Search code examples
postgresqlaws-lambdaserverless

How should I manage postgres database handles in a serverless environment?

I have an API running in AWS Lambda and AWS Gateway using Up. My API creates a database connection on startup, and therefore Lambda does this when the function is triggered for the first time. My API is written in node using Express and pg-promise to connect to and query the database.

The problem is that Lambda creates new instances of the function as it sees fit, and sometimes it appears as though there are multiple instances of it at one time.

I keep running out of DB connections as my Lambda function is using up too many database handles. If I log into Postgres and look at the pg_stat_activity table I can see lots of connections to the database.

What is the recommended pattern for solving this issue? Can one limit the number of simultaneous instances of a function in Lambda? Can you share a connection pool across instances of a function (I doubt it).

UPDATE

AWS now provides a product called RDS Proxy which is a managed connection pooling solution to solve this very issue: https://aws.amazon.com/blogs/compute/using-amazon-rds-proxy-with-aws-lambda/

Solution

  • There a couple ways that you can run out of database connections:

    1. You have more concurrent Lambda executions than you have available database connections. This is certainly possible.
    2. Your Lambda function is opening database connections but not closing them. This is a likely culprit, since web frameworks tend to keep database connections open across requests (which is more efficient), but on Lambda have no opportunity to close them since AWS will silently terminate the instance.

    You can solve 1 by controlling the number of available connections on the database server (the max_connections setting on PostgreSQL) and the maximum number of concurrent Lambda function invocations (as documented here). Of course, that just trades one problem for another, since Lambda will return 429 errors when it hits the limit.

    Addressing 2 is more tricky. The traditional and right way of dealing with database connection exhaustion is to use connection pooling. But with Lambda you can't do that on the client, and with bare RDS you don't have the option to do that on the server. The good news is that AWS has started offering their own RDS connection pooling solution. If you don't want to use that you could set up your own intermediary persistent connection pooler, but that makes for a more complicated setup.

    In the absence of pooling, one option is to create and destroy a database connection on each function invocation. Unfortunately that will add quite a bit of overhead and latency to your requests.

    Another option is to carefully control your client-side and server-side connection parameters. The idea is first to have the database close connections after a relatively short idle time (on PostgreSQL this is controlled by the tcp_keepalives_* settings). Then, to make sure that the client never tries to use a closed connection, you set a connection timeout on the client (how to do so will be framework dependent) that is shorter than that value.

    My hope is that AWS will give us a solution for this at some point (such as server-side RDS connection pooling). You can see various proposed solutions in this AWS forum thread.