We moved connection setup out of the request path and start compute on the first DNS lookup. What worked, what didn't, and the graphs from the rollout.
Copy-on-write storage lets every pull request get its own database in about a second. A tour of the page server behind it.
Add a replica close to your users from the dashboard or the API. Replication lag stays under 100 ms in our tests.
Expand, backfill, contract. The three-step pattern we use for every schema change, with the SQL for each step.
New projects start on 18 today. Existing projects can upgrade in place with a branch-first dry run.
Default settings fall behind once a table takes thousands of updates a second. The four parameters we change first.
Paid take-home projects, written interviews and why we stopped asking people to whiteboard B-trees.