Found during integration runs for draft PR #712 (branch chtypes-v2).
Problem. With one idle HTTP keep-alive connection held open to the server, stopping ClickHouse 26.8 with SIGTERM took about 30 s, against about 5 s on 26.6. Measured in integration runs on both versions. WaveHouse's HTTP reader holds idle keep-alives (see also the client-pruning gap in the reader issue filed alongside this), so a WaveHouse that has queried a server can slow that server's shutdown. Whether WaveHouse is the only client that does this is unverified.
Repro. Start clickhouse/clickhouse-server 26.8 and 26.6, open one HTTP connection with keep-alive and leave it idle, send SIGTERM, time to exit.
Possible direction: close idle reader connections on shutdown, or bound the reader's idle timeout.
Related: #383, #592 (other shutdown-latency reports).
Found during integration runs for draft PR #712 (branch
chtypes-v2).Problem. With one idle HTTP keep-alive connection held open to the server, stopping ClickHouse 26.8 with SIGTERM took about 30 s, against about 5 s on 26.6. Measured in integration runs on both versions. WaveHouse's HTTP reader holds idle keep-alives (see also the client-pruning gap in the reader issue filed alongside this), so a WaveHouse that has queried a server can slow that server's shutdown. Whether WaveHouse is the only client that does this is unverified.
Repro. Start
clickhouse/clickhouse-server26.8 and 26.6, open one HTTP connection with keep-alive and leave it idle, send SIGTERM, time to exit.Possible direction: close idle reader connections on shutdown, or bound the reader's idle timeout.
Related: #383, #592 (other shutdown-latency reports).