Deterministic tests

Each test starts from the schema and nothing else, so it can be run alone, in any order, and in parallel. A few habits keep it that way:

• Defaults live in the schema (DEFAULT now(), gen_random_uuid(), sequences, constraints), so every clone has them. A default set by application code is one the tests have to remember to set.
• Don't assert values the database makes up (now(), generated ids) unless the test passes them in. Assert that an id is a UUID, or that a time is after the test started.
• The server's configuration doesn't leak in. Every database gets the same settings, including while the template is built: by default UTC, ISO dates, the C locale for money, numbers and times, and the simple text search configuration, rather than whatever initdb took from the machine. Set others, such as search_path, there too.
• Data a test needs is inserted by the test, or by a fixture builder next to it, after cloning, so what it starts from is visible where it matters. Only data nearly every test needs belongs in a seed.
• Keep pools small (a few connections): every test of every parallel run shares the server's max_connections.
• Close connections before close(). Dropping a database ends any connection left open, so a leak can't hang the run, but the client then sees an error.

Debugging a failed test

Keep a failed test's database with keep(). Its name is printed, and it is left on the server for retainMs (a day by default):

psql "postgres://postgres:postgres@localhost:5432/pg_dolly_…"

A later run drops it once retainMs has passed. To clear everything sooner, drop the databases starting with the prefix, or reset a server used only for tests.