Automatically repair app when the HealthCheck goes down (Not Responding)

[Lonkle]

This would be a precaution added until @girish is ready to add what he wrote below (I don't know if dockerode has it which would be a huge difference in difficulty in adding that as a feature):

@girish said in Automatically "/repair" app when the HealthCheck goes down (Not Responding):

I would ideally like to remove Cloudron's healthcheck field and replace it with Docker's own HEALTHCHECK (https://github.com/moby/moby/pull/22719). When we started out, that feature didn't exist in docker and maybe it replaces what Cloudron does internally. Once we do that, we can get automatic restarts etc from upstream docker. Even though I note that https://github.com/moby/moby/issues/28400 is open for over 2 years now.

[nebulon]

@ruihildt ideally we would know what the issue was if it affected so many apps. Do you still know the error? If you do maybe we can discuss that in a new forum thread to fix what caused it for the next update.

[d19dotca]

I think this is a great conversation, and great points from different points of view. For what it's worth, my two cents is that the health check should indeed do more than just simply log it, it should take a predetermined action. This is how other health checks work on different platforms and I think is the general expectation of users coming to Cloudron from managing a Docker Swarm cluster, or using Kubernetes, etc.

I think the appropriate measure would be for health checks to be logged as they already are, and for a new option for an admin to set an action to be taken whenever a health check state goes to "Error", for example, such as automatically restarting the node. I've personally never encountered an issue yet where an app in "Error" state isn't fixed by a simple reboot of the app, and even fi it wasn't and was then in a reboot loop, well if it wasn't working anyways then not much has really changed so it can only really help in a way not harm.

[Lonkle]

@nebulon And that's what you should do, of course, but it'd be a lot easier of him to do if he'd gotten a notification with the portion of the log of when the app went down, and had it automatically repaired (both the logs shortly before it repaired and shortly after would be delivered with the notification). He could get you that info at anytime (if a notification system like I proposed is created). It's far too late for that now - making his ability to report what happened, effectively more diminished than if we agreed on a position in this thread on where we should go in terms of reporting to devs and pre-determined actions. I argue pre-determined actions are the point of a healthcheck but, hmm.

This is an ideal you and I fundamentally disagree on and no matter what you have the final say, but I'm feeling meeting in the middle with my proposal to increase the ease of use / ability for users to report to the devs when apps go down, and it increases uptime.

That was me taking what you wanted (accountability and log reporting), and what I wanted (higher uptime, even if it is only a 0.0001% difference, it'll make me and other users feel safe in that way). I'm a developer, but user experience is key - so I always put myself in their shoes while debating the way to make these kinds of infrastructure changes.

Basically, @nebulon, what I'm asking you is - does my proposal not only solve both of our problems at once, but actually increases the ability for users to do what you wanted them to do in the end anyway (hit repair manually and report the log) as long as we do it in the right way as has been discussed?

[Lonkle]

I would also like to point out that @girish said in this thread that this is going to be a feature down the road as Cloudron gets more Docker features. So, it's already a Docker feature, and they must have also discussed this. I wonder if I can find the ticket for their discussion on this very subject. But thanks @robi for finding and posting this, I thought it was a good read about this whole issue:

https://github.com/moby/moby/issues/28400#issuecomment-712510304

[girish]

@nebulon probably didn't communicate properly but it's not about implementing the feature or how it can be implemented. He is saying that we need to understand the root cause of apps not responding in the first place. Once we understand the problem, we can think of the correct solution.

We have 5-6 Cloudrons ourselves and essentially never have to repair/restart apps. In fact, this whole repair stuff was only added some releases ago and we thought even that was uncommon

Anyway, this thread exists so people can tell us if apps are not responding often. Like things stop working every day? every week? every month? Depending on the various experiences, we can try to figure out how to solve it. For example, if you had to restart like once a month, it's already not a priority with respect to our massive back log. So far, I have noted two, let's hear from more users. But we made Cloudron so one doesn't have to deal with all this stuff about apps going up and down.

[Lonkle]

@girish Like I said, it happened to me only once between updates so the system is pretty stable. I outlined a theoretical situation where it would be needed. Then another user mentioned an important use case for simply adding a single endpoint to reaction to a Not Responding status dashboard function.

