intermittent: write EPIPE + Cannot read property 'onError' of undefined

Bug description

These two exceptions happen in the same flow:

ERROR	Invoke Error 	{"errorType":"Error","errorMessage":"write EPIPE","code":"EPIPE","clientVersion":"2.12.1","stack":["Error: write EPIPE","    at PrismaClientFetcher.request (/var/task/dist/webpack:/mnt/ramdisk/project/node_modules/@prisma/client/runtime/index.js:79355:15)","    at processTicksAndRejections (internal/process/task_queues.js:97:5)"]}

2 milliseconds later:

ERROR	Uncaught Exception 	{"errorType":"TypeError","errorMessage":"Cannot read property 'onError' of undefined","stack":["TypeError: Cannot read property 'onError' of undefined","    at Socket.ce (/var/task/dist/webpack:/mnt/ramdisk/project/node_modules/@prisma/client/runtime/index.js:25673:45)","    at Socket.emit (events.js:314:20)","    at Socket.EventEmitter.emit (domain.js:483:12)","    at Pipe.<anonymous> (net.js:676:12)"]}

This was previously masked by #4117, which was fixed by this commit; now the real error is visible.

Possibly related to #4400

How to reproduce

This is an intermittent error on AWS Lambda, could not reproduce it on a dev machine.

Expected behavior

• No exception (duh)
• It would be nice to have source mapping to link the stack trace into a specific spot in the code (instead of @prisma/client/runtime/index.js:25673:45).
• It would be nice to have more logs from the Rust process; IIUC there’s no way to set RUST_LOG to something other than info

Environment & setup

• OS: AWS Lambda (linux-musl), node 12.x
• Database: PostgreSQL 12 on RDS
• Node.js version: 12.18
• Prisma version: 2.12.1

[pantharshit00]

I have a theory that lambda is cleaning up the /tmp directory and cleaning up our unix domain socket which we use now in the latest version(we were using a TCP port before for Rust <-> communication).

Can you all try installing the dev channel and turning off UDS and see if this shows up again:

const prisma = new PrismaClient({
  __internal:{
    useUds: false
  }
} as any) // remove as any if you using js

Similar bug report about uds: https://github.com/prisma/prisma/issues/4744

[royra]

I have a theory that lambda is cleaning up the /tmp directory and cleaning up our unix domain socket which we use now in the latest version(we were using a TCP port before for Rust <-> communication).

Looks like prisma itself is deleting its socket file on the beforeExit event of the node process. According to this blog post:

The Lambda runtime uses the beforeExit event to complete the last request and start processing the next request.

If this is true, it explains the intermittent error:

1. The event loop for the lambda container is emptied and beforeHook is called
2. Prisma deletes its own socket file
3. More requests are queued to the event loop, causing the same prisma instance to attempt more work
4. EPIPE

Reverting to tcp socket is a workaround, but the real fix could be unhooking from the beforeExit event.