[META] Filesync fails silently and breaks watches

[grayside]

Problem

@illepic reports that filesync's silent failures cause developers to work for an indefinite amount of time with uncertainty if troubleshooting fixes are failing because they are wrong or because the filesync isn't carrying changes properly.

As a result, we have wasted time from troubleshooting after a fix has been found, as well as a growing distrust in the filesync system.

This has been seconded by a number of other developers, raising the profile of this issue to seemingly the largest source of trouble for Outrigger users. Thank you to everyone that has spoken up about this problem.

Solution

With such a sweeping problem statement, it is impossible to declare a single solution. Rather, we will treat this like a "meta" issue, a bug that will require multiple changes to fully address. The definition of done should be that this problem stops being encountered for a reasonable length of time.

Related Issues

Here are the issues so far identified to help support this goal:

• Add notes about care and feeding of unison during filesystem operations outrigger-docs#30: Write a temporary PSA in the documentation
• Detect unison failure: unison doctor command? #163: Add a tool to detect sync failures that can be run manually or in the background.
• Automatically restart broken unison processes #164: Automatically restart broken unison processes

Use of This Issue

1. Report specific reproduction steps that cause Unison to crash
2. Report any steps/upgrades to rig that make your problem go away.
3. Suggest changes to rig or the documentation here so they can be coordinated with efforts underway.

[grayside]

As an alternative to building out the restart of unison process inside the container (assuming it's crashing of the server and not of the client we need to be concerned with) we could take steps to build a smarter healthcheck into our unison container image and set something up to auto-restart when it reports unhealthy. This has some answers on how we might do that: https://stackoverflow.com/questions/47088261/restarting-an-unhealthy-docker-container-based-on-healthcheck

[grayside]

I'm working on a rig project sync:check to operate as a sort of doctor check of the unison process.

[grayside]

Collecting some research avenues:

• rig project sync defaults fs.inotify.max_user_watches to 100,000 for the docker-machine. Is this enough? Probably.
  • Do we need to increase this number inside the container? Maybe.
• When lots of files change, or a file is changed while being synced from a change, it can cause the high CPU spikes mentioned in the referenced issue. https://github.com/EugenMayer/docker-image-unison/pull/11/files demonstrates using Monit to monitor for performance and use supervisord to restart the unison server process.

[mkochendorfer]

This has happened to me and other developers countless times. It is incredibly frustrating and wastes untold hours of time going down the wrong paths debugging things that are really just that your code changes are not making it into the container. This is by far the highest priority issue with rig currently today.

[srjosh]

I've run into this quite frequently when working on client work; it definitely seems tied in my case to my host machine going to sleep/waking up. It is definitely frustrating.

[grayside]

Note: This issue is now: support request, problem research, "doctor" research, and autoheal research. I will probably split this apart in the next few days. I'm breaking the "doctor" angle here to #163

[grayside]

I have converted this issue to a METABUG, please re-read the issue summary for details on what we are doing so far and what this issue should continue to be used for.

[grayside]

Further discussion with afflicted users has pointed out one of the major error cases is resume-from-sleep. Improved handling of sleep/suspend/hibernation operations may go a long way to address this problem.

[crittermike]

Some of us have gotten in the habit of just assuming it's broken both when starting dev (for the day or after a break) and also whenever something unexpected happens, and running sync:start proactively before doing anything else.