+1 have seen scenarios where operators that don't utilize large state management take a while to load state from recovery.
On Tue, Oct 11, 2016 at 5:05 PM, Sandesh Hegde <[email protected]> wrote: > Hi All, > > Operators that are not making progress are killed after > TIMEOUT_WINDOW_COUNT. But this doesn't take the operator recovery into > account. Operators that are recovering may need more time than what is set > in TIMEOUT_WINDOW_COUNT. > > The new attribute, RECOVERY_WINDOW_TIMEOUT_COUNT, will be used to decide > the timeout when the operators are recovering. > > Thoughts? >
