| It's part of a live video platform for a gaming company which needs extra information to be overlaid on top of their live stream (for the customer in-video frame performance is incredibly important: the stream run at 30fps (so a maximum of 33ms per frame) and the overlay information has to be generated in 11ms or less) We did originally consider writing it in Rust (it would have been my personal preference) but the consensus was that since the rest of the tooling for this customer is in PHP we didn't want to add a skillset they didn't have internally if they were to move away from us in the future (we aren't big on needless vendor lock-in and prefer to compete on quality and service). We experimented with PHP7.4 and Symfony / Laravel and found that we didn't have the performance leeway we needed, even with preloading / opcache. We opted for Either Swoole or Phalcon, tried them both and found that Swoole was more to our liking and was somewhat more performant than Phalcon in our real-world testing. We are exposing Swoole directly to them within their VPC, so we were able to do away with NGINX as a reverse proxy as well and have Swoole handle the HTTP request phase and HTTPS. We were able to make heavy use of Swoole\Table to save superfluous trips to Redis and run a Redis synchronisation process in a coroutine thread when the concurrent number of network tasks drops below 16. Our end service response times (including network overhead) are around here: - Mean response time: 4ms - Fastest: 2ms - 99th percentile: 9ms - 999th permille: 10ms One unforeseen side effect is that it has a crazy amount of scalability built in by default, the 8core machine the process runs on can easily scale up to 15,000 concurrent requests (as in 15,000 requests per second) without going over 15ms for response times, and can scale up to 45,000 concurrent requests before things start to "break". |