But you yourself said Docker supports this, and multiple people have expressed that customizable actions based on "status changes" of apps are beneficial. I even outlined how to make @nebulon's and your main issue with this work even better while still increasing uptime. Increasing uptime for users was important enough for Docker to implement it themselves with their heathchecks. Even supervisor.

The notifications can literally say "We restarted your app x, if you notice this happening often with this app - click this to send the log directly to the developers."

Only @mehdi appears opposed to this though I couldn't tell, I'd say it's almost unanimous and why wouldn't we compromise to make everyone happy app devs, Cloudron developers (you and @nebulon), and especially users.

But the point you guys made (yourself and @nebulon) is valid, it's just overruled by using the notification system you already have in place to make this situation so much easier for the users to repot more to mitigate your concerns while simultaneously increasing user reporting and increasing app uptimes. Win-win - does anyone disagree with that?

Note: I agree with @girish that we should hear from more users and their opinions btw.

[Lonkle]

Also, @girish, I wasn't saying that /repair wasn't overkill, I was saying, why not use the overkill option that already exists and has claimed (by users) to fix more issues than /restart to try to fix something as bad as literal website downtime. What if @ruihildt had been on vacation and Cloudron auto-updates (as it does) and it caused the need for a manual repair of all of those apps (he said 20 per installation). Yes, a dumb admin user shouldn't exist but what if he didn't know what to do and then @ruihildt had to be interrupted to fix these apps, each one by one. When he could have, after his vacation, clicked the "Report to Developers" button in his Notification Center when he sees them. He's more likely to report, you're more likely to have the data you want, he's a happier user. And the rest of the users feel comfortable with that one more protection against downtime.

Btw, the one time it happened to me was after a Cloudron auto-update, with the Wordpress Unmanaged app. So there could be a common factor here. But the fact that we don't know it and this solution accounts for this and other potential user-downtime-impacting situations. It's just very important to consider for developers and users alike. But I think we've all made it a point to consider this so maybe we'll revisit it later.

[mehdi]

@Lonk said in Automatically repair app when the HealthCheck goes down (Not Responding):

Only @mehdi appears opposed to this though I couldn't tell, I'd say it's almost unanimous and why wouldn't we compromise to make everyone happy app devs, Cloudron developers (you and @nebulon), and especially users.

Nope, not opposed to it ! I just did not understand what /repair does, more precisely what the difference was between it and /restart

To be clear, I'm not opposed, but I am not lobbying for implementing this either. I've literally never had to restart / repair an app to fix it => I don't really have an opinion on the matter.

[Lonkle]

@mehdi That makes perfect sense, since I've encountered it personally; it makes me a little more opinionated on the resilience and uptime of the apps in an app manager. I'm only proposing a single repair as soon as an app's status moves to status of "Failed" or "Not responding...". If it doesn't work, it doesn't work. If it does, then we only have uptime to gain here.

[robi]

It sounds like we need a multi-step approach.

First, improve logging, capturing and surfacing of the actual errors in general, per cloudron. Notifications that an app went down then came back up are next to useless.

Second, add more resilience in an escalating manner, and at different timeout lengths and counts. Maybe restart first and if that doesn't do it, repair.

Third, have a way to get telemetry for the Cloudron team from all running Cloudrons, so they have an auto generated view of which apps need more attention because of triggered restarts, repairs and errors.

Less chaff/noise, more signal/automation. That's what we love about Cloudron.

[Lonkle]

@robi Exactly, I want to add automation to this without losing what @girish and @nebulon want, a compromise where we all win.

I really liked where you were going with that analytics of apps thing. If an app restarts more than once a day, in a week we can send a report to the admin.

I'm advocating we treat a suddenly unresponsive app in the same way we treat an app that takes up too much memory. We restart it, and the process might fix itself.

Girish did say when we update to the latest Cloudron we'll eventually get this anyway. I'm just saying, why wait a year and instead hash out the fundamental disagreements between developers and users so that everyone could benefit. I liked whenere @fbartels was going with it in the beginning of this thread, and I like how the analytics you bring up could help send admin maybe a weekly notification that their app went down a few time, it may be "unhealthy" and need to be looked into. Something like that. I think notifications have their place, just not sure 100% how which is why I like this thread, to hash it out.

