A better workaround for atomic writes

[mitar]

This fixes #742 and fixes #237.

[coveralls]

Coverage increased (+0.4%) to 96.62% when pulling 8d33e01 on mitar:atomic-workaround into c7cf6e9 on paulmillr:master.

[coveralls]

Coverage increased (+0.4%) to 96.62% when pulling 4a3db5a on mitar:atomic-workaround into c7cf6e9 on paulmillr:master.

[SGrondin]

Posting here to confirm that this PR fixed our issues in Parcel.

[paulmillr]

@SGrondin fixed issues in Parcel and broke in Gatsby...

[cmcaine]

@paulmillr do you have any ideas for how to fix this without breaking other tools? It's unclear to me why gatsby failed...

I might even have the time to do a PR ;)

Refs:
gatsbyjs/gatsby#12759
parcel-bundler/parcel#221

In the meantime #742 and #237 should probably be re-opened, right?

[wpyoga]

Hello. I noticed a problem with docusaurus a few days ago, and traced it back to this apparent problem in chokidar. This is the issue: facebook/docusaurus#5252 . I've set up a repo to reproduce the issue: wpyoga/repro-docusaurus-5252, hopefully that can help to diagnose the problem. I've only tried it on my Linux laptop though.

The problem manifests itself in this way: when I run the docusaurus development server, and then make changes to its config file docusaurus.config.js using vim, only the first change will trigger a rebuild of the site. Further changes wouldn't. Also, other editors including nano and VS Code don't trigger this issue.

Based on my experiments, what I think happened was:

1. When vim saves a file, it will write the data to a temporary file, then remove the original file, and then replace the original file with the temporary file.
2. This setTimeout()'s callback runs before the delay value passed into setTimeout(): https://github.com/paulmillr/chokidar/blob/master/index.js#L579-L585 . This caused the file to be removed from the watch list, so further changes to the file does not trigger an event.
3. I was able to reproduce this in my repo.
4. Outside of chokidar, I could not make setTimeout() execute its callback before the delay specified.

Since this is actually a PR, should I post this on an issue somewhere? Maybe on #742? (but that issue has been closed)