ML Ops

When this skill is activated, always start your first response with the 🧢 emoji.

ML Ops

A production engineering framework for the full machine learning lifecycle. MLOps bridges the gap between model experimentation and reliable production systems by applying software engineering discipline to ML workloads. This skill covers model deployment strategies, experiment tracking, feature stores, drift monitoring, A/B testing, and versioning - the infrastructure that makes models trustworthy over time. Think of it as DevOps for models: automate everything, measure what matters, and treat reproducibility as a first-class constraint.

When to use this skill

Trigger this skill when the user:

• Deploys a trained model to a production serving endpoint
• Sets up experiment tracking for training runs (parameters, metrics, artifacts)
• Implements canary or shadow deployments for a new model version
• Designs or integrates a feature store for online/offline feature serving
• Sets up monitoring for data drift, prediction drift, or model degradation
• Runs A/B or champion/challenger tests across model versions in production
• Versions models, datasets, or pipelines with DVC or a model registry
• Builds or migrates to an automated training/retraining pipeline

Do NOT trigger this skill for:

• Core model research, architecture design, or hyperparameter search (use an ML research skill instead - MLOps starts after a candidate model exists)
• General software observability (logs, metrics, traces for non-ML services - use the backend-engineering skill)

Key principles

1. Reproducibility is non-negotiable - Every training run must be reproducible from scratch: fixed seeds, pinned dependency versions, tracked data splits, and logged hyperparameters. If you cannot reproduce a model, you cannot debug it, audit it, or roll back to it safely.

2. Automate the training pipeline - Manual training is a one-way door to undocumented models. Build an automated pipeline (data ingestion -> preprocessing -> training -> evaluation -> registration) from day one. Humans should only approve a model for promotion, not run the steps.

3. Monitor data, not just models - Model metrics degrade because the input data changes. Track feature distributions in production against training baselines. Data drift is usually the root cause; model drift is the symptom.

4. Version everything - Models, datasets, feature definitions, pipeline code, and environment configs all deserve version control. An unversioned artifact is a liability. Use DVC for data/models, a model registry for lifecycle state, and git for code.

5. Treat ML code like production code - Tests, code review, CI/CD, and on-call rotation apply to training pipelines and serving code. The "it works in the notebook" standard is not a production standard.

Core concepts

ML lifecycle describes the end-to-end journey of a model:

Experiment -> Train -> Validate -> Deploy -> Monitor -> (retrain if drift)

Each stage has gates: an experiment produces a candidate; training on full data with tracked params produces an artifact; validation gates on held-out metrics; deployment chooses a serving strategy; monitoring decides when retraining is needed.

Model registry is the source of truth for model lifecycle state. A model moves through stages: Staging -> Production -> Archived. The registry stores metadata, metrics, lineage, and the artifact URI. MLflow Model Registry, Vertex AI Model Registry, and SageMaker Model Registry are the main options.

Feature stores decouple feature computation from model training and serving. They have two serving paths: an offline store (columnar, batch-oriented, used for training and batch inference) and an online store (low-latency key-value lookup, used at prediction time). The critical guarantee is point-in-time correctness - training features must only use data available before the label timestamp to prevent target leakage.

Data drift occurs when the statistical distribution of input features in production diverges from the training distribution. Concept drift occurs when the relationship between features and labels changes even if feature distributions are stable (e.g., user behavior shifts after a product change).

Shadow deployment runs the new model in parallel with the live model, receiving the same traffic, but its predictions are not served to users. Used to compare behavior before any real traffic exposure.

Common tasks

Design an ML pipeline

Structure pipelines as discrete, testable stages with explicit inputs/outputs:

Data ingestion -> Validation -> Preprocessing -> Training -> Evaluation -> Registration
     |                |               |              |             |
  raw data      schema check     feature eng      model       go/no-go
  versioned     + stats           artifact       artifact      gate

Orchestration choices:

Gate evaluation: define a go/no-go threshold before training starts. A model that does not beat baseline (or the current production model) should never reach the registry.

Set up experiment tracking

Track every training run with: parameters (hyperparams, data version), metrics (loss curves, eval metrics), artifacts (model weights, plots), and environment (library versions, hardware).

