By name: retrain-pipeline-spec description: Specify when and how to retrain the model. Use when mlops needs a retrain strategy.
Retrain Pipeline Spec
Produce a specification for retraining and rollout.
Role
You define triggers, pipeline steps, and validation for retraining.
Input
- Path to training script, data source description, and experiment config
- Optional: current deployment and monitoring context
Output
Write to the path provided by the caller (e.g. output/{project-slug}/mlops/retrain-pipeline.md) with:
Structure
# Retrain Pipeline: {Project}
## Triggers
- Schedule: e.g. weekly, monthly
- Conditional: e.g. when drift > X, when accuracy drops below Y
- Manual: on-demand
## Pipeline steps
1. Data pull: source, date range, validation
2. Preprocessing: same as feature spec
3. Training: command, config, artifact output
4. Evaluation: metrics, gate (e.g. must beat current model)
5. Rollout: canary, A/B, or full replace
6. Rollback: condition and procedure
## Validation and gates
- Minimum performance to promote
- Data quality checks before training
- Post-deploy checks
Rules
- Pipeline must be executable from the training script and config.
- Caller provides output path.