[rush] rushd: a client that stops reading its output keeps the batch/lease after its operations finish; every other client blocks and fails with wait-timeout #6071

Summary

If one client stops reading its stdout (a pager, Ctrl+S, a suspended or slow agent), the daemon keeps that request's batch, and its execution lease, open until the client has drained all of its output. This continues even after that client's operations have finished. Every other request to the workspace queues behind it and fails with rush-client: daemon admission failed (wait-timeout). after 30 s, even an unrelated no-op build of a different project. The daemon buffers the client-bound frames in memory without bound in the meantime.

Repro steps

Two-project synthetic workspace; p01 prints 100k lines (3.7 MB) with --verbose; RUSH_DAEMON=1:

rush-client rebuild --verbose --only p01 2>&1 | (sleep 45; wc -c) &   # client A: its reader stalls for 45 s
sleep 3; rush-client build --only p02                                  # client B: unrelated no-op, 1.4-1.7 s when run alone

Expected result: A slow consumer only slows itself down. Once its operations have finished, the batch and lease are released and its remaining output drains independently (a bounded buffer, a spill to or replay from the operation log file, or detaching the client after a limit). Unrelated requests never wait on another client's terminal.

Actual result: p01's child process finished by t=16 s, but client B fails with exit 1 at 34.5 s (wait-timeout). With a 30 s stall, B finishes at 32.2 s instead of about 1.5 s. Daemon RSS grows with the buffered output (+10 MB for 15 MB).

Details

Root cause (main @ 60007c9): libraries/rush-daemon/src/PhasedRequestEventSink.ts:42-84 (OrderedClientWriter) chains every event and log chunk onto an unbounded promise tail, and writeLogChunk returns void, so the engine never feels backpressure. flushAsync() is then awaited before the request result and batch completion, so batch completion depends on the slowest client's terminal.

Suggested fix: decouple batch completion from per-client output draining. Release the lease when execution finishes, and drain each client's remaining output on that client's own connection with a bounded buffer. If a limit is exceeded, spill to the log file and tell the client, or disconnect it with an explicit error.

This was found during an automated performance/behavior analysis of rush-client/rushd on Linux (multi-agent scenario) and independently reproduced.

Standard questions