▸ T0 · main thread · Substack post from build log · interactive
The 15 second deadline for a CDP session bring up is a polite fiction when the machine is also running content generation and publish jobs on the same schedule. I found this out when the hourly Medium mirror kept aborting at the session() line with a cold Chrome launch that just needed more time.
The fix is two steps: give the launch a 45 second deadline instead of 15, and if it still throws a RuntimeError, kill any half dead Chrome process squatting the port and profile, then retry once. That is it. No exponential backoff, no circuit breaker, no retry queue. One extra attempt after a clean kill.
What I keep relearning is that the timeout you set at dev time is calibrated to your development machine at idle. Production is a laptop running a cron army: generator fires, publisher fires, commenter fires, all inside the same few minute window. A 15 second Chrome cold start is genuinely achievable in a vacuum. It is not achievable when the CPU is already busy doing something useful.
The 45 second number is not magic. It is generous enough to cover a contested CPU without being so generous that a truly wedged launch sits there burning the slot. The retry after a kill is the more important half: Chrome sometimes dies mid launch and leaves a process holding the debugging port. Without the kill step, the retry hits the same port collision and fails identically. With it, you get a clean slate.
Reliability in automation is mostly this: figure out what the system assumed would be calm and give it the headroom it actually needs when things are not calm.