studio

In Operations → Models, identify the registered version you want to use. An active ONNX row offers Score a query, Operations, and Monitor. Expand the row to see available actions for its versions. To inspect saved SQL, metrics, and runs, click the model name or open its Workflow tab. See Models in Studio.

Preview and explain predictions​

  1. Choose Score a query on the model version and supply read-only Input SQL.
  2. Choose Preview columns and map each positional model input to the matching query column. If the version declares units, enter the units of the source columns.
  3. Choose Preview predictions. Inspect the results and use Open in SQL editor or Copy SQL to reuse the generated query.
  4. For a version with a saved experiment reference, choose Explain predictions. Each delta is the row's original prediction minus the prediction after replacing one input with its holdout mean. Use it to inspect the model's sensitivity to that substitution.

Use Look up feature group to compose point-in-time features before mapping the inputs. The generated lookup matches data by entity key and event timestamp. See Pin sources and features.

Set a version contract​

Expand a row and select Contract for the version. Set the units, nullability, and optional minimum and maximum for each input. For freshness checks, name the observation timestamp column and set the maximum age. Choose the output meaning and unit. Output meanings include a numeric value, probability, and binary label.

Choose Save version contract. A saved contract belongs to that registered version. Create a new model version to change the contract. Studio's prediction preview and batch runs validate input values and check the prediction against the declared output. For example, map an invoice amount to an input with unit USD, declare that unit in the scoring definition, and supply a recent feature timestamp for a version that requires freshness.

Publish a batch scoring window​

  1. On the intended version, choose Operations → New batch scoring job. Enter a name, a qualified prediction table, and the Scoring SQL.
  2. Choose Preview columns. Map the entity ID and event timestamp, then map the model inputs in order. Use a source query with one row per entity in the scoring window. Supply source units if the model contract requires them.
  3. Choose Save batch job. Select it under Saved operation, enter the UTC window, and choose Run window. The end date is exclusive. For windows larger than 10,000 scoring rows, use narrower time windows.
  4. Open Run history to see saved predictions and the number of new rows. The prediction table includes the batch, run, and model version identifiers along with the entity, event timestamp, and prediction.

Studio stages the complete scoring window and publishes it with one destination write. Repeating the same job keeps previously published rows for the same entity ID and event time. A retry can therefore save no additional rows when those predictions were already published.

Watch, cancel, or resume a run​

Open Operations console on the Models page. Search for the job and filter by status. Expand a run to see its window, attempts, phase, progress, and errors. The console covers jobs across your workspace.

  • Cancel run stops a run before publication begins.
  • Resume run retries an eligible failed, cancelled, or stalled run for the same window. When Studio has a complete staging checkpoint, it reuses it.
  • Open the console from a Studio notification to investigate a failed operation or an outcome threshold alert.

Schedule runs​

Manual runs use your signed-in connection and grants. To schedule a batch job or outcome monitor, have a deployment administrator assign a scoped service identity to that operation. Open Service identity approval to find the operation ID and definition hash. When that identity is ready, set Repeat (hours) and choose Enable schedule.

Match predictions to delayed outcomes​

For an invoice model, payment dates may arrive after predictions have been published. In Operations, choose New outcome monitor and select the batch job that wrote the predictions.

Supply Actual outcomes SQL with the entity ID, the actual outcome, and the time the outcome was observed. Choose Preview columns, map those fields, and set the task, quality threshold, and lookback. Return one current outcome per entity. Use the same entity identifiers as the scoring job and an observation timestamp at or after scoring.

Choose Save outcome monitor, then run or schedule it for the prediction window. Run history shows how many predictions matched outcomes and reports mean absolute error for regression or accuracy for binary labels. The Workflow's Outcomes node links the outcome run to its prediction table and recorded result.

Monitor quality and feature drift​

Use the version's Monitor action for recurring comparisons on a query that contains features and observed outcomes together. Select Add monitor, map the inputs and Actual outcome, and optionally choose a segment column. Set the quality and drift thresholds, then Save monitor or use Run now on a saved monitor. Review sample coverage, feature and prediction drift, and quality by segment. Recurring checks use the monitor creator's active Studio sign-in in the workspace.

Compare a candidate with a shadow rollout​

Open a saved batch job and expand Shadow rollout. Set the baseline and candidate versions. Pick minimum row count, minimum paired coverage, maximum mean prediction difference, and maximum candidate errors. Choose Save shadow plan.

Use Check selected window to run a comparison on demand. A matching completed batch can also trigger the check. Studio evaluates both saved artifacts on the same replay inputs and records their paired coverage, errors, prediction differences, and replay timing.

Enable Restore the baseline automatically if the active candidate breaches these thresholds when you want Studio to restore a previously active baseline. The check verifies the active candidate and the baseline before rollback, and records the result in the job's shadow history.