Abhishek Pal created RATIS-2637:
-----------------------------------
Summary: Netty RPC lacks per-connection read backpressure
Key: RATIS-2637
URL: https://issues.apache.org/jira/browse/RATIS-2637
Project: Ratis
Issue Type: Improvement
Components: Netty
Reporter: Abhishek Pal
Assignee: Abhishek Pal
After Netty RPC handling is offloaded to a worker pool,
NettyRpcService.InboundHandler.channelRead0 reads every inbound request off the
socket immediately and chains it as a pending CompletableFuture stage (holding
the request proto) on the shared request executor (unbounded
LinkedBlockingQueue).
Reads are never paused, so there is no TCP backpressure.
We should cap in-flight requests per channel and when the cap is reached set
`ctx.channel().config().setAutoRead(false)` and re-enable once drained. This
bounds memory, restoring TCP backpressure.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
