0
Challenge · 59RLS at Scale
RLS performance: (select auth.uid()) & indexes
Perfused in ~85% of apps
Every policy you've written so far has been correct. None of them has had to survive a table with two million rows. A policy is invisible in your code, which is exactly what makes it easy to forget it isn't invisible to the query planner — it runs on every row a query touches, and the three ways to make that expensive (a function call that can't be cached, a predicate column with no index, a subquery shaped the wrong way) are all things you can spot, and fix, in under a minute with EXPLAIN.
Core idea. RLS performance comes down to three concrete, checkable decisions. First: wrap `auth.*()` calls in `(select ...)` so the planner can cache them as an InitPlan — but only calls that don't vary per row benefit; a helper function called with a row's own column, like `is_org_member(org_id)`, is still re-run once per row no matter how you wrap it, because the wrapping doesn't remove the correlation. Second: index every column a policy predicate touches, exactly as you would for a WHERE clause. Third: prefer a subquery shape where the caller's own data ("which orgs am I in") is computed ONCE and matched against every row, over a shape that asks "is this specific row's org mine" once per row.
poke it, drag it,
try to break it
try to break it
no slides here.
you drive.
you drive.
Playground, this isn't a screenshot
Live demointeractive
Real Postgres · in your browseridle
team_posts has a REAL policy installed — using ((select is_org_member(org_id))) — on 3,000 rows across 5 orgs. Run this as alice (impersonated) and note the total Execution Time and the loops count next to SubPlan. Later in this concept you'll run the identical query against a rewritten policy and compare.Querying as
Cmd/Ctrl + Enter
↑ play a little first, then dig in