There are use cases where a Task reads from a rocksDB table in its init and sets timers.
The problem is the TimerListener is initialized later in the task-initialization flow.
This causes problems because invocations corresponding to timestamps in the past get ignored.
One job is using this timer to issue re-transmissions to kafka (after reading the re-transmission list from rocksDB).
So currently, on container-failoer, some of their retransmissions get ignored.
Current hacky-fix is to add a 2-second offset to the timer-timestamps to try and ensure task-init and timerlistener completes before the timers-invocation.