CLI Reference¶

All commands are available under kedro azureml. Run kedro azureml --help to list them.

kedro azureml init¶

kedro azureml init

Initializes the plugin in the current Kedro project. Creates:

• conf/base/azureml.yml: configuration file with placeholder values
• .amlignore: file exclusion list for code upload

No flags.

kedro azureml compile¶

kedro azureml compile (-j JOB_NAME | --all) [options]

Compiles job(s) into Azure ML pipeline YAML definitions without submitting them. Select jobs with one or more -j names, or compile every resolved job with --all. Add --check to validate that jobs compile without writing any files. See Compile and inspect for a walkthrough.

kedro azureml run¶

kedro azureml run -j JOB_NAME [options]

Submits named job(s) to Azure ML managed compute immediately, ignoring any configured schedule.

In the code flow (execution.code_directory set), one invocation stages the code snapshot once and registers it as one code asset per workspace before any job is generated; every step references that asset. kedro azureml schedule does the same. See the execution configuration for details.

Batch submission is fail-fast¶

When you pass several -j names (or run all configured jobs), they are submitted in the order given, one after another. Jobs in a batch usually feed each other (for example snapshot then training then inference), so if one job fails to submit, run stops immediately and skips every remaining job rather than launching work that would consume missing outputs.

The skipped jobs are listed, and the final summary counts succeeded, failed, and skipped jobs:

Aborting batch: 'training' failed; skipping 1 remaining job(s): inference

Run summary: 1 succeeded, 1 failed, 1 skipped (out of 3 selected)

The command exits non-zero if any job failed or was skipped.

Concurrent submission for independent jobs¶

When the jobs in a batch do not depend on each other (for example one backtest per model against the same data), pass --concurrent. The jobs are then submitted from a worker pool of up to CONCURRENT_SUBMIT_WORKERS threads (a constant in kedro_azureml_pipeline.constants, 4 by default, never more than the number of jobs), every job is attempted whatever the others do, and the summary reports succeeded and failed jobs with none skipped. The exit status follows the same rule as the serial mode: non-zero if any job failed.

--concurrent disables the fail-fast ordering, so do not use it for chains. It cannot be combined with --wait-for-completion: streaming several runs' logs into one terminal is not useful, and the command refuses the pair with a usage error. One credential, and one client per workspace, serve the whole batch in both modes.

Per-job display names¶

A job's display_name in azureml.yml is one template per job, resolved once per invocation. When a caller needs distinct names for jobs submitted together (for example an attempt number computed per job before submit), pass --display-name JOB=NAME once per job. The override replaces both the Azure ML pipeline job's display name and the MLflow root run name for that job; jobs without an override keep their configured name. The value is split on the first =, so NAME may contain =.

Every JOB must be among the -j jobs of the same invocation; an unknown job, an entry without =, or an entry whose job or name half is empty is a usage error, raised before any job is prepared or any code is staged. --dry-run prints each job's display name next to its job name, so a computed override can be checked without submitting. The schedule and compile commands do not take this option.

kedro azureml schedule¶

kedro azureml schedule -j JOB_NAME [options]

Creates or updates persistent Azure ML schedules for named job(s). Every selected job must have a schedule configured in azureml.yml. In the code flow, the snapshot is staged once and registered once per workspace before the schedules are created, exactly as for run (see execution).

kedro azureml resolve-patterns¶

kedro azureml resolve-patterns

Lists the concrete jobs derived from the job factories and the pipeline namespaces (the job analogue of kedro catalog resolve-patterns). For each job it prints the name, pipeline, node namespaces, and schedule; literal (non-factory) jobs are included. No Azure ML connection is made. Add -e ENV to inspect a specific environment.

The names printed here are exactly what you pass to -j on run, schedule, and compile.

kedro azureml list-patterns¶

kedro azureml list-patterns

Lists the job-factory patterns, the jobs keys containing {placeholder} markers (the analogue of kedro catalog list-patterns). Literal job keys are not listed; use resolve-patterns to see the concrete jobs the factories produce.

kedro azureml execute¶

Used internally by Azure ML pipeline steps to run individual Kedro nodes on compute. Not intended for direct use. You may see this command in Azure ML step logs when inspecting pipeline run output in Azure ML Studio.

Common flags¶

The following flags are accepted by compile, run, and schedule: