How does CI/CD for machine learning differ from CI/CD for traditional software?

Traditional CI/CD tests and deploys code. ML CI/CD must also handle data and models, which adds pipeline stages. Beyond unit tests, CI validates data schemas, checks data quality, and tests feature-engineering code. It often includes CT (continuous training): an automated pipeline that retrains, evaluates against a baseline, and only promotes a model that beats the current one on holdout metrics. Deployment ships a model artifact plus its serving code, not just code. Extra concerns include validating that the model meets latency and accuracy thresholds, guarding against training-serving skew, and monitoring after deploy for drift. Because behavior depends on data, ML systems need triggers to retrain and redeploy when data changes, not only when code changes. Tools include Kubeflow Pipelines, TFX, GitHub Actions, and MLflow.