AWS Lambda Durable Functions are a powerful new tool for orchestrating multi-step workloads — write sequential code, get built-in retries, checkpointing, and executions that run for up to a year, all without Step Functions or external orchestration.

But once you adopt them, a new question arises: how do I actually observe this? What's in the box?

AWS provides a solid starting point with execution-level metrics and lifecycle events. But distributed tracing is where the gaps show. X-Ray gives you per-invocation traces — not per-execution. And when you try to close that gap with OpenTelemetry, the checkpoint-and-replay model introduces real challenges: phantom spans from replay invocations, trace context that breaks at every freeze/resume boundary, and no semantic conventions to build on.

This talk covers how durable functions work, what observability AWS ships out of the box, and how to extend it — including what I learned trying to add proper OTel tracing to a runtime that replays your code.