[Lonkle]

Just read this:

https://github.com/moby/moby/issues/28400#issuecomment-713457999

And liked this quote: “Ideological answers are not very useful when people ask for pragmatic solutions to real-life problems.” Although that person is fighting for configurable options on healthcheck fail. I just want a single repair.

Because the healthcheck isn’t any more useful than UptimeRobot otherwise. And you handle running out of RAM the same way (by restarting and notifying the user it happened).

I still want to find a solution where users, app devs, and the main developers all can compromise and figure out a solution to fit our needs / wants on this platform.

[nebulon]

As @girish mentioned, the repair was added as a last resort and nearly always it just covers up a real underlying issue, which should be tackled to avoid repair runs in the future. Given that those issues are not well understood and known currently, essentially what it does is to tear everything down and start the app fresh. It may be docker issues or other things, if we would know what logs or hints we should attach then we would do that, but it usually isn't that trivial.

So for future reference, if you hit a situation needing a repair, please copy the error shown for the last task (visible in the repair UI) and also download app logs and ideally do some basic investigation on the server, like running cloudron-support and save the resulting link. Do all this before hitting repair, since by using it errors might be obfuscated and hard to find afterwards. Also if multiple apps are in this state, see if there is some correlation between the error. For example it could just be that the system got overloaded temporarily after a reboot or such. In many cases there are solutions we can build into the platform, but we have to first understand the underlying issue.

[Lonkle]

@nebulon said in Automatically repair app when the HealthCheck goes down (Not Responding):

if you hit a situation needing a repair, please copy the error shown for the last task (visible in the repair UI) and also download app logs and ideally do some basic investigation on the server, like running cloudron-support and save the resulting link

Since you were able to explain how to do it manually, I thought about it and literally everything you just said could be automated and shown in a notification and there could be a “Send log to developers button.” So, again, if trying to auto-repair once doesn’t work; that’s it. I don’t see -a single con but a huge pro being more uptime by removing an unnecessary human first debug step. Automation is why we use Cloudron in the first place.

Let me put it this way, why do you not just keep apps that run out of memory, stopped, why do you restart them? Uptime is the answer and it’s what the users care about. This isn’t a developer platform where we stare and love logs and make a million forum posts (like me). Users want a real solution to a real-world problem.

This is already built into Docker now - so we’ll eventually get it anyway when you guys update Docker. That’s probs a year away though, so why not band aid it till we all have official Docker support for repair on health check fail.

Also, don’t think I don’t understand why you don’t want to band aid an issue. But instead of band aid-ing your suggestion is to tell users to do work that could be automated and save them time. And like @ruihildt proved. Users don’t actually report this stuff unless it actually is consistent. They never would have told their story without my post.

What I’m proposing is automatic and only serves to increase uptime. Your solution increases downtime and user annoyance if it was a one off platform thing which appears that it is. It only happened to me once after an auto-update but I only had one app. Having 20 do all the same thing and me not being there to repair them and their in a production environment. Yikes.

I’m not asking for an endless loop. Just a preemptive action to see if we can keep an app up without human intervention when it would otherwise stay up with my proposal.

I don’t believe Docker should remove this behavior, nor should supervisor. I think there’s a solution out there for all of us and I want to discuss in good faith ways to solve your problems with the proposal as well as putting users over developers in terms of UX.

[ruihildt]

@nebulon Would it be complicated to automate that 3 steps reporting? (For example, clicking repair would trigger this by default)

Like I said, if I have clients complaining and if I'm not in front of my computer with free time on hand, I'm going to hit that repair button to get back online ASAP.

This was my situation yesterday, I'm sick and stuck in bed with just my smartphone, it would have helped if it was automated, for you and for me.

[Lonkle]

@ruihildt said in Automatically repair app when the HealthCheck goes down (Not Responding):

@nebulon Would it be complicated to automate that 3 steps reporting? (For example, clicking repair would trigger this by default)

Like I said, if I have clients complaining and if I'm not in front of my computer with free time on hand, I'm going to hit that repair button to get back online ASAP.

This was my situation yesterday, I'm sick and stuck in bed with just my smartphone, it would have helped if it was automated, for you and for me.

Exactly, what @nebulon and @girish want could be automated and in doing so - it increases what they want (more data to fix underlying problems) while at the same time keeping the users apps uptime as high as possible. I want everyone to win in this scenario so I want to hear everyone’s pros and cons.

