sglang-prod-incident-triage

Debug SGLang serving incidents using a replay-first workflow that collects baseline data and captures failing requests to replay on a clean target. This approach isolates the root cause of health degradation or post-deploy behavioral changes.

Reproduce a CUDA crash or distributed hang by saving the failing request or crash dump from the live server. Replay this bundle on a clean target to trigger specialized tools for deterministic debugging.

Triage an SGLang incident by collecting a live bundle, saving the failing request, replaying it on a clean target, and then applying targeted tools for CUDA crashes or profiling to identify the root cause.

You need the standard Python requests library to run the SGLang incident triage workflow. The process relies on standard Python tooling and integrates optional scripts and references for end-to-end support.

Yes, you can use this replay debugging flow for post-deploy behavioral changes in SGLang. It captures the failing request from the live environment and replays it to resolve serving problems rapidly.

Replaying a failing request on a clean target helps troubleshoot SGLang serving problems by isolating the issue from live environment noise. This deterministic step pinpoints whether CUDA crashes or distributed hangs are caused by specific inputs.