sglang-prod-incident-triage

To debug SGLang serving incidents, you can use replay-based debugging. Capture a baseline bundle from the running server, save the failing request or crash dump, then replay it on a clean target to isolate the issue.

Replay-based debugging for serving incidents involves capturing a baseline bundle from a live server and replaying the failing request on a clean target. This isolates health failures, latency regressions, and crashes into repeatable workflows.

Collect a read-only baseline bundle directly from the running SGLang server. Preserve the failing request or crash dump within this bundle to replay the exact failure state later on a clean target.

Yes, you can replay SGLang incidents to diagnose distributed hangs. After replaying on a clean target, the workflow guides you to switch to targeted debug tools like CUDA crash analyzers or profilers as needed.

Yes, the requests library is required as a dependency. It supports the scripts and references used to capture the baseline bundle and execute the replay-based debugging workflow.

The SGLang incident triage process provides a structured output detailing the problem class, the strongest signal identified during replay, and recommended next steps to reach a production-ready resolution.