Türker Yıldırım

performance

  • Alerting that people do not learn to ignore

    Deep dive10 min read

    340 alerts in a month, four of which mattered. An alert channel nobody reads is the same as no alerting, and the cause is alerting on causes.

  • Measure at the percentile that hurts

    Note1 min read

    A mean response time of 180ms is compatible with 5% of requests taking four seconds, and it is the 5% who write in. The average is the one…

  • MEMORY USAGE tells you which key is the problem

    Note1 min read

    –bigkeys reports the largest key by element count, which is not the same as the largest by memory — a hash with 500 short fields is smaller than…

  • Blackfire compares two profiles, which is the useful part

    Note2 min read

    A single profile tells you where the time goes and leaves the question of whether that is unusual. A comparison between two profiles answers the question you actually…

  • OPcache is per-process-pool, not per-server

    Note1 min read

    Two PHP-FPM pools have two separate OPcache shared memory segments, so a cache reset in one does nothing to the other — and the site serves a mix…

  • Compose file version 3, and what swarm changed

    Deep dive10 min read

    v3 exists to be deployable to a scheduler, and that removes things a single-host file relied on. What went, why it had to, and what replaces it.

  • Code splitting with dynamic import

    Note1 min read

    Shipping the checkout code to everyone who visits the homepage is the default, because a single entry point produces a single bundle. A dynamic import is a split…

  • Prefetch decides whether one slow consumer blocks the rest

    Note1 min read

    RabbitMQ pushes messages to a consumer up to the prefetch limit without waiting for acknowledgements. With the default of unlimited, one consumer takes the whole queue into its…

  • A cache stampede on a single hot key

    Note1 min read

    A cached value expiring under load means every request that arrives in the next few hundred milliseconds recomputes it simultaneously — so the moment of highest cost is…

  • Centralised logging that answers a question

    Deep dive17 min read

    Four servers, eleven log files, and an incident reconstructed from three of them. Shipping first and parsing second, structured events, and the correlation id that makes it useful.