Writing

Notes on building systems that have to stay up. Plain language, real sources, and the numbers that go with them.

Written by Rahul Kumar Choudhary. Each piece comes out of a system that had to keep running, and the figures on every card are the ones the article has to defend.

RSS feed
12 min readAI AgentsDistributed SystemsArchitecture

The agent pauses. The human approves. The approval lands on a different machine.

Making an agent stop and ask a human is four lines of code, and every framework shows you how. None of them mention that the approval arrives as a brand new request, so in production it reaches a machine that has never heard of the paused job. It fails by quietly starting over, not by crashing.

ways a paused agent breaks in production
3
ways a paused agent breaks in production
of them raise an error
0
of them raise an error
emails one approval can send
2
emails one approval can send
  • 8 min readArchitectureTestingEngineering Practice

    Docs are a wish. Review does not scale. So the rules run.

    Architectural decisions die two ways: people leave, and people tidy up. Documentation survives neither. So I stopped writing the decisions down and started writing them as tests that fail. There are 855 of them, including one that caught a screen describing a connection flow that a good share of the services it connects to do not use.

  • 8 min readDatabasesDistributed SystemsArchitecture

    Amazon left relational. Google came back. Both were right.

    Most of what you believe about SQL and NoSQL came from blog posts written around 2012. They were right then. They are not right now, and the two companies everyone quotes as proof both wrote down what actually happened.

One long piece, when there is one worth sending

Agents and distributed systems, argued from a codebase rather than from opinion. It goes out when a piece is finished and not on a schedule, so it will not fill your inbox.

Free. Unsubscribe whenever. Hosted on Substack.

Writing — Rahul Kumar Choudhary