You can have timestamps that are very much out-of-order (in the future, specifically). The window operator assigns them to the specific window. The window operators can hold many windows concurrently, which are in progress at the same time.
Windows are then flushed once the triggers fire (after a time, or at a watermark). On Wed, Feb 3, 2016 at 12:11 PM, Aljoscha Krettek <aljos...@apache.org> wrote: > Hi, > with TPS you mean tuples-per-second? I have an open pull request that > changes the WindowOperator to work on a partitioned state abstraction. In > the pull request I also add a state backend that uses RocksDB, so it it > possible. > > The size of the windows you can keep also depends on the window function, > if it is a ReduceFunction then the window result can be incrementally > computed and the state that we have to keep is very small. For a > WindowFunction that takes an Iterable of all the window elements the state > can grow very large, of course. > > Cheers, > Aljoscha > > On 03 Feb 2016, at 11:17, Anwar Rizal <anriza...@gmail.com> wrote: > > > > point > >
