InfrastructureSubsection
MLOps
Everything that has to be true before a model is allowed near production, and everything that keeps being true afterwards. Three topics: the pipeline that makes a training run reproducible, the serving path where latency and cost become somebody’s problem, and the index that makes similarity search fast enough to use. Running an LLM feature specifically is LLMOps, in Development.
What is inside
Every page in this group, with what each one covers.
- ML PipelinesLineage as the deliverable: which data, which code, which parameters produced this model — and can you do it again.LineageReproducible
- Model ServingThe serving path end to end, plus what to observe about a model specifically — including the three-way drift diagnosis.LatencyDrift
- Vector DatabasesApproximate nearest neighbour search: what "approximate" gives up, and how to tell when it gave up too much.ANNRecall
Where to start
- Models are trained by handML pipelines. Lineage is the thing that turns a notebook result into something you can ship and later explain.
- A model is already liveModel serving, and the drift section in particular — it is the page that owns what to watch about a model.
- Building retrievalVector databases here, then RAG architecture in Development — the store and the design that uses it.
Elsewhere in the section
- InfrastructureInfrastructure is everything that turns code into something running that you can observe, change and roll back. This section covers 11 topics in four groups: DevOps and the pipeline that builds one artifact and promotes it unchanged, the three major clouds and how each contains a mistake, MLOps — the same problems again for a model instead of a binary — and the monitoring that makes an incident answerable at three in the morning.
- AI & LLMBuilding features on a component that gives a different answer to the same question twice. Six topics in the order they become relevant: what the model is actually doing, how to ask it properly, how to give it your own data, when to change the model itself, how to run the result in production, and what changes when it can call tools. Testing these features is a discipline of its own and lives in QA.
- MonitoringTwo pages with a clean line between them. Logging owns the record — what gets written down and whether a question asked three weeks later has an answer. Alerting starts where a condition has to reach a person, which makes it a question about interrupting somebody rather than about thresholds. Read them in that order: you cannot alert usefully on data you never structured.
Found this useful?
Share it with someone who is working on the same problem.