Prisma Connection Pool Exhaustion

Prisma Connection Pool Exhaustion in Serverless

Problem

Serverless functions create a new Prisma client instance on each cold start. Each instance opens multiple database connections (default: 5 per instance). With many concurrent requests, this quickly exhausts the database's connection limit (often 20-100 for managed databases).

Context / Trigger Conditions

This skill applies when you see:

• P2024: Timed out fetching a new connection from the connection pool
• PostgreSQL: FATAL: too many connections for role "username"
• MySQL: Too many connections
• Works fine locally with npm run dev but fails in production
• Errors appear during traffic spikes, then resolve
• Database dashboard shows connections at or near limit

Environment indicators:

• Deploying to Vercel, AWS Lambda, Netlify Functions, or similar
• Using Prisma with PostgreSQL, MySQL, or another connection-based database
• Database is managed (PlanetScale, Supabase, Neon, RDS, etc.)

Solution

Step 1: Use Connection Pooling Service

The recommended solution is to use a connection pooler like PgBouncer or Prisma Accelerate, which sits between your serverless functions and the database.

For Supabase:

# .env
# Use the pooled connection string (port 6543, not 5432)
DATABASE_URL="postgresql://user:pass@db.xxx.supabase.co:6543/postgres?pgbouncer=true"

For Neon:

# .env  
DATABASE_URL="postgresql://user:pass@ep-xxx.us-east-2.aws.neon.tech/dbname?sslmode=require"
# Neon has built-in pooling

For Prisma Accelerate:

npx prisma generate --accelerate

Step 2: Configure Prisma Connection Limits

In your schema.prisma:

datasource db {
  provider = "postgresql"
  url      = env("DATABASE_URL")
  // Limit connections per Prisma instance
  relationMode = "prisma"
}

In your connection URL or Prisma client:

// lib/prisma.ts
import { PrismaClient } from '@prisma/client'

const globalForPrisma = global as unknown as { prisma: PrismaClient }

export const prisma = globalForPrisma.prisma || new PrismaClient({
  datasources: {
    db: {
      url: process.env.DATABASE_URL + '?connection_limit=1'
    }
  }
})

if (process.env.NODE_ENV !== 'production') globalForPrisma.prisma = prisma

Step 3: Singleton Pattern (Development)

Prevent hot-reload from creating new clients:

// lib/prisma.ts
import { PrismaClient } from '@prisma/client'

const globalForPrisma = globalThis as unknown as {
  prisma: PrismaClient | undefined
}

export const prisma = globalForPrisma.prisma ?? new PrismaClient()

if (process.env.NODE_ENV !== 'production') globalForPrisma.prisma = prisma

Step 4: URL Parameters

Add these to your connection string:

?connection_limit=1&pool_timeout=20&connect_timeout=10

• connection_limit=1: One connection per serverless instance
• pool_timeout=20: Wait up to 20s for available connection
• connect_timeout=10: Fail fast if can't connect in 10s

Verification

After applying fixes:

1. Deploy to production
2. Run a load test: npx autocannon -c 100 -d 30 https://your-app.com/api/test
3. Check database dashboard—connections should stay within limits
4. No more P2024 errors in logs

Example

Before (error under load):

[ERROR] PrismaClientKnownRequestError:
Invalid `prisma.user.findMany()` invocation:
Timed out fetching a new connection from the connection pool.

After (with connection pooling):

# Using Supabase pooler URL
DATABASE_URL="postgresql://...@db.xxx.supabase.co:6543/postgres?pgbouncer=true&connection_limit=1"

Database connections stable at 10-15 even under heavy load.

Notes

• Different managed databases have different pooling solutions—check your provider's docs
• PlanetScale (MySQL) uses a different architecture and doesn't have this issue
• connection_limit=1 is aggressive; start there and increase if you see latency
• The singleton pattern only helps in development; in production serverless, each

instance is isolated

• If using Prisma with Next.js API routes, each route invocation may be a separate

serverless function

• Consider Prisma Accelerate for built-in caching + pooling: https://www.prisma.io/accelerate

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: AlexMikhalev
• Source: AlexMikhalev/claude-code-continuous-learning-skill
• License: MIT

Install and usage instructions live in the source repository linked above.