Re-using the features from execaNode()

[ehmicky]

execaNode() has some nice features:

• prepend 'node' as the binary argument
• re-use the current Node.js binary, including its version
• re-use the current Node.js CLI flags, without getting any issues when using --inspect
• forces shell: false

However, those features are not available for users who either:

• use $ or execaCommand()
• use synchronous methods like execaSync()
• do not wish to add an ipc channel

We could fix this by adding a { node: true } option (any other option name works too), which would do the same thing as what execaNode() is doing, except for adding an ipc channel.

What do you think?

[sindresorhus]

Why would you not want a ipc channel though? If there's no good reason, then having {node: true} be the same as execaNode() would make the docs simpler.

[ehmicky]

The only reason I was thinking of was to avoid the cost of setting up the IPC channel.

A named pipe is created (or is a TCP Unix socket used? I am not sure from looking at the libuv code). I am not sure whether this counts towards ulimit -n. 🤔

Performance-wise, the child process does some additional setup when it starts. I tried benchmarking it with the following code.

import {spawn} from 'node:child_process'
import {once} from 'node:events'

console.time()
for (let i = 0; i < 1e3; i += 1) {
  const childProcess = spawn('node', ['empty.js'], {stdio: ['pipe', 'pipe', 'pipe', 'ipc']})
  await once(childProcess, 'exit')
}
console.timeEnd()

console.time()
for (let i = 0; i < 1e3; i += 1) {
  const childProcess = spawn('node', ['empty.js'], {stdio: ['pipe', 'pipe', 'pipe']})
  await once(childProcess, 'exit')
}
console.timeEnd()

Which gives me (Node 21.1.0, Ubuntu):

default: 18.379s
default: 15.602s

I am not sure though whether the reasons above are enough to justify creating an ipc channel by default or not. 🤔 How do you feel about this?

[sindresorhus]

People can just do {node: true, ipc: false} if they don't want it, right? If so, I think the consistency is worth the tiny performance overhead by default.

[ehmicky]

Yes, that's a good point. Also, the benchmark is on an empty file. The time to actually load the file should be much higher, making the time spent setting up the ipc channel not as significant.

So, what are you suggesting is the following, is this correct?

• ipc option defaults to true if {node: true} is used
• execaNode(...) is the same as execa(..., {node: true})

[sindresorhus]

Yes

And make sure there is a test for {node: true, ipc: false}.