I tried to reproduced that one with several tactic, and never succeeded. I will close it and hope for more details if it happen again.
I learned some things along the way, noting them for post-mortem:
- there used to be a bug in old jetty: https://github.com/eclipse/jetty.project/issues/1357
- this bug was corrected in https://github.com/eclipse/jetty.project/commit/0d4fcd0cb4330e1895def6f5daa8996f8e23a6de, ie jetty 9.2.22 and 9.3.21 ,
- tests were latter added in https://github.com/eclipse/jetty.project/commit/b0bead2bc0de39c57eb8e531d72bd3f201ce7018 (jetty 9.4.7)
- rudder is in jetty 9.4.11 or up since at least 5.0 branche
So it is unlikely that the problem is due to that bug, and that jetty bug lead to torought testing.
Nonetheless, the behavior of the rotation is VERY surprising. It is not linked to the actual hour of the system, but to a latency since last rotation (or app start):
- when the app start or a rotation happens, jetty looks at current time and schedule a rotation in (24h-current time).
- that means that is time change, rotation will happen not at 00:00 but after computed latency is finished,
- that also may mean that IF a rotation can't happen, none will happen ever again.
That last option is what could explain the observed bug, but I can't see how it could happen without a big problem (like thread killed or something). And we didn't saw anything like that in logs.
Any clue on that problem would be much appreciated.