Eight tracks

Each one is prose with a live database beside it. Every claim a lesson makes is a statement you can run against a forty-one-million-row table with a workload going through it.

• Locks

  0/2

  What is waiting behind your migration?

  The eight lock levels and what conflicts with what, the queue they wait in, and head-of-line blocking watched as it happens. Ends with lock_timeout, which belongs at the top of every migration you will ever write.

  • Eight levels, and what conflicts with what9 min
  • Head-of-line blocking10 min

• Rewrite or metadata

  0/2

  Is this statement copying every row?

  Constant defaults against volatile ones, type changes, NOT NULL and its two-step form, and what a full-table rewrite does to replication lag. The single most valuable distinction in the whole subject.

  • The default that rewrites, and the one everybody warns you about11 min
  • Which type changes are free10 min

• Indexes

  0/1

  How do you build one without stopping writes?

  CONCURRENTLY and what it costs, invalid indexes and how to clean them up, unique indexes on live data, and dropping the index nobody uses.

  • Building an index without stopping writes9 min

• Constraints

  0/1

  How do you add a rule to a table already full of rows?

  NOT VALID then VALIDATE, foreign keys on large tables and the lock they take on both sides, and check constraints as the cheap alternative to a type change.

  • NOT VALID, then VALIDATE10 min

• Expand and contract

  0/1

  How do you change a column the running code still reads?

  Adding beside, dual writes, backfill, switch, drop — and why a rename is five migrations rather than one. Views as a shim when the old name has to keep working.

  • Why a rename is five migrations11 min

• Backfills

  0/1

  How do you touch ten million rows without anybody noticing?

  Batch size and what decides it, sleeping between batches, writing a backfill you can safely run twice, and watching p99 while it runs.

  • Ten million rows, and nobody notices10 min

• Schemaless drift

  0/1

  Where does the schema go when there isn't one?

  Versioned documents, keeping a reader for every version still in the store, backfill as a batch job, and deciding when a version is finally safe to stop supporting.

  • Where the schema goes when there isn't one11 min

• Reading a generated migration

  Lands in a later milestone

  What will this do to production?

  The track the rest exists for. You are handed the migration file an agent wrote for a stated change and a production-sized table, and asked what it locks, for how long, whether it rewrites, and what the old code does mid-migration. Then you run it and watch.

  It lives at /audit

9 lessons built so far. The rest land with their milestones.