andygrove opened a new pull request, #2857: URL: https://github.com/apache/datafusion-ballista/pull/2857
# Which issue does this PR close? Follow-up to #2816 (part of #2808). This branch is stacked on #2816, so until that merges the diff here also includes its commits. Only the last commit is new. # Rationale for this change #2816 added `tcp_incoming_from_listener` so call sites with a pre-bound listener get the same `TCP_NODELAY` and keepalive that `create_grpc_server` would set, and documented `tcp_keepalive_seconds = 0` as disabling keepalive. Two listeners in the standalone executor were not switched over: - The Flight (shuffle) server served its listener through `TcpListenerStream`, so accepted connections got neither `TCP_NODELAY` nor TCP keepalive. tonic ignores the builder's settings on a pre-bound listener. - The gRPC server built `TcpIncoming` by hand with `Some(Duration::from_secs(..))`, so `0` produced a zero idle time instead of disabling keepalive. # What changes are included in this PR? - Both standalone executor listeners now go through `ballista_core::utils::tcp_incoming_from_listener`. - A comment on the `socket2` dev-dependency in `ballista-scheduler` explaining why it enables the `all` feature. # Are there any user-facing changes? Standalone executor Flight connections now get `TCP_NODELAY` and a 3600 s TCP keepalive, matching the executor binary. No API changes. 🤖 Generated with [Claude Code](https://claude.com/claude-code) -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
