Türker Yıldırım

system-design

  • Backpressure is a design decision, not an accident

    Note2 min read

    A producer faster than its consumer fills a queue, and what happens when the queue is full is a decision somebody makes — or a decision the infrastructure…

  • Idempotency for an endpoint that takes money

    Deep dive14 min read

    The client timed out, retried, and the customer was charged twice. A timeout tells the client nothing about what the server did.

  • The dual write problem, and why the outbox exists

    Note2 min read

    Writing a row and publishing an event are two systems, and there is no ordering of the two that is safe. Neither ordering is correct and both are…

  • A read replica that was silently stale

    Deep dive15 min read

    The replica was up, accepting connections and serving data from an hour ago. Seconds_Behind_Master reported zero, honestly, and was measuring the wrong thing.

  • An SLO is a number somebody agreed to

    Deep dive14 min read

    Everyone wanted the site to be fast and nobody could say what fast meant. An indicator, an objective and a window — and the error budget is what…

  • Read-your-own-writes is a product requirement, not a bug

    Note1 min read

    A user who submits a form and immediately sees stale data reports it as a bug, and it is a consequence of a decision somebody made about replication.…

  • The error budget is what the objective is for

    Note2 min read

    An objective of 99.5% is also a statement that 0.5% of requests may fail, and that allowance is a resource with a use. The budget turns reliability from…

  • A stampede lock, and what the losers should do

    Note2 min read

    An expensive cached value expires under load and every concurrent request misses simultaneously, so forty workers run the same four-second query at once. What the losers do is…

  • Timeouts bound one request; only a cap bounds concurrency

    Note1 min read

    Setting a thirty-second timeout feels responsible and does nothing about the number of workers waiting simultaneously, which is what actually takes a site down. Dropping the timeout from…

  • Redis 6 and the shared password that was everywhere

    Deep dive10 min read

    One password, six services, and FLUSHALL available to all of them. 6.0 landed in April with users — and the keyspace pattern is where the isolation actually is.