PARAS MADAN
All writing

Latest in AI

CloudWatch Omni for Azure: Verify One Cross-Cloud AI Request Before Expanding Ingestion

Pilot Azure Application Insights traces in CloudWatch Omni, check which records become spans, and verify request correlation before relying on it.

An AI application can begin a request in Azure and wait on a dependency in AWS. Seeing both clouds in one console helps, but it doesn't prove the records describe the same request. That's what I'd verify before relying on CloudWatch Omni in an incident.

AWS added Azure telemetry collection to CloudWatch Omni on November 21. It includes platform metrics, resource logs, virtual network flow logs and Application Insights traces. Rather than importing all of them, start with one inference path and decide what you need to see when it slows down.

Pick a request you can follow

Suppose a document summarizer runs a frontend service in Azure and calls a model proxy in AWS. Choose a synthetic document and a test request you can identify in both systems. Record the expected request path and which timestamps or identifiers you can use to find it.

Get a trustworthy baseline in the source systems before changing collection. Find the Azure request and its dependency record, then find the corresponding AWS-side record. If one side doesn't generate the signal you need, adding a connector won't create it.

Choose the right Azure source

The Azure telemetry setup guide distinguishes existing platform telemetry from telemetry collected inside a VM or application. For traces, the source is the Application Insights resource. Its AppRequests and AppDependencies records can become spans, while records such as AppTraces are imported as log events. AWS also requires Transaction Search to be enabled in the receiving AWS account and Region for trace collection.

For the pilot, scope the connection to the test subscription, Region and supported resource type. Confirm the actual Application Insights resource is included rather than selecting only the service hosting the code. Ask the environment owners to review what will leave Azure and what is already recorded there. Keep prompts, model outputs and customer data out of the test traffic and review logs for fields that should not be forwarded.

Check collection before looking for correlation

Azure resource logs and traces pass through Event Hubs. The AWS guide tells you to look for an ACTIVE telemetry pipeline and match its active source count to its total. Then look for imported spans in Transaction Search with cloud.provider set to azure.

That gives you two separate checkpoints: the intake is running, and the expected test signal arrived. If the pipeline is active but the request isn't visible, check the source records, scope and AWS destination before concluding that the inference request failed.

For the first test, record at least one successful request and one controlled slowdown in staging. Make the slowdown easy to explain, for example by returning a delayed response from a test dependency. This gives the on-call reviewer a known answer without disrupting real workloads.

Don't confuse an Azure dependency span with an AWS trace

An Azure dependency span can tell you how long the Azure caller waited. It doesn't automatically explain how the AWS service spent that time. Check whether the AWS-side span exists and whether the trace context actually ties it to the Azure request.

If you only see two nearby timestamps, label that as a manual correlation, not a complete trace. Check instrumentation, context propagation and sampling in the participating services before using the combined view to assign ownership of a slowdown.

Have a second engineer repeat the investigation from the shared console. Give them the time window and test request, not the expected answer. If they can't distinguish a slow AWS dependency from missing AWS telemetry, the pilot needs another pass.

Decide what to keep before expanding

Measure what arrived and decide how much of it is useful. Budget for both cloud providers' collection resources and for AWS ingestion and retention. Do not add flow logs or unrelated subscriptions to a trace pilot just because they are available.

Write down who owns both sides of cleanup. AWS notes that removing items from collection scope does not delete existing Azure resources or diagnostic settings. Review those separately when ending the pilot.

The rollout is ready when an engineer can follow the same request, point to evidence for the slowdown and say which parts of the path are still missing.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams explain technical products through clear documentation, practical guides and developer-focused content.

Book a 30-minute call.