Cloud and data
Cloud services, data storage, and MLOps: the basic model and terms.
Cloud
How object storage differs from a filesystem, what a bucket and an object are, and why it matters for backups and static files.
How IaaS, PaaS, and SaaS differ, who's responsible for what in each model, and how to choose between "take it as a service" and "run it yourself".
Data and databases
How to connect to Postgres, look around with psql, and run basic SQL queries. The minimum a DevOps engineer needs.
Why schema changes go through versioned migrations, how up differs from down, and how to fit this into a deploy with no downtime.
Why database connections are expensive, how a pool reuses them, and why pgbouncer sits in front of Postgres.
What Redis is, why you cache in memory, how to store sessions, and why TTL and invalidation matter.
Why you decouple services with a queue, how a producer differs from a consumer, and what sets RabbitMQ apart from Kafka.
MLOps
How a service with a model differs from a regular one, why a model degrades with no code changes, and what tasks that creates for an engineer.
Three maturity levels from the Google paper: a manual process, automatic retraining, and CI/CD for the pipeline. The signs of each level and what breaks at it.
What steps the training pipeline consists of, what's checked before and after training, and why you need a model registry and metadata.
Where training data lives: data warehouse, data lake, lakehouse, the Parquet and Iceberg formats. How to version datasets with DVC and lakeFS, and why you need a feature store.
How MLflow records experiments and stores models, how an orchestrator runs the training pipeline, and how Airflow differs from Kubeflow Pipelines.
How a model answers requests in production: batch and online, a service in a container, KServe, Triton, and vLLM. How to roll out a new model gradually and what to monitor besides CPU.