We should have all parts of Flink documented that well. Very nice overview, Stephan!
Two comments: Considering the lazy memory segment allocation. Besides, the problem of initial garbage collection of these dynamically allocated segments, wouldn't it also have the disadvantage that we could not guarantee a fixed amount of memory? How would we check that the user memory does not take away the managed memory? I would add a note to why java.misc.Unsafe is faster for longs. I would guess because of the coding of the long (little or big endian). Best, Max On Wed, Mar 18, 2015 at 5:21 PM, Stephan Ewen <se...@apache.org> wrote: > Hi all! > > Here is a first version of the documentation how memory management works in > Flink. > > I hope it sheds some light on the magic we do. Let me know if certain > sections are still confusing. > > https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=53741525 > > Greetings, > Stephan >
