Every index is a write tax you pay forever in exchange for reads you may or may not run. I keep meeting teams who have ten indices on a table and use three.

If the service can only be operated by the person who built it, you have not finished building it. The runbook is the part where engineering becomes operations.

A deploy that takes six hours to roll back is not a deploy, it is a renovation. Optimize for the size of the change you can revert in one click.

Every log line is a tiny API your future self will depend on. Treat them like a public interface — versioned, structured, ruthlessly pruned.

I keep watching teams rip out a working REST API for a GraphQL one they will replace with tRPC inside a year. The shape of your network calls is rarely the bottleneck.

Most operational outages I have been part of were ultimately fixed by two engineers and a shared shell. The tooling we wrap around that fact matters less than we admit.

The best reviews I have received did not just look at the diff — they looked at why the file existed at all, and gently asked whether it still should.

Every flag is a promise to delete a branch of your code later. Treat the flag's ticket and its deletion ticket as the same ticket.

Postgres, a queue, and a script that runs every minute will get you further than you think before the architecture diagram needs to grow.

If a change spans three repos and you cannot land it in one PR, you do not have services. You have a distributed monolith and worse tooling.

You can afford to log everything until you cannot. Decide the budget before the bill arrives. Sampling is not a failure, it is the answer.

Half the bugs I fix get fixed in the paragraph where I describe the bug to someone else. The other half get fixed because I had to type the description.