NSFW CPU consumption higher in v. 1.34.2

[colin-grant-work]

Bug Description:

We have performance monitoring that tracks the resource consumption in our Theia-based application. We recently uplifted from Theia 1.28 to Theia 1.34 and saw a precipitous increase in the CPU usage of the file watching service. In the same large workspace, the Theia 1.28 version uses between 0.5 and 10% of a single processor core for watching. The Theia 1.34 version of the application uses > 10% consistently, and it regularly shoots up to 100% for several seconds at a time; using language features, it will go up to 100% and stay there.

Steps to Reproduce:

1. Open a large workspace in earlier Theia (tested 1.26)
2. Run a resource consumption tracking utility like HTOP.
3. Observe that the ipc-bootstrap process for NSFW resource consumption is moderate. If you don't interact with the workspace, it will fluctuate between about 20% and 60% of a core before settling down to effectively 0.
4. Repeat 1-2 with Theia 1.34.
5. Observe that the ipc-bootstrap process consumes multiples more resources than before. It will fluctuate a bit, but after a couple of minutes, it will go to 100% and pretty much stay there.

Additional Information

• Operating System: RHEL
• Theia Version: 1.34.2

[colin-grant-work]

With git bisect, I've confirmed that the problem in our downstream application came with the uplift to 1.34; in the Theia example application, it goes back to dc62cda. Likely, the version changes in that commit only propagated down to our application with its own lock file when a range requirement was changed later that forced a new version resolution.

[colin-grant-work]

Downstream, reverting to nsfw@2.2.0 and node-addon-api@4.3.0 (the version predating that yarn upgrade PR) seems to fix the issue, though it means we miss out on any improvements in NSFW in the meantime.

[tsmaeder]

@paul-marechal was Axosoft/nsfw#124 and attempt to fix this problem? Did anyone ever understand why the CPU consumption increased?

[paul-marechal]

The goal of that PR was to save on inotify handles because it is a limited resource on linux. I don't remember anyone looking into high CPU consumption in our team.