The runtime had to move without a one-shot cutover.
The application was on PHP 7.3 and served about 100,000 daily active users. Staging could not exercise every runtime assumption or dependency path. A single cutover would have turned an unknown incompatibility into a full outage.
Two pools, and a rollback that was just a traffic change.
Bhavin led the migration. Two Cloudflare Load Balancer pools ran side by side: a stable pool on PHP 7.3, and a candidate pool on PHP 8.4. Traffic started at 95% on 7.3 and 5% on 8.4. The candidate pool grew only when production behavior showed the compatibility work was holding.
If a route failed on 8.4 and stayed healthy on 7.3, the rollback was mechanical: disable the candidate pool, send traffic back to the stable pool, fix the issue, and resume the canary. That kept rollback as a traffic-control step, not an emergency deploy.
The result
The application moved from PHP 7.3 to PHP 8.4 while serving those ~100,000 daily users. The rollout began at 95/5 and advanced as live traffic supported it.
On client systems that cannot afford a big-bang cutover, Developer Labs uses the same pattern: introduce the change under limited exposure, watch it, and keep a way back.
Bhavin wrote a longer engineering account on his personal site: PHP 8.4 production migration.