`io.grpc.internal.ServerImpl$JumpToApplicationThreadServerStreamListener$1OnReady` occupy too much memory

[lulu2panpan]

my grpc verseion is :<grpc.version>1.30.0</grpc.version>
my question is ：io.grpc.internal.ServerImpl$JumpToApplicationThreadServerStreamListener$1OnReady occupy too much memory

 num     #instances         #bytes  class name (module)
-------------------------------------------------------
   1:        166890     1899695256  [B (java.base@11.0.19.19-AJDK)
   2:      56705537     1360932888  java.util.concurrent.ConcurrentLinkedQueue$Node (java.base@11.0.19.19-AJDK)
   3:      56705251     1360926024  io.grpc.internal.ServerImpl$JumpToApplicationThreadServerStreamListener$1OnReady
   4:         29061       26383648  [Ljava.lang.Object; (java.base@11.0.19.19-AJDK)
   5:         13494        8852064  io.grpc.netty.shaded.io.netty.util.internal.shaded.org.jctools.queues.MpscArrayQueue
   6:           857        4990072  [C (java.base@11.0.19.19-AJDK)
   7:         28351        2590160  [J (java.base@11.0.19.19-AJDK)
   8:         93460        2243040  java.lang.String (java.base@11.0.19.19-AJDK)
   9:         28646        1833344  io.grpc.netty.shaded.io.netty.buffer.PoolSubpage
  10:         65536        1572864  com.aliyun.polardbx.binlog.dumper.dump.logfile.parallel.EventData
  11:         12101        1445832  java.lang.Class (java.base@11.0.19.19-AJDK)
  12:         42613        1363616  java.util.concurrent.ConcurrentHashMap$Node (java.base@11.0.19.19-AJDK)

[lulu2panpan]

Haha, I encountered this problem three years ago，but i forgot it，should do my logic in another thread

#7839
#7361

[ejona86]

Yes, blocking the callback thread the way you were can accumulate onReady() callbacks. There could be some ways to fix it in gRPC, but it isn't probably worth it given how pathological your approach is (you don't even learn if the RPC died, because onClose() can't be called).

You should take a look at 1.70 though (the next release), because we merged #10318 which has a blocking API for streaming. That will probably yield much better results than hard-coded polling.