databricks-ml-serving-ops

# Databricks ML & Serving Ops Skill ## Purpose Manage Databricks ML experiments, feature engineering, model registry, and serving endpoints via the Databricks CLI (v0.205+). ## Owner databricks-cli-operator ## Preconditions - Databricks CLI v0.205+ installed - Authentication configured with MLflow and serving permissions - Unity Catalog enabled for model registry (recommended) ## Covered Operations ### Experiments (MLflow) - `databricks experiments list` — list experiments - `databricks experiments get --experiment-id <id>` — get experiment details - `databricks experiments create --name <name>` — create experiment - `databricks experiments delete --experiment-id <id>` — delete experiment (soft delete) ### Feature Engineering - `databricks feature-engineering list-tables` — list feature tables (Unity Catalog) - Feature table operations are primarily API/SDK-driven; CLI provides listing and metadata ### Model Registry (Unity Catalog) - `databricks registered-models list` — list registered models - `databricks registered-models get --full-name <catalog.schema.model>` — get model details - `databricks registered-models create --name <name> --catalog-name <cat> --schema-name <schema>` — register model - `databricks registered-models delete --full-name <catalog.schema.model>` — delete model (destructive) - `databricks model-versions list --full-name <catalog.schema.model>` — list model versions - `databricks model-versions get --full-name <catalog.schema.model> --version <v>` — get version details ### Serving Endpoints - `databricks serving-endpoints list` — list serving endpoints - `databricks serving-endpoints get --name <name>` — get endpoint details - `databricks serving-endpoints create --json <spec>` — create serving endpoint - `databricks serving-endpoints update-config --name <name> --json <spec>` — update endpoint config - `databricks serving-endpoints delete --name <name>` — delete endpoint (destructive) - `databricks serving-endpoints query --name <name> --json <input>` — query endpoint for inference ## Disallowed Operations - Direct model training (use Jobs/notebooks, not CLI) - Experiment run creation (use MLflow SDK in notebooks/jobs) - Workspace artifact management (use databricks-workspace-ops) ## Output Contract - All commands use `--output json` - Experiments return `experiment_id`, `name`, `lifecycle_stage` - Models return `full_name`, `creation_timestamp`, `comment` - Serving endpoints return `name`, `state.ready`, `config.served_models` ## Verification - After model register: `registered-models get` returns model metadata - After endpoint create: poll `serving-endpoints get` until `state.ready` is `READY` - After endpoint query: response contains `predictions` array