What is implicit in the parent task T964:
- out of ram worker is killed (it cannot clean up since it's killed)
- the node running the worker is then mostly idle for that particular work (in regards to the other sister nodes)
- so it starts consuming the queue faster than the other workers (since they do actual work)
- and fails faster
- resulting in an empty queue in the end
That is what i was trying to solve in T964 (well finding proper solution to implement for the moment).
As I realized it was not explicitly mentioned, opening a dedicated issue for it.