[aletorrado]

Hi! I work in production with workers in cluster_mode + wait_ready + kill_timeout + listen_timeout. I've been noticing that sometimes, after just after a cold start (while booting a server), some workers randomly were marked as online but with memory = 0.

I've debugged this issue with the help of Claude, and came to the conclusion that there was a bug in PM2 which got fixed by commit c87dd75. Nevertheless, Claude found another bug introduced in that same commit in which callbacks are clashing. I'm leaving the full description co-authored with Claude, including reproduction code below for your reference.

Problem

Since c87dd75 ("process flow refactor"), when a process exits before it is ready, both exit handlers in God.executeApp (cluster and fork) call readyCb() to answer the caller. readyCb() also sets status = online (and logs/notifies online) for a process that has just exited.

When the exit was caused by pm2 stop, this overwrites the stopping status. handleExit() then no longer sees a stop, treats the exit as a crash, and restarts the app.

Reproduction (7.0.4 and development):

// app.js: never sends 'ready'
setInterval(() => {}, 1000);

// ecosystem.config.js
module.exports = { apps: [{ name: 'repro', script: 'app.js', exec_mode: 'cluster',
  wait_ready: true, listen_timeout: 5000 }] };

pm2 start ecosystem.config.js & sleep 2; pm2 stop repro; sleep 5; pm2 ls

Expected: stopped. Actual: the app was restarted (↺ 1) and is online. The log shows Stopping app, exited with code [0], online, starting in -cluster mode-.

Fix

Add exitedCb(), which concludes the caller (keeping the no-hang guarantee from c87dd75) without marking the process online. Both exit handlers use it, so handleExit() sees the real status.

The cb_called guard still makes the pending ready timer / process:msg listener a no-op once the process has exited.

Context

On 6.x the same path showed up differently: the ready timer was not cleared on exit, so listen_timeout later marked an exited process online with pid 0 / memory 0 (e.g. after pm2 stop during boot, or during a restart_delay wait). c87dd75 fixed that part; this PR fixes the restart regression it introduced.

Tests

• New test/programmatic/stop_before_ready.mocha.js (cluster and fork mode), registered in test/unit.sh. It fails without the fix (expected 'online' to equal 'stopped') and passes with it.
• Still passing: cluster_exit_before_online, issue_6142_stale_exit_event, issue_6129_reload_collision, graceful, cluster, god, auto_restart, exp_backoff_restart_delay.

[CLAassistant]

All committers have signed the CLA.