Section · 8 articles
Systems & Data
Distributed systems, storage, consistency and the failures that follow from an unreliable network.
For a separate people-operations application of the same measurement discipline, see remote workforce management software.
Transaction Isolation Levels: What Each One Actually Guarantees
The ANSI levels describe anomalies, not mechanisms, and no major database implements them the way the standard reads. What you actually get, per engine.
02Every Request Will Be Retried: Idempotency in Practice
You cannot tell a lost request from a lost response, so duplicates are structural rather than a bug. What idempotency actually requires to implement.
03Consistency in Replicated Stores: Reading the Guarantees
Vendors use the same words for different guarantees. What each model actually promises, which anomalies survive, and the questions that get a straight answer.
04Clocks in Distributed Systems: Why You Cannot Trust Timestamps
Wall clocks jump backwards, drift, and disagree between machines. What breaks when you order events by timestamp, and what to use instead.
05Two-Phase Commit and Why Almost Nobody Uses It
It gives you atomicity across systems and takes your availability in exchange. Why the blocking case is fatal, and what people use instead.
06Schemaless Storage Still Has a Schema
Removing the schema from the database does not remove it — it moves it into application code, where nothing checks it and every version disagrees.
07Indexes: What Happens on Write
Every index makes reads faster and writes slower. What a write actually does, and how to find the indexes you are paying for and not using.
08Exactly-Once Delivery Does Not Exist
The guarantee vendors advertise is exactly-once processing under specific conditions, which is a different and narrower thing. What is actually available.
For primary background on this topic, consult RFC Editor.