Connection management

Databases can handle a limited number of concurrent connections. Each connection requires RAM, which means that simply increasing the database connection limit without scaling available resources:

  • โœ” might allow more processes to connect but
  • โœ˜ significantly affects database performance, and can result in the database being shut down due to an out of memory error

The way your application manages connections also impacts performance. This guide describes how to approach connection management in serverless environments and long-running processes.

This guide focuses on relational databases and how to configure and tune the Prisma connection pool (MongoDB uses the MongoDB driver connection pool).

Long-running processes

Examples of long-running processes include Node.js applications hosted on a service like Heroku or a virtual machine. Use the following checklist as a guide to connection management in long-running environments:

The recommended connection pool size (connection_limit) to start with for long-running processes is the default pool size (num_physical_cpus * 2 + 1) รท number of application instances.

If you have one application instances:

  • The default pool size applies by default (num_physical_cpus * 2 + 1) - you do not need to set the connection_limit parameter.
  • You can optionally tune the pool size.

If you have multiple application instances:

PrismaClient in long-running applications

In long-running applications, we recommend that you:

Re-using a single PrismaClient instance

To re-use a single instance, create a module that exports a PrismaClient object:

client.ts
1import { PrismaClient } from '@prisma/client'
2
3let prisma = new PrismaClient()
4
5export default prisma

The object is cached the first time the module is imported. Subsequent requests return the cached object rather than creating a new PrismaClient:

app.ts
1import prisma from './client'
2
3async function main() {
4 const allUsers = await prisma.user.findMany()
5}
6
7main()

You do not have to replicate the example above exactly - the goal is to make sure PrismaClient is cached. For example, you can instantiate PrismaClient in the context object that you pass into an Express app.

Do not explicitly $disconnect()

You do not need to explicitly $disconnect() in the context of a long-running application that is continuously serving requests. Opening a new connection takes time and can slow down your application if you disconnect after each query.

Prevent hot reloading from creating new instances of PrismaClient

Frameworks like Next.js support hot reloading of changed files, which enables you to see changes to your application without restarting. However, if the framework refreshes the module responsible for exporting PrismaClient, this can result in additional, unwanted instances of PrismaClient in a development environment.

As a workaround, you can store PrismaClient as a global variable in development environments only, as global variables are not reloaded:

client.ts
1import { PrismaClient } from '@prisma/client'
2
3// add prisma to the NodeJS global type
4interface CustomNodeJsGlobal extends NodeJS.Global {
5 prisma: PrismaClient
6}
7
8// Prevent multiple instances of Prisma Client in development
9declare const global: CustomNodeJsGlobal
10
11const prisma = global.prisma || new PrismaClient()
12
13if (process.env.NODE_ENV === 'development') global.prisma = prisma
14
15export default prisma

The way that you import and use the client does not change:

app.ts
1import prisma from './client'
2
3async function main() {
4 const allUsers = await prisma.user.findMany()
5}
6
7main()

Serverless environments (FaaS)

Examples of serverless environments include Node.js functions hosted on AWS Lambda, Vercel or Netlify Functions. Use the following checklist as a guide to connection management in serverless environments:

The serverless challenge

In a serverless environment, each function creates its own instance of PrismaClient, and each client instance has its own connection pool.

Consider the following example, where a single AWS Lambda function uses PrismaClient to connect to a database. The connection_limit is 3:

An AWS Lambda function connecting to a database.

A traffic spike causes AWS Lambda to spawn two additional lambdas to handle the increased load. Each lambda creates an instance of PrismaClient, each with a connection_limit of 3, which results in a maximum of 9 connections to the database:

Three AWS Lambda function connecting to a database.

200 concurrent functions (and therefore 600 possible connections) responding to a traffic spike ๐Ÿ“ˆ can exhaust the database connection limit very quickly. Furthermore, any functions that are paused keep their connections open by default and block them from being used by another function.

  1. Start by setting the connection_limit to 1
  2. If a smaller pool size is not enough, consider using an external connection pooler like PgBouncer

The recommended pool size (connection_limit) in serverless environments depends on:

Without an external connection pooler

If you are not using an external connection pooler, start by setting the pool size (connection_limit) to 1, then optimize. Each incoming request starts a short-lived Node.js process, and many concurrent functions with a high connection_limit can quickly exhaust the database connection limit during a traffic spike.

The following example demonstrates how to set the connection_limit to 1 in your connection URL:

PostgreSQL
MySQL
postgresql://USER:PASSWORD@HOST:PORT/DATABASE?schema=public&connection_limit=1

If you are using AWS Lambda and not configuring a connection_limit, refer to the following GitHub issue for information about the expected default pool size: https://github.com/prisma/docs/issues/667

With an external connection pooler

If you are using an external connection pooler, use the default pool size (num_physical_cpus * 2 + 1) as a starting point and then tune the pool size. The external connection pooler should prevent a traffic spike from overwhelming the database.

