drizzle-orm

Fixes #6241 What changed in the , and sessions ran before the that releases the pool client. When threw (statement timeout, connection dropped right after connect, admin shutdown), never ran and the leaked — one client per failure until hung on an empty pool. This moves the statement inside the in all three sessions. That's the whole code change (3 lines). The existing still runs and rethrows. Outside a transaction is a WARNING, not an error. The error the caller gets is unchanged ( wrapping the driver error). Non-pooled clients never went through , so nothing changes for them. and don't acquire a pooled client and aren't touched. Tests — vitest, no database or Docker. It subclasses each driver's , overrides to return a fake client whose throws , and runs two cases per driver via : 1. fails → error propagates with , the callback never runs, the client is acquired once and released once. 2. Happy path → , , released once. Guards the normal flow. On case 1 fails for all three drivers ( is 0). With this change all 6 pass. is clean and has no errors. Changelog One line under a new "Bug fixes" heading in (the unreleased version on ). No version bump. Notes (1.0.0-rc.4) has the same three lines; happy to open the same change there if you want it ported. Found this while moving a service from raw to drizzle — the raw code had around , so every in the migration silently picked up the leak. This was created with the help of AI (Claude), reviewed and tested by me.

Fixes #6214. was mangling any check constraint defined with NOT VALID. The introspector assumed the definition always looked like and pulled out the expression with a fixed slice, which works fine normally but chops the last letter off "NOT VALID" when it's there, e.g. CHECK ((version >= 0)) NOT VALID -> sql which is just invalid SQL once it lands in the generated migration. My first pass just stripped the NOT VALID suffix and kept the constraint, but I switched to excluding it entirely instead. There's nowhere in drizzle-kit that tracks whether a check constraint is validated, so keeping a "fixed" version around means a future push could quietly re-create it as validated - that's a real behavior change nobody asked for, on top of just fixing corrupted SQL. I noticed there's already a test for this exact issue in the repo ( in pull.test.ts, date-gated to turn on today) that expects exactly this - NOT VALID checks excluded from introspection. This PR is written to make that pass. Couldn't run it end to end locally though - it goes through a tsc check that needs the built drizzle-orm package, and building drizzle-orm on rc5 right now hits an unrelated rolldown error (MISSING_EXPORT on sql/expressions/index.ts) even on a clean checkout with nothing changed. I checked the actual behavior works directly against a real Postgres instance instead - a NOT VALID constraint gets excluded, a normal one on the same table comes through untouched.

drizzle-orm: 0.45.2 — same code on (0.45.3) and (1.0.0-rc.4) drizzle-kit: 0.31.10 Other packages:** pg 8.22.0 Describe the Bug If runs on a and the statement fails, the client taken from the pool is never released. is executed before the that calls : does fail in practice: (), a connection dropped right after connect (, "Connection terminated unexpectedly"), admin shutdown. Each time, the pool loses one client for good. Under load eventually hangs on an empty pool. Repro without a database — subclass and make throw: Expected: the client is released on every exit path, including a failing . Moving inside the is enough: the existing runs (only a WARNING when no transaction is open) and releases. The error the caller sees stays the same . I have the fix plus a driver-mocked test for the three sessions; PR follows.