Hints for a Custom Marshaller(KnownLength vs Drainable)

Using gRPC version 1.7.0 with Netty Epoll transport.

I am using a custom marshaller for non-proto messages. For InputStream provided by Marshaller.stream(message), currently I don’t have a way to provide length of the message without serializing it. So, I have these two options:

1. Support KnownLength by serializing the message to byte[] which would be helpful in reducing the number of ByteBuf allocations (in MessageFramer when draining the message to OutputStream) as length is known. But this approach loses the benefit of Drainable, where we would be copying the byte[] to OutputStream.

2. Don’t support KnownLength and support only Drainable to avoid additional copy when draining to OutputStream. But this may lead to lot more ByteBuf allocations as message length is not known. (Due to usingBufferChainOutputStream in MessageFramer).

Could you share your thoughts on which approach provides lower latency? (Especially if we have large messages of 10 MB or so.)

[ejona86]

It seems this is answered reasonably for now. @srujann, comment more if you need more help on this topic and we can reopen. Closing.

[ejona86]

@srujann, since you can’t easy obtain the length, I’d suggest not implementing KnownLength. I’d suggest implementing Drainable if it is the more “natural” interface for your serialization logic, and if it avoids allocating temporary byte[]s. Basically, do what’s natural/cheap and let gRPC handle the rest.