Thanks for the reply. If what I'm understanding is correct there's no chance of an OOM, but since direct memory is for I/O, it being completely filled may be a sign of backpressure? Currently one of our operators takes a tremendous amount of time to align during a checkpoint. Could increasing direct memory help checkpointing by improving I/O performance across the whole plan (assuming I/O is at least part of the bottleneck)?
On Tue, Dec 15, 2020 at 10:37 PM Robert Metzger <rmetz...@apache.org> wrote: > Hey Rex, > > the direct memory is used for IO. There is no concept of direct memory > being "full". The only thing that can happen is that you have something in > place (Kubernetes, YARN) that limits / enforces the memory use of a Flink > process, and you run out of your memory allowance. The direct memory is > allocated outside of the heap's upper limit, thus you could run out of the > budget. > But Flink is usually properly configuring the memory limits correctly, to > avoid running into those situations. > > tl;dr: you don't need to worry about this. > > > On Tue, Dec 15, 2020 at 8:38 AM Rex Fenley <r...@remind101.com> wrote: > >> Hello, >> >> Our job consistently shows >> Outside JVM >> Type >> Count >> Used >> Capacity >> *Direct* 32,839 1.03 GB 1.03 GB >> for direct memory. >> >> Is it typical for it to be full? What are the consequences that we may >> not be noticing of direct memory being full? >> >> Thanks! >> >> -- >> >> Rex Fenley | Software Engineer - Mobile and Backend >> >> >> Remind.com <https://www.remind.com/> | BLOG <http://blog.remind.com/> >> | FOLLOW US <https://twitter.com/remindhq> | LIKE US >> <https://www.facebook.com/remindhq> >> > -- Rex Fenley | Software Engineer - Mobile and Backend Remind.com <https://www.remind.com/> | BLOG <http://blog.remind.com/> | FOLLOW US <https://twitter.com/remindhq> | LIKE US <https://www.facebook.com/remindhq>
