[quinnj-bot]

Why

Once MySQL.jl has opened a connection, a Ctrl-C during a long statement can vanish, and reaping stops afterwards (#254). Julia throws a SIGINT's InterruptException into whichever task is resident on thread 1. The reaper wakes every 0.5 s and parks again, so during a long statement it is usually that task. Its catch around wait(t) took the exception, ended the loop, and left the timer open.

What changed

• reaper_loop catches nothing around wait(t). An InterruptException now leaves the task, which is what Base's task_done_hook needs to hand it to the REPL backend.
• ensure_reaper! tracks the task in REAPER_TASK and starts a new one on the next connect when the old one is done.
• The reaper task goes on the default pool, so with an interactive pool (-t N,M, or -t auto on 1.12) it never parks on thread 1 while the user's work runs there.

Scope

• Reaper only. The statement-id parking path has no background task.
• Does not change where Julia delivers a SIGINT. With one thread the reaper is still the resident task most of the time. The interrupt then ends the reaper; Base forwards it to the REPL when one is evaluating, and a script sees what it sees for any finished task on thread 1: the first Ctrl-C is lost, the second terminates the process. Base's own Timer(cb) task and the Profile SIGINFO listener behave the same way.
• Between the interrupt and the next connect, queued transports wait. The restarted reaper or the atexit hook closes them.

Tradeoffs

• Keeping the catch and continuing (option 2 in The reaper task swallows a Ctrl-C (InterruptException) meant for the main task, and then stops reaping #254) keeps the reaper alive, but every later Ctrl-C is lost too, and a one-thread script can no longer be stopped with Ctrl-C at all. Today its second Ctrl-C terminates it.
• Threads.@spawn :default is the public way to pick a pool, but its closure thunk is not reachable from the static call edge --trim needs. Task(reaper_loop) plus Threads._spawn_set_thrpool keeps the trim story as it is.

Blast Radius

The reaper's task lifecycle only. The CAS states, exactly-once close, reap_now!, and the atexit hook are untouched. The reaper now runs on default threads when an interactive pool exists; transport close was already reached from any thread.

Verification

• The repro from The reaper task swallows a Ctrl-C (InterruptException) meant for the main task, and then stops reaping #254 (no server), Julia 1.12.6, macOS. Before, with -t 2,1, 3 of 3 runs lost the interrupt and left the reaper done with its timer open. After, 2 of 3 runs delivered InterruptException to the main task after 1.1 s, and the reaper kept reaping in all 3; the third run lost the interrupt inside the runtime, the same 1-in-3 drop a MySQL-free control shows on this Mac. With -t 1, before: swallowed and reaping stopped; after: the interrupt ends the reaper task and ensure_reaper!() restarts it.
• The new testset in test/protocol/native_tests.jl fails on main (5 of 8 assertions: the task ends normally, nothing restarts it, a queued entry is never closed) and passes with the fix on -t 1 and on -t auto,2, on 1.12.6 and 1.10.11. The test finds the parked reaper through the timer's wait queue under its lock, which holds tasks before Julia 1.15 and WaitEntry records on 1.15-dev (the first CI round failed only the nightly jobs, on that check).
• Serverless suite (MYSQL_INTEGRATION=0 Pkg.test()), macOS arm64: Julia 1.12.6 passed 1823 of 1823 (the --trim=safe lane compiled and ran the workload, which exercises the reaper), Julia 1.10.11 passed 1816 of 1816.

Review notes

Checked as a reviewer: ensure_reaper! has one caller (register! in connect.jl); wait(::Timer) throws only EOFError for a closed timer, which nothing closes, or an exception thrown into the task; the restart runs under REAPER_SETUP_LOCK, so two concurrent connects start one task; atexit is still registered once; the test holds the timer's lock between the parked check and the throw, so schedule(..., error=true) never reaches a running task; _spawn_set_thrpool has the same signature on 1.10, 1.12, and master.

Fixes #254

🤖 Generated with Claude Code