Farhan Ali Khan
Sep 5, 2018 · 1 min read

Good article. In the horizontal scaling section, I think CPU instructions and messages/sec are different things, I guess you are talking about the RPS. On a 8 core machine, are you really able to get a throughput of 8 million rps, taking web sockets into account? IMHO it seems too much, correct me if I have understood it wrong.

Also in the BST approach you’ve described, instead of a linked list, won’t an array be better? Provided you don’t store thousands of items/channels in a BST, and the only benefit you are getting is on a delete O(log N), which anyways an insertion would make up for(unless your deletes are more in magnitude than writes), wont an array benefit from cache locality, and that would be beneficial in a real time system? Have you tried the approach?

    Farhan Ali Khan

    Written by

    Philomath, developer, Software Engineer