persona-pass 2026-11-12: PgBouncer in transaction mode: clients intermittently get 'prepared statement does not exist' e

Not specified by the original author.

TL;DR: Add ?pgbouncer=true to your DATABASE_URL. PgBouncer in transaction mode cannot do prepared statements, and that one parameter tells Prisma to stop using them. Keep a separate direct connection (no pooler) for migrations.

Error: prepared statement "s0" already exists

Steps

1. Confirm you are behind PgBouncer in transaction mode: your connection string points at a pooler port (for example Supabase port 6543) or your provider says 'transaction mode'. Expected: the error appears on queries, with Postgres code 42P05 in the logs.

2. Add the Prisma PgBouncer flag to the app's connection string:

   postgresql://YOUR_HOST:6543/postgres?pgbouncer=true

   Keep your real username, password, and host; only the parameter is new. Expected: Prisma logs no longer prepare statements on this datasource.

3. Keep migrations on a direct connection. Set a separate variable for them:

   DIRECT_URL="postgresql://YOUR_HOST:5432/postgres"

   Expected: prisma migrate deploy runs against the direct port, the app runs against the pooler.

4. Re-run the failing queries. Expected: the 'already exists' errors stop.

When to use

• The error is prepared statement ... already exists with code 42P05 and you connect through PgBouncer / Supavisor / a pooler
• Supabase, Heroku, Railway, or any hosted Postgres with a transaction-mode pooler

When not to use

• Direct connections to Postgres: prepared statements work fine there, so this error means something else
• P2024 pool timeouts: that is pool exhaustion, not prepared statements

Compatibility

Prisma Client 4.x through 6.x. Any PgBouncer deployment in transaction pooling mode.

Variant phrasing: prepared statement "$z" already exists

Same error, different generated statement name. The fix is identical.

Variant phrasing: error only in production, not locally

Local dev usually connects directly (no pooler), so the bug only appears where the pooler sits. That mismatch is expected.

Why it happens

Prisma names its prepared statements per connection. PgBouncer in transaction mode reuses server connections across clients, so a statement name prepared by one client collides with the same name from another, and Postgres rejects the duplicate with 42P05. The pgbouncer=true flag makes Prisma use simple query protocol instead, which needs no prepared statements.

Edge cases

• Session-mode poolers (Supabase port 5432) support prepared statements; you only need the flag for transaction mode.
• The flag must be on the URL the Prisma Client uses, not just the migration URL.
• Some providers need &statement_cache_size=0 style extras on other drivers; for Prisma the single flag is the documented fix.

Supabase documents this exact failure: connecting Prisma through PgBouncer produces 'prepared statement XXX already exists' because PgBouncer in transaction mode does not support prepared statements. Their documented fix is the Prisma-specific query parameter: append pgbouncer=true to the connection string, which makes Prisma disable prepared statements for that datasource. Teams that applied the parameter report the 42P05 errors stop immediately. Source: https://supabase.com/docs/guides/troubleshooting/error-prepared-statement-xxx-already-exists-3laqeM

Not specified by the original author.

TL;DR: # psycopg + pooled Neon: handle prepared statements ## The trap psycopg3 (and asyncpg-style drivers) use server-side prepared statements for repeated queries. Against Neon's pooler in transaction mode, the prepared statement lives on a Postgres backend that PgBouncer hands to a different client next transaction, so you get "prepared statement does not exist" style failures under concurrency, passing in single-threaded tests. ## The rule Two supported options, from Neon's pooling semantics: 1.

Fix

1. psycopg + pooled Neon: handle prepared statements ## The trap psycopg3 (and asyncpg-style drivers) use server-side prepared statements for repeated queries.

   Expected: this specific failure stops.

2. Use the direct connection string for anything that relies on prepared statements. Expected: the request routes instead of erroring.

3. Or disable client-side preparation (psycopg3:. Expected: this specific failure stops.

Details

Against Neon's pooler in transaction mode, the prepared statement lives on a Postgres backend that PgBouncer hands to a different client next transaction, so you get "prepared statement does not exist" style failures under concurrency, passing in single-threaded tests. ## The rule Two supported options, from Neon's pooling semantics: 1. 2. prepare_threshold=0 / prepare_threshold=None on the connection) so every query goes through as a simple query against the pooled endpoint. ## Checklist - Intermittent prepared-statement errors only under load + only on the pooled string: this is it. - SQLAlchemy users: this is separate from pool_pre_ping ; you can need both. - PREPARE / DEALLOCATE in raw SQL never works on pooled; that is documented, not a bug.

When to use

You hit exactly this: psycopg + Neon pooler: disable prepared statements or go direct in psycopg + Neon pooler.

When not to use

A different error, or the same symptom in a different tool. This page only covers the failure above.

Compatibility

psycopg + Neon pooler.

Root cause

Session state survives because there is no pooler in the middle.

Neon's pooling docs (verified live 2026-09-26): PREPARE/DEALLOCATE (SQL-level prepared statements) are in the documented unsupported list for pooled connections in transaction mode; direct connections have no pooler in the middle.

Not specified by the original author.

TL;DR: Features that assume "my session" break in ways that look like app bugs: search_path resets, LISTEN never fires, temp tables vanish. ## The rule: the unsupported list (from Neon's docs) Not supported with pooled connections: - SET / RESET (session variables) - LISTEN / NOTIFY - WITH HOLD cursors - PREPARE / DEALLOCATE - Temporary tables with ON COMMIT PRESERVE ROWS / DELETE ROWS quirks - LOAD - Session-level advisory locks ## Limits to know -

Fix

1. Features that assume "my session" break in ways that look like app bugs: search_path resets, LISTEN never fires, temp tables vanish. ## The rule: the unsupported list (from Neon's docs) Not supported with pooled connections: - SET / RESET (session variables) - LISTEN / NOTIFY - WITH HOLD cursors - PREPARE / DEALLOCATE - Temporary tables with ON COMMIT PRESERVE ROWS / DELETE ROWS quirks - LOAD - Session-level advisory locks ## Limits to know -. Expected: this specific failure stops.

2. : direct-to-Postgres limit by compute size; beyond it "too many clients already". ## Checklist - Hitting any item on the unsupported list: switch that workload to the direct connection string. - Hitting the queue: raise compute size (raises max_connections, hence pool size) or cut per-instance pool sizes. - Hitting 10,000 clients: you need read replicas or fewer, better-pooled app instances, not a bigger pooler setting. Expected: the request routes instead of erroring.

Details

pool_mode=transaction max_client_conn : 10,000 client connections to PgBouncer; beyond that clients get "no more connections allowed". - default_pool_size : 90% of max_connections , per user per database; beyond that clients queue with a 2-minute timeout. - max_connections

When to use

You hit exactly this: Neon PgBouncer runs transaction mode: know the unsupported list in Neon PgBouncer runs transaction mode.

When not to use

A different error, or the same symptom in a different tool. This page only covers the failure above.

Compatibility

Neon PgBouncer runs transaction mode.

Root cause

means the server-side connection is reassigned after every transaction.

Edge cases

• # The pooler's transaction mode and what it forbids ## The trap.

Neon's connection pooling docs (verified live 2026-09-26): PgBouncer in transaction mode; the full unsupported-features list; the three limit tiers (maxclientconn 10,000, defaultpoolsize 90% of maxconnections, maxconnections by compute size).