Optimizing for parallel requests

If you rarely or never exceed the database connection limit with the pool size set to 1, you can further optimize the connection pool size. Consider a function that is sends queries in parallel:

Promise.all() {
query1,
query2,
query3
query4,
...
}

If the connection_limit is 1, this function is forced to send queries serially (one after the other) rather than in parallel. This slows down the function's ability to process requests, and may result in pool timeout errors. Tune the connection_limit parameter until a traffic spike:

  • Does not exhaust the database connection limit
  • Does not result in pool timeout errors

PrismaClient in serverless environments

Instantiate PrismaClient outside the handler

Instantiate PrismaClient outside the scope of the function handler to increase the chances of reuse. As long as the handler remains 'warm' (in use), the connection is potentially reusable:

import { PrismaClient, Prisma } from '@prisma/client'
const client = new PrismaClient()
export async function handler() {
/* ... */
}

Do not explicitly $disconnect()

You do not need to explicitly $disconnect() at the end of a function, as there is a possibility that the container might be reused. Opening a new connection takes time and slows down your function's ability to process requests.

Other serverless considerations

Container reuse

There is no guarnatee that subsequent nearby invocations of a function will hit the same container - for example, AWS can choose to create a new container at any time.

Code should assume the container to be stateless and create a connection only if it does not exist - Prisma Client JS already implements this logic.

Zombie connections

Containers that are marked "to be removed" and are not being reused still keep a connection open and can stay in that state for some time (unknown and not documented from AWS). This can lead to sub-optimal utilization of the database connections.

A potential solution is to clean up idle connections (serverless-mysql implements this idea, but cannot be used with Prisma).

Concurrency limits

Depending on your serverless concurrency limit (the number of serverless functions running in parallel), you might still exhaust your database's connection limit. This can happen when too many functions are invoked concurrently, each with its own connection pool, which eventually exhausts the database connection limit. To prevent this, you can set your serverless concurrency limit to a number lower than the maximum connection limit of your database divided by the number of connections used by each function invocation (as you might want to be able to connect from another client for other purposes).

Optimizing the connection pool

If the query engine cannot process a query in the queue before the time limit , you will see connection pool timeout exceptions in your log. A connection pool timeout can occur if:

  • Many users are accessing your app simultaneously
  • You send a large number of queries in parallel (for example, using await Promise.all())

If you consistently experience connection pool timeouts after configuring the recommended pool size, you can further tune the connection_limit and pool_timeout parameters.

Increasing the pool size

Increasing the pool size allows the query engine to process a larger number of queries in parallel. Be aware that your database must be able to support the increased number of concurrent connections, otherwise you will exhaust the database connection limit.

To increase the pool size, manually set the connection_limit to a higher number:

datasource db {
provider = "postgresql"
url = "postgresql://johndoe:mypassword@localhost:5432/mydb?schema=public&connection_limit=40"
}

Note: Setting the connection_limit to 1 in serverless environments is a recommended starting point, but this value can also be tuned.

Increasing the pool timeout

Increasing the pool timeout gives the query engine more time to process queries in the queue. You might consider this approach in the following scenario:

  • You have already increased the connection_limit.
  • You are confident that the queue will not grow beyond a certain size, otherwise you will eventually run out of RAM.

To increase the pool timeout, set the pool_timeout parameter to a value larger than the default (10 seconds):

datasource db {
provider = "postgresql"
url = "postgresql://johndoe:mypassword@localhost:5432/mydb?connection_limit=5&pool_timeout=20"
}

Disabling the pool timeout

Disabling the pool timeout prevents the query engine from throwing an exception after x seconds of waiting for a connection and allows the queue to build up. You might consider this approach in the following scenario:

  • You are submitting a large number of queries for a limited time - for example, as part of a job to import or update every customer in your database.
  • You have already increased the connection_limit.
  • You are confident that the queue will not grow beyond a certain size, otherwise you will eventually run out of RAM.

To disable the pool timeout, set the pool_timeout parameter to 0:

datasource db {
provider = "postgresql"
url = "postgresql://johndoe:mypassword@localhost:5432/mydb?connection_limit=5&pool_timeout=0"
}

External connection poolers

Connection poolers like PgBouncer prevent your application from exhausting the database's connection limit.

PgBouncer

PostgreSQL only supports a certain amount of concurrent connections, and this limit can be reached quite fast when the service usage goes up โ€“ especially in serverless environments.

PgBouncer holds a connection pool to the database and proxies incoming client connections by sitting between Prisma Client and the database. This reduces the number of processes a database has to handle at any given time. PgBouncer passes on a limited number of connections to the database and queues additional connections for delivery when connections becomes available.

To use PgBouncer, see: Configure Prisma Client with PgBouncer

AWS RDS Proxy

Due to the way AWS RDS Proxy pins connections, it does not provide any connection pooling benefits when used together with Prisma Client.

Edit this page on GitHub