Observability training is usually terrible. It is either a product demo pretending to be a course or an abstract lecture about the three pillars that never touches a keyboard. The Datadog Learning Center avoids both by giving you a live sandbox with real telemetry flowing into it and asking you to go and find something. That single design decision is why this is worth your time.
The AI observability collection is the reason to look now. Getting Started with LLM Observability covers exploring trace data containing the inputs and outputs of model calls, observing performance through metrics, and identifying errors in applications built on models. This is the operational discipline nobody teaches. There is a large amount of material available on building with models and almost none on running them in production, where the actual questions are why the answers got worse last Tuesday, which prompt version was live when it happened, and how you would even know.
The course takes those questions seriously. The concepts transfer. A span, a trace, an evaluation attached to a generation, the difference between a latency problem and a quality problem, all of that is vendor-neutral thinking dressed in Datadog's interface. If you run LLM applications and you monitor them with Langfuse or Arize or a pile of custom logging, you will still come out of this course better at the job.
That is a high compliment for free vendor training and I do not extend it often. Datadog recently restructured the Learning Center around learning paths, fourteen in the general curriculum and three aimed at certification preparation, and the improvement is noticeable. Previously the courses existed as a flat list and the sequencing was left to you. Now there is a persona-based route through, which matters when the product surface is as large as Datadog's.
Certification is the paid layer and the pricing is not surfaced from the course pages, which is mildly annoying. Datadog also runs pilot versions of updated exams for limited windows, which is good practice for exam quality and slightly frustrating if you plan around one and find the window closed. The Fundamentals certification path is the sensible entry point. The obvious structural limitation is that this teaches Datadog.
Every exercise happens in their interface, every concept arrives wearing their naming, and the implicit answer to every problem is a Datadog feature. That is not dishonest, it is what a vendor academy is, but somebody learning observability from scratch here will need to do some translation work before those skills read as general competence rather than tool familiarity. The gap I would most like closed is cost. Observability spend is one of the great unpleasant surprises in modern infrastructure, and LLM observability adds a new axis to it, since capturing full prompt and completion payloads at volume is not cheap.
The training covers how to instrument thoroughly and says comparatively little about how to instrument affordably. Teams discover that distinction on an invoice. Four point zero. Genuinely hands-on, genuinely free, and currently one of the few structured routes into LLM observability as an operational skill.
Marked down for vendor framing and for not teaching the cost discipline that goes with the instrumentation discipline.