Hi, the result of splitting by key is that processing can easily be distributed among the workers because the windows for individual keys can be processed independently. This should improve cluster utilization.
Cheers, Aljoscha On Tue, 28 Jun 2016 at 17:26 Vishnu Viswanath <vishnu.viswanat...@gmail.com> wrote: > Hi, > > Thank you for the responses. > I am not sure if I will be able to use Fold/Reduce function, but I will > keep that in mind. > > I have one more question, so what is the implication of having a key that > splits the data into window of very small size(=> large number of small > windows) ? > > Thanks and Regards, > Vishnu Viswanath, > > On Tue, Jun 28, 2016 at 5:04 AM, Aljoscha Krettek <aljos...@apache.org> > wrote: > >> Hi, >> one thing to add: if you use a ReduceFunction or a FoldFunction for your >> window the state will not grow with bigger window sizes or larger numbers >> of elements because the result is eagerly computed. In that case, state >> size is only dependent on the number of individual keys. >> >> Cheers, >> Aljoscha >> >> On Tue, 28 Jun 2016 at 10:36 Kostas Kloudas <k.klou...@data-artisans.com> >> wrote: >> >>> Hi Vishnu, >>> >>> RocksDB allows for storing the window contents on disk when the state of >>> a window becomes too big. >>> BUT when you have to trigger and apply the computation of your window >>> function on that big window, >>> then all of its state is loaded in memory. >>> >>> So although during the window formation phase, RocksDB allows you to not >>> worry about storage space, >>> when it is time to fire your computation, then you have to consider how >>> much RAM you have and if the >>> window fits in it. >>> >>> Regards, >>> Kostas >>> >>> >>> On Jun 28, 2016, at 1:25 AM, Vishnu Viswanath < >>> vishnu.viswanat...@gmail.com> wrote: >>> >>> Hi Kostas, >>> >>> Thank you. >>> Yes 2) was exactly what I wanted to know. >>> >>> - So if I am using RocksDB as state backend, does that mean that I don't >>> have to worry much about the memory available per node since RocksDB will >>> use RAM and Disk to store the window state? >>> >>> Regards, >>> Vishnu >>> >>> >>> On Mon, Jun 27, 2016 at 6:19 PM, Kostas Kloudas < >>> k.klou...@data-artisans.com> wrote: >>> >>>> Hi Vishnu, >>>> >>>> I hope the following will help answer your question: >>>> >>>> 1) Elements are first split by key (apart from global windows) and then >>>> are put into windows. In other words, windows are keyed. >>>> 2) A window belonging to a certain key is handled by a single node. In >>>> other words, no matter how big the window is, its >>>> state (the elements it contains) will never be split between >>>> two or more nodes. >>>> 3) Where the state is stored, depends on your state backend. Currently >>>> Flink supports an in-memory one, a filesystem one, and >>>> a rocksDB one which is in the middle (first in-memory and then >>>> disk when needed). Of course you can implement your own. >>>> >>>> From the above, you can see that if you use the memory-backed state >>>> backend, then your window size is limited by the memory >>>> available at each of your nodes. If you use the fs state backend, then >>>> your state is stored on disk. Finally, rocksDB will initially >>>> use RAM and then spill on disk when no more memory is available. >>>> >>>> Here I have to add that the window documentation is currently being >>>> re-written to explain new features introduced in Flink 1.1, >>>> which include more flexible handling of late events and more explicit >>>> state garbage collection. >>>> >>>> So please stay tuned! >>>> >>>> I hope this helps at answering your question, >>>> Kostas >>>> >>>> > On Jun 27, 2016, at 8:22 PM, Vishnu Viswanath < >>>> vishnu.viswanat...@gmail.com> wrote: >>>> > >>>> > Hi All, >>>> > >>>> > - Is there any restriction on the size of a window in Flink with >>>> respect to the memory of the nodes? >>>> > - What happens if a window size grows more than size of a node, will >>>> it be split into multiple nodes? >>>> > >>>> > if I am going to have a huge window, should I have fewer nodes with >>>> more memory. >>>> > Is there any documentation on how memory is managed/handled in the >>>> case of windows and also in the case of joins. >>>> > >>>> > Regards, >>>> > Vishnu >>>> >>>> >>> >>> > >
