For future reference here is a cross link to the referred ML thread discussion [1].
[1] http://mail-archives.apache.org/mod_mbox/flink-user/201904.mbox/%3cm2ef5tpfwy.wl-nings...@gmail.com%3E Cheers, Till On Wed, Apr 24, 2019 at 4:00 AM Ning Shi <nings...@gmail.com> wrote: > Hi Congxian, > > I think I have figured out the issue. It's related to the checkpoint > directory > collision issue you responded to in the other thread. We reproduced this > bug on > 1.6.1 after unchaining the operators. > > There are two stateful operators in the chain, one is a > CoBroadcastWithKeyedOperator, the other one is a StreamMapper. The > CoBroadcastWithKeyedOperator creates timer states in RocksDB, the latter > doesn’t. Because of the checkpoint directory collision bug, we always end > up > saving the states for CoBroadcastWithKeyedOperator. > > After breaking these two operators apart, they try to restore from the > same set > of saved states. When the StreamMapper opens the RocksDB files, it doesn’t > care > about any of the column families in there, including the timer states. > Hence the > error. > > -- > Ning >
