Amazon Web Services has added OpenTelemetry (OTel) support to Lambda durable functions. A new plugin for the Durable Execution SDK produces one trace for a whole durable execution, even when that execution pauses and resumes across many Lambda invocations. AWS announced the feature on November 19, 2026, alongside a detailed post on its Compute blog.
Durable functions can run multi-step workflows for up to a year, with checkpoints and automatic retries. That model is resilient, but it has made tracing hard. Each invocation used to produce a trace of its own, so a single workflow showed up as several unrelated pieces in an observability tool.
What Changed
The OpenTelemetry plugin sits next to the Durable Execution SDK and plugs into the durable handler. It is available for the TypeScript, Python and Java SDKs. The documentation says Go, Rust and C# do not offer this integration yet.
Traces can go to Amazon CloudWatch or to any other tool that accepts OpenTelemetry data. AWS lists these core capabilities:
- Deterministic trace IDs derived from the X-Ray trace header, so every invocation of one execution shares a trace
- A span for each step, wait and callback wait, with real timings
- Continuation spans that link an operation finishing in a later invocation back to where it began
- Trace and span IDs added to log records automatically
- Standard OpenTelemetry sampling to control trace volume
The feature works in every Region where Lambda durable functions are offered.
How It Works
The Compute blog walks through a document review workflow with a human approval step. The first invocation extracts text and writes a summary, then suspends while it waits for a reviewer. When the reviewer answers hours later, Lambda calls the function again, replays the finished steps and publishes the result.
Without the plugin, that run produced two traces with different IDs. With it, both invocations land under one trace ID, and the long wait appears as a span without compute cost. During replay, the plugin spots operations that already ran and does not emit duplicate spans.
The blog describes two ways to view the data. An execution view groups every operation under a top-level Workflow span. An invocation view groups operations under the invocation that ran them, which suits workflows that may outlast a vendor's span ingestion limit, such as a cap of a few hours.
What It Means for Developers
For teams already running durable functions in production, this fills a real gap. Finding where a stuck order or a slow approval spent its time no longer means matching request IDs across separate traces. Step durations, retries and failures now line up on one timeline.
There are costs to plan for. AWS says the plugin and the ADOT layer add roughly 40 to 60 MB of memory use and lengthen cold starts. CloudWatch Application Signals charges per trace recorded and scanned, so high-volume workflows should set a sampling rate from the start.
The blog also gives practical advice. Always pass a name to each step and callback wait, since those names become span names. Pair the automatic trace IDs in logs with structured JSON logging to make Logs Insights queries easier.
If you are new to durable functions, the site has covered the Durable Execution SDK for Rust and the ability to pause, resume and delete executions.
How to Get Started
For Python, the blog sets up the plugin in four steps:
- Install the aws-durable-execution-sdk-python-otel package next to the SDK
- Add the AWS Distro for OpenTelemetry (ADOT) Lambda layer and set the exec wrapper variable
- Turn on X-Ray active tracing and give the role permission to write trace segments
- Pass the plugin factory to the plugins argument of the durable execution decorator
The plugin can also be switched on through an environment variable instead of code. Java users add a plugin artifact and load its JAR as an extension of the ADOT Java agent. AWS also published a sample application on GitHub showing a full traced workflow.
After deploying, publish a function version, because durable functions need a qualified ARN. Then invoke the function and look for a single trace in the CloudWatch Application Signals traces view.
Conclusion
OpenTelemetry support makes Lambda durable functions much easier to run in production. One trace now covers a whole execution across pauses, retries and callbacks, and teams that budget for the extra memory and sampling costs gain a clear view of long workflows.
Sources
- AWS What's New: AWS Lambda durable functions now supports OpenTelemetry tracing
- AWS Compute Blog: AWS Lambda Durable Functions observability with OpenTelemetry
- AWS Durable Execution SDK Reference: OpenTelemetry
See my articles more often when you search Google.
