Skip to content

Latest commit

 

History

History
59 lines (40 loc) · 3.85 KB

README.md

File metadata and controls

59 lines (40 loc) · 3.85 KB

WALRUS regression suite

pg_regress tests inherited from supabase/walrus.

They cover the SQL side of Postgres Changes end to end: a real logical replication slot, wal2json output, realtime.apply_rls, realtime.subscription_check_filters, realtime.is_visible_through_filters, RLS, column permissions, selected_columns, action_filter and the filter operators.

What they run against

The tenant realtime schema, built the way a fresh tenant gets it. Nothing here builds the schema independently, so a change to the tenant migrations that breaks Postgres Changes shows up as a diff in expected/.

Realtime.Tenants.Migrations builds it in one of two ways, and the suite runs in either:

Mode Schema comes from Which tenants get it
--dump priv/repo/tenant_db_dump_<major>.sql every tenant on a supported major (the default, and the common case)
--migrations replaying every migration in that module orioledb tenants, where the dump is refused, and any tenant whose dump won't load

Both modes share one expected/: the two schemas are meant to be the same schema, so a mode that diffs is itself the finding. --migrations runs the migrations through dev/scripts/migrate_tenant_db.exs, so it needs Elixir on the host; --dump needs only docker.

setup.sql adds the two things the supabase/postgres image doesn't already provide: a for all tables publication, and the auth.uid() a real project has (the image ships an older one that can't see the claims apply_rls sets. See the comment in that file). fixtures.sql defines the walrus / polling_query views and the norm(), clear_wal(), seed_uuid() helpers the tests are written against.

Running

mise run walrus                     # whole suite, from the dump
mise run walrus --migrations        # whole suite, from the migrations
mise run walrus test_simple_insert  # one or more tests
mise run walrus --major 15          # a specific Postgres major

The image is resolved from TENANT_DUMP_IMAGES in mise.toml. With no --major the newest entry in that list wins.

POSTGRES_IMAGE overrides the lot, which is how you reach an image with no committed dump:

POSTGRES_IMAGE=supabase/postgres:17.11.0.019-orioledb mise run walrus --migrations

test/walrus/run.sh takes the same arguments and works on its own; the only thing the mise task adds is resolving POSTGRES_IMAGE, which the script otherwise leaves to the default in compose.walrus-db.yml.

Every run drops what actually happened in results/ (gitignored). On a failure the diff is printed, and once you've read it and agree the new output is correct:

cp test/walrus/results/<test>.out test/walrus/expected/<test>.out

Layout

Path Purpose
sql/ one file per test, run in alphabetical order in a shared database
expected/ the psql transcript each test must reproduce, byte for byte
setup.sql the publication and auth.uid() the image doesn't provide
fixtures.sql helper views and functions the tests call
run.sh boots the database, builds the schema, drives pg_regress
results/ last run's actual output, gitignored

Each test is responsible for its own cleanup (drop table, pg_drop_replication_slot, truncate realtime.subscription) because they all share one database.