Kafka has two sink transforms: KafkaSinkSingle and KafkaSinkCluster
KafkaSinkCluster has already been ported to use SinkConnection::try_recv instead of SinkConnection::recv.
But we still need to port KafkaSinkSingle which is the target of this issue.
SinkConnection::recv is an async method that asynchronously waits until at least 1 new response has arrived before returning.
SinkConnection::try_recv is a non-async method that immediately returns 0 or more responses.
recv introduces huge delays because it does not return until the server can respond to the request which will be in the milliseconds, which is an eternity for a proxy.
During this time shotover cannot process any new requests coming in from the client on this connection.
Previously we had to use the async wait in recv to avoid spin waiting.
But due to recentimprovements to shotover internals we can now return without waiting for all responses, so it is much more efficient to use try_recv instead.
Since we expect performance improvements from this PR, appropriate windsock benchmarks should be run to demonstrate the change actually improves performance.
Take a screenshot and post it in the PR.
prereq: https://github.com/shotover/shotover-proxy/issues/1609
This is part of https://github.com/shotover/shotover-proxy/issues/1507. Specifically the checkbox
swap single sinks to use SinkConnection::try_recv instead of SinkConnection::recvKafka has two sink transforms: KafkaSinkSingle and KafkaSinkCluster KafkaSinkCluster has already been ported to use
SinkConnection::try_recvinstead ofSinkConnection::recv. But we still need to port KafkaSinkSingle which is the target of this issue.SinkConnection::recvis an async method that asynchronously waits until at least 1 new response has arrived before returning.SinkConnection::try_recvis a non-async method that immediately returns 0 or more responses.recvintroduces huge delays because it does not return until the server can respond to the request which will be in the milliseconds, which is an eternity for a proxy. During this time shotover cannot process any new requests coming in from the client on this connection.Previously we had to use the async wait in
recvto avoid spin waiting. But due to recent improvements to shotover internals we can now return without waiting for all responses, so it is much more efficient to use try_recv instead.Since we expect performance improvements from this PR, appropriate windsock benchmarks should be run to demonstrate the change actually improves performance. Take a screenshot and post it in the PR.