Briefing
The order-scoring workers were redeployed at 09:23 as part of a manifest cleanup.
Since then both pods have restarted six times, each run lasting only a few seconds, and the orders.score backlog has passed a thousand messages.
The worker image is the same 14 January build that has run all week.
Dev points out that the data team retrained the scoring models on Monday and suspects the bigger models are the problem, so he would like to pin the old model files.
The fix is yours to find: hints, the recovery checks and the debrief unlock inside the incident.