My opinion: If this change allows to reduce GC challenges (so more memory can be allocated to the heap) and increases the performance, this might be managable...
2016-02-02 14:47 GMT-05:00 Enis Söztutar <enis....@gmail.com>: > BTW, we should also be looking at https://google.github.io/flatbuffers/ or > https://capnproto.org/ for serialization as an option. The idea is to not > allocate objects and prevent allocations altogether. > > We are allocating PB objects for every Get / Put, then we allocate our Get > / Put objects. At least we can save on 1. > > Although, just switching our serialization format again will be a huge > undertaking with obvious wire-incompatibility issues. If PB3 or 2.x gives > us what we want in terms of preventing big byte[] allocations, we would > gain regardless. > > Enis > > On Tue, Feb 2, 2016 at 11:41 AM, Enis Söztutar <enis....@gmail.com> wrote: > > > Google guys over at > > https://github.com/grpc/grpc-java/issues/1054#issuecomment-147295224 are > > saying that CIS changes may be coming to 2.x from what I understand. If > so, > > our life would be easier. Even so, I'm 100% sure we have to do shading > > since Hadoop will not change it's PB dependency anytime soon. > > > > We have to do this before doing shading: > > https://issues.apache.org/jira/browse/HBASE-15174 > > > > Enis > > > > On Tue, Feb 2, 2016 at 8:15 AM, Stack <st...@duboce.net> wrote: > > > >> Thanks Duo. If proto3 had what we wanted, you are suggesting we might > move > >> to proto3 setting it to do proto2 support and shade it so we don't clash > >> with other includes of pb? > >> > >> Regards Anoop comment, the note on the end of this issue looks promising > >> but I don't know when it'd see the light of day: > >> https://github.com/grpc/grpc-java/issues/1054#issuecomment-147295224 > >> > >> St.Ack > >> > >> > >> On Mon, Feb 1, 2016 at 10:49 PM, Anoop John <anoop.hb...@gmail.com> > >> wrote: > >> > >> > UnsafeByteStrings - This may help us to avoid copy even with out our > >> > HBaseZeroCopyByteString stuff. But with a DirectByteBuffer, it has to > >> copy > >> > data to onheap byte[]. We even want a DBB backing ! > >> > > >> > -Anoop- > >> > > >> > On Tue, Feb 2, 2016 at 12:07 PM, 张铎 <palomino...@gmail.com> wrote: > >> > > >> > > https://groups.google.com/forum/#!topic/protobuf/wAqvtPLBsE8 > >> > > > >> > > PB2 and PB3 are wire compatible, and of course, protobuf-java is not > >> > > compatible so dependency will be a problem... But I think the shaded > >> > client > >> > > and server can solve the problem? > >> > > > >> > > Thanks. > >> > > > >> > > 2016-02-02 14:27 GMT+08:00 Stack <st...@duboce.net>: > >> > > > >> > > > We are running into a few issues with protobufs. > >> > > > > >> > > > + PB always copies all data before making a Message. This > generates > >> > > garbage > >> > > > unnecessarily. > >> > > > + CodedInputStream does not support ByteBuffers in 2.5. In 2.6 it > >> does > >> > > but > >> > > > again, copies the data out of the BB always; this is especially > >> painful > >> > > > when the BB is a DBB with its data offheap and intent is to keep > >> data > >> > > > offheap. > >> > > > > >> > > > There are other issues. CIS allocates 4k buffers regardless (See > >> > > > HBASE-15177). > >> > > > And then there was the HBaseZeroCopyByteString fun and games we > had > >> a > >> > > while > >> > > > back. > >> > > > > >> > > > 3.0 PB adds UnsafeByteStrings so can do zero copy. Thats good. But > >> PB3 > >> > is > >> > > > incompatible with PB2 (or at least, it looks like PB2 clients > can't > >> > talk > >> > > to > >> > > > PB3 [1]). > >> > > > > >> > > > There is javanano protobufs. All is open access, but it too looks > >> > > different > >> > > > to PB2 (i've not tried it). > >> > > > > >> > > > Protostuff seems really quiet these times [2]. > >> > > > > >> > > > Fork (and shade)? > >> > > > > >> > > > Thoughts? > >> > > > > >> > > > St.Ack > >> > > > > >> > > > 1. https://github.com/google/protobuf/releases > >> > > > 2. https://groups.google.com/forum/#!forum/protostuff > >> > > > > >> > > > >> > > >> > > > > >