MLflow pattern:

import mlflow

mlflow.set_experiment("fraud-detection-v2")

with mlflow.start_run(run_name="xgboost-baseline"):
    mlflow.log_params({
        "max_depth": 6,
        "learning_rate": 0.1,
        "n_estimators": 200,
        "data_version": "2024-03-01"
    })

    model = train(X_train, y_train)

    mlflow.log_metrics({
        "auc_roc": evaluate_auc(model, X_val, y_val),
        "precision_at_k": precision_at_k(model, X_val, y_val, k=100)
    })

    mlflow.sklearn.log_model(
        model,
        artifact_path="model",
        registered_model_name="fraud-detector"
    )

Key discipline: log the data version (or dataset hash) as a parameter. Without it, you cannot reproduce the run.

Compare runs on the same held-out test set. Never tune on the test set. Use validation for selection, test set for final reporting only.

Deploy a model with canary rollout

Choose a serving infrastructure before choosing a rollout strategy:

Canary rollout stages:

v1: 100% traffic
  -> v2 shadow: 0% served, 100% shadowed (compare outputs)
  -> v2 canary: 5% traffic -> monitor error rate + latency
  -> v2 staged: 25% -> 50% -> 100% with automated rollback triggers

Define rollback triggers before deploying: error rate > X%, prediction latency p99 > Y ms, or business metric (e.g., conversion rate) drops > Z%.

Implement model monitoring

Monitor three layers - input data, predictions, and business outcomes:

Population Stability Index (PSI) thresholds:

PSI < 0.1  -> No significant change, model stable
PSI 0.1-0.2 -> Moderate drift, investigate
PSI > 0.2  -> Significant drift, retrain or escalate

Monitoring setup pattern:

# On each prediction batch, compute and log feature stats
baseline_stats = load_training_stats()  # saved during training
production_stats = compute_stats(current_batch_features)

for feature in monitored_features:
    psi = compute_psi(baseline_stats[feature], production_stats[feature])
    metrics.gauge(f"drift.psi.{feature}", psi)

    if psi > 0.2:
        alert(f"Significant drift on feature: {feature}")

Set up scheduled monitoring jobs (hourly/daily depending on traffic volume) rather than per-prediction to avoid overhead. Load the references/tool-landscape.md for monitoring platform options.

Build a feature store

Separate feature computation from model code to enable reuse and prevent leakage.

Architecture:

Raw data sources
      |
Feature computation (Spark, dbt, Flink)
      |
      +-----------> Offline store (Parquet/BigQuery) -> Training jobs
      |
      +-----------> Online store (Redis, DynamoDB)  -> Real-time serving

Point-in-time correctness - the most critical correctness property:

# WRONG: uses future data at training time (target leakage)
features = feature_store.get_features(entity_id=user_id)

# CORRECT: fetch features as they existed at the event timestamp
features = feature_store.get_historical_features(
    entity_df=events_df,  # includes entity_id + event_timestamp
    feature_refs=["user:age", "user:30d_spend", "user:country"]
)

Feature naming convention: <entity>:<feature_name> (e.g., user:30d_spend, product:avg_rating_7d). Version feature definitions in a registry (Feast, Tecton, Vertex Feature Store). Never hardcode feature transformations in training scripts.

A/B test models in production

A/B testing models requires statistical rigor. A "better offline metric" does not guarantee better business outcomes.

Setup:

1. Define the primary metric (business metric, not model metric) and a guardrail metric before the test
2. Calculate required sample size for desired power (typically 80%) and significance level (typically 5%)
3. Randomly assign users/sessions to treatment/control - sticky assignment (same user always gets the same model) prevents contamination
4. Run for full business cycles (minimum 1-2 weeks for weekly seasonality)

Traffic splitting options:

Option A: Load balancer routing (simple %, stateless)
Option B: User-ID hashing (sticky, consistent assignment)
Option C: Experimentation platform (Statsig, Optimizely, LaunchDarkly)

Stopping criteria: Do not peek at p-values daily. Pre-register the minimum runtime and only stop early for clearly harmful outcomes (guardrail breach). Use sequential testing methods (mSPRT) if early stopping is required by business needs.

A model that improves AUC by 2% but reduces revenue is not a better model. Always tie model tests to business metrics.