Triggering: schedule, new data, performance decay or a manual run - SkilBrill

MLOps Training Triggering: schedule, new data, performance decay or a manual run

MLOps Training

0% Completed

M1 — The ML Lifecycle and What Breaks

From notebook to service: every handoff where a model usually dies

Why software testing practice fails on a model

Reproducibility: seeds, versions, environment and data snapshot

The artifact set: data, code, model, metrics and configuration

Roles: who owns training, delivery and monitoring

Choosing the smallest MLOps footprint the problem deserves

M2 — Pipelines and Orchestration

Turning a notebook into typed, parameterised pipeline steps

Orchestration: scheduling, dependencies and retries

Idempotent steps and safe reruns of a partial pipeline

Data validation at ingestion before training begins

Feature computation and keeping training and serving consistent

Triggering: schedule, new data, performance decay or a manual run

M3 — Experiment Tracking and Feature Stores

Logging parameters, metrics and artifacts for every run

Comparing runs and reproducing the one you actually shipped

Feature stores: shared definitions, point-in-time correctness

Training/serving skew and how a feature store prevents it

Feature freshness, backfills and the cost of a wrong join

Lineage: which features produced which model

M4 — Model Registry and Delivery

The registry as the source of truth for what is in production

Model stages, signatures and the metadata a stage requires

CI for ML: tests for data, tests for code, tests for the model

Validation gates that can block a release on a metric

Packaging: container, bundle or managed endpoint

Rollback when the new model is worse and you have proof

M5 — Serving and Performance

Batch inference versus online endpoints and choosing deliberately

Latency, throughput and the cost curve of a bigger model

A/B testing and canary releases for a model

Shadow deployment: scoring live traffic with no user impact

Autoscaling, cold starts and request batching

Feature computation at request time without adding latency

M6 — Monitoring, Drift and Retraining

Data drift: the input distribution moved

Concept drift: the relationship the model learned changed

Performance decay when labels arrive late or not at all

Building monitoring that alerts on cause rather than noise

The retraining decision: evidence, cadence and approval

Post-retrain validation and comparing against the incumbent

M7 — Applied Skills

Reproducing a colleague's run from its recorded lineage

Diagnosing a pipeline failure: data, code, environment or model

Reducing training cost without changing the algorithm

Writing a model card someone outside the team can read

Reviewing an ML platform design for the failure it will hit first

Estimating the operational cost of a model over a year

M8 — Capstone

Package a training run so it reproduces from its own record

Build an orchestrated pipeline with data validation and retries

Track experiments and register a model with full lineage

Add CI with a metric gate that can block a bad release

Serve behind a versioned endpoint with a working rollback

Wire drift monitoring and a documented retraining trigger

Announcements Course Info