Why is the server streaming responses in batches instead of one by one?

Yan Yan

unread,

Aug 3, 2016, 11:18:00 PM8/3/16

to grpc.io

commit 7f6c7798b08afcbf77e675e0bc4e8ce5b6ba580d

Date: Thu Jul 14 18:10:19 2016 -0700

route_guide.proto

rpc ListFeatures(Rectangle) returns (stream Feature) {}

I am running the above version of gRPC python. I have tried throttling the server to send 100, 500 or 1000 responses per second. I have found the client to receive responses in 4k batches. How can I send them ASAP? Is it related to TCP_NODELAY? Does gRPC c++ offer a way to send them ASAP? Is there a way to flush earlier? Is there away to adjust the batch size or buffer size? Thx.

Yan Yan

unread,

Aug 3, 2016, 11:26:59 PM8/3/16

to grpc.io

https://github.com/grpc/grpc-go/issues/524#issuecomment-180536815

Go's http2 server does this aggregation automatically. It has a write scheduler that only flushes only when a packet is full or there's nothing else to send.

Yan Yan

unread,

Aug 8, 2016, 9:51:05 PM8/8/16

to grpc.io

I am now able to do 1 streaming response/sec by NOT reading data from stdin. Python stdin probably has a 4096 line buffer.

On Thursday, August 4, 2016 at 11:18:00 AM UTC+8, Yan Yan wrote:

Reply all

Reply to author

Forward