[nebulon]

Well those steps are just generic debugging and investigation steps for sysadmins. They may or may not apply and are certainly not exhaustive, just what came to mind while writing this. Plus we don't generally just send information wholesome from your server to us also for privacy reasons. You can still manually issue a support ticket for that app from the support view, which will include the app logs.

The out-of-memory restart is something different though. The underlying issue is known here and restarting due to out-of-memory is the correct thing to do. There is also no solution code-wise, since we can't just up the memory limit automatically, risking over-provisioning the server. Nor can we add memory to the server automatically.

Again if you have concrete situations where a repair may solve the issue, we have to investigate. Keep in mind that in our experience this is not at all common across our users. Can you imagine Cloudron just issuing a server restart automatically since we found that this often fixed issues in the past?

[Lonkle]

@nebulon said in Automatically repair app when the HealthCheck goes down (Not Responding):

Can you imagine Cloudron just issuing a server restart automatically since we found that this often fixed issues in the past?

Well, of course not. That would defeat the entire purpose of this thread which is to have your apps have as much uptime as needed. Your opposition is, “we, developers need to make a perfect system and fix things so that our healthcheck isn’t even needed.” I agree that that’s the goal but you’re not there yet (both my and the other users on this thread attest to that) so why not protect users to these once off oddities that will never be reported if it doesn’t happen more than once? Because users are still going to hit repair and not do a single thing about it. But if in the notification there was an “ignore” button or “Send before and after logs to developers” then they may just press that and automate all of your manual steps. Because you could make this more complicated if you want - but in the end, it boils down to a user choosing to send you logs to fix an underlying issue. Even if you choose not to repair automatically once per “Non-responsive” / “Failed” app, you should then go further in your direction and make it easy for users to log these events and report to you. I still think a single courtesy attempted repair is best and Docker already has that built in so we’re gonna get it. Well, actually, I’m now making “Dot - The Repair Bot” to monitor the status of my apps and repair them if needed. This was always something I could do - even from inside a Cloudron app. And I intend to do it and release it for everyone. But this seems more suited as a built in feature, a checkbox that says “Attempt a single repair on app failure.”

And sure your steps weren’t exhaustive. But @ruihildt put it best. Average users don’t have time for any of that. I wasn’t advocating for automated sending of the logs. I was advocating for a button in the Notification Center that would send you, the developers, the relevant data you need - because users do not do what devs do (report bugs), but they would if it was a button click. Regardless of where you land on this matter, making it easier for devs to fix their apps is best if you’re going to say the only con to the pro of “more uptime” is “less developer communication.”

You guys for sure won’t be niche forever. So scaling ideas to reduce support requests should be considered heavily. Turns out for everyone in this thread, they repaired all of their apps and it never happened again. So no support requests when I could totally see someone complaining about having to manual repair 20 apps. The fact that you have a health check and this one-off repair has worked for all of us and we never reported it, means a singular automatic repair has no real-life con (aside from so far your theoretical con), but real life pros (uptime and reliability).

Coincidentally, this unintentionally happened to me again while rebooting the server. The app booted up not responding (1 of 4) and I repaired it and went on with my day. I’m a developer and didn’t think anything of it except that I wished it had been automatic in case I hadn’t noticed.

So I just want to ask, do you feel like my goal and your goal can possibly coexist? I can wait for the Docker update which adds this but I think I’ll create Dot to keep everything healthy automatically (which is something I would expect from an app manager). Average customer UX, saying this even as a developer, is more important than developer UX.

[nebulon]

Well really if you hit this just now on a reboot, I don't understand why this would not be investigated. We are happy to help. This is simply not acceptable having to run a repair after reboot. Repair in this case would just mean it recreates the containers after they were already just recreated after reboot. This makes no sense why it would fail first and then not, maybe there is some timing or resource issue lurking. I suspect there is an issue in the healthchecker state and the apps would have become healthy after some time. But all this is up in the air without further info. Repair may actually re-runs the last task! So after a reboot it may run some random unrelated app configuration task.

I have seen many times where users would restart/repair apps just because they felt it took too long for them to start up, such actions are sometimes even risky if an app is starting up. This is also why we don't show the task cancellation button immediately. Randomly interrupting processes is not a good idea if it can be avoided, as it heavily depends on how an app is written and how it can recover from for example interrupted database migrations.