system-design – Page 18 – Türker Yıldırım

A saga is a compensation, not a distributed transaction

NoteMay 25, 20212 min read

A sequence of steps across services cannot be rolled back, so each step has a compensating action that undoes its effect after the fact. The compensations are business…

A Redis Lua script is atomic and blocks everything

NoteMay 25, 20212 min read

A script runs to completion with nothing interleaved, which is exactly why it is useful and exactly why a slow one is an outage. The atomicity removes the…

Compose v2 is a plugin, and the hyphen finally goes

Deep diveMay 19, 20219 min read

The same command, spelled two ways, on eight machines. A Python program became a Go plugin, and the container names changed underneath every script.

systemd Restart=on-failure and the loop it creates

NoteMay 17, 20212 min read

A unit that restarts on failure and fails immediately will restart as fast as systemd allows, until the rate limit stops it — and the rate limit is…

The service that should not have been a service

Deep diveMay 14, 202110 min read

Four services, three of which deploy together and share a database. The costs of distribution with none of the independence.

Kafka partitions are the parallelism limit

NoteMay 13, 20212 min read

A consumer group can have at most one consumer per partition, so the partition count is the maximum useful number of consumers. The inability to reduce the count…

Alert on the symptom, never on the cause

NoteMay 12, 20212 min read

An alert on a cause fires when nothing is wrong and stays quiet when something is, because the mapping from causes to consequences is not one to one.…

A recursive CTE that carries its own path to detect a cycle

NoteMay 11, 20212 min read

A depth limit stops a recursive query running forever and does not tell you the data has a cycle in it, which is a different problem with a…

The unique index arbitrates; the SELECT does not

NoteApril 29, 20212 min read

Checking whether a key exists and then inserting it is two statements and a race, and the race is precisely the concurrent-retry case the key exists to handle.…

Backpressure is the queue depth you agreed to refuse at

NoteApril 28, 20212 min read

A queue with no bound converts an overload into a latency problem that appears hours later and cannot be reasoned about. Framing the bound as a staleness budget…