Expand description
Recovery of runs abandoned by a dead worker.
A worker attaches a lease to every run it picks up and refreshes it while it
executes. When the worker dies (OOM, evicted pod, hard shutdown), the lease
stops being refreshed and the run would otherwise stay Running forever.
The Reaper periodically requeues those runs.
Runs executed in-process by the API server (inline execution, resume after approval) never hold a lease and are never touched by the reaper.
Structs§
- Reaper
- Periodic task that requeues runs whose worker lease expired.
Constants§
- DEFAULT_
REAPER_ BATCH_ SIZE - How many runs a single tick recovers.
- DEFAULT_
REAPER_ INTERVAL - How often expired leases are collected.