[utay]

Summary

• Enable client-side gRPC keepalive: ping after 30s of silence, treat the connection as dead if the ping goes unanswered for 20s (matching the server's settings).
• Add a regression test that cuts a connection mid-statement and expects FlightUnavailableError instead of a hang.

If a connection is cut while the client waits for a statement result, only the server notices. The client has nothing to send, so its TCP stack never retransmits and the call hangs forever. Callers such as dbt-altertable get the fix without any change on their side.

Channel options added to FlightClient:

• grpc.keepalive_time_ms=30000 and grpc.keepalive_timeout_ms=20000.
• grpc.http2.ping_timeout_ms=20000: the gRPC bundled with recent pyarrow releases times keepalive pings out with this option (default 60s), not keepalive_timeout_ms.
• grpc.http2.max_pings_without_data=0: by default gRPC stops pinging after two pings while the client sends no data, which is exactly the long-statement case.
• grpc.keepalive_permit_without_calls=1: also detect dead idle connections between statements.

The server's tonic/h2 stack does not reject frequent client pings, so a 30s interval is safe.

Testing

tests/test_keepalive.py runs a minimal grpcio Flight server behind a TCP proxy. It uses grpcio rather than pyarrow's server because older pyarrow servers (21, used on Python 3.9) reject frequent pings with too_many_pings, unlike the production tonic server. It keeps the connection healthy for 5s while a statement blocks, then silently stops forwarding bytes and asserts the call fails within 15s. The test fails if either ping_timeout_ms or max_pings_without_data is removed.

Made with Cursor