Mailgun downtime resulting from Rackspace cloud server reboots | Hacker News Reader

Mailgun downtime resulting from Rackspace cloud server reboots

status.mailgun.com

41 pointsjhealy34 comments

Expect this to happen on more services, too. Rackspace customer are reporting uncontrollable downtimes (because the physical hardware is down) ranging from "a few minutes" to "six hours." Search twitter for [rackspace reboot].

Also, since Rackspace didn't communicate it that effectively: if you're on their first-gen infrastructure (your servers have an asterix next to them in the list), you won't get rebooted. I had to hear this from a fellow customer who heard it from support, as I was searching my inbox frantically wondering why they hadn't given me a heads-up about an incoming-any-minute-now reboot yet.

> since Rackspace didn't communicate it that effectively

Unless I missed something here, I didn't receive the notice until going home for the weekend that this was happening this weekend. Hopefully someone from Rackspace is reading this: If you're going to reboot every VM on a weekend, tell us before the weekend, please.

I can picture this happening

Rackspace PR: lets release the notice Friday night so we don't get bad press.

I think it's more like "lets release the notice and reboot the servers as quickly as possible to urgently fix this urgent security problem".

Although I do agree that more notice would have been helpful.

As quickly as possible? This is the same Xen issue that made Amazon reboot their EC2 instances. They announced it significantly before Rackspace. Both RS and Amazon have the info from the same source -- the Xen vulnerability pre-disclosure list which keeps this issue (XSA-108) under wraps until Oct 1st while major cloud providers can apply the patch.

Rackspace may have thought that they could mitigate the issue without taking this drastic step. They certainly didn't want to do global reboots.

The engineering team probably spent some time running tests and scribbling on whiteboards, trying to prove that the boat wasn't going to sink. In hindsight, they should have just sounded the klaxon and started handing out life jackets, but you know what they say about hindsight. And there are lots of reasons why the typical engineering organization struggles to accept the inevitable and call for an evacuation. Nobody likes Cassandra. Everybody wants to be a hero. Didn't you say this boat was unsinkable? It's hard to get all the decision-makers into one room. The show must go on. It isn't obvious that this complicated problem leads to our certain doom. Et cetera.

The key to making these things go smoothly is the Chaos Monkey, a.k.a. "conduct constant drills of your emergency responses". If you don't rehearse the response, you shy away from trying it. AWS halts or reboots EC2 instances all the time, and lo and behold, when it comes time to reboot all EC2 instances they don't flinch. Or they flinch less visibly, anyway.

I got bit by this over this weekend, too. I received the initial email September 27 @ 1:35 AM local time. I called Rackspace to see if they would be sending out emails prior to each VPS being rebooted 30 minutes or so before just as a heads up for human monitoring purposes. The representative told me that it "was not feasible" for them to do that for every VPS. Instead... I've been camped by my computer all of today (Sunday) to monitor the reboots since I have servers at DFW & ORD and they have a 24 hour time window for those regions.

While their status page was somewhat helpful, I find it absolutely absurd that they can only update it once every 60 minutes to cue customers in. In addition, their Cloud Control Panel doesn't reflect the reboots. When a VPS goes down for a reboot... the CP shows the server as "Active - Online and functioning as properly". Thankfully great third party monitoring services (using Scout) exist so they can notify in-place of Rackspace's incompetency.

As someone who shells out a significant amount of money to them each month, this is pretty disheartening. That being said... I seem to have survived the Great Rackspace Reboot of 2014 and can only hope they handle the next event better.

We've been given a 6 hour maintenance window. They won't give us a smaller window or tell us when it's finished. Apparently, we should apply our own monitoring - fortunately, my tech chaps are psychic and can tell if monitoring is down because Rackspace maintenance is taking place or because it's taken place and something's broken.

We're a MG user via Heroku.

Our Heroku MG logs indicate that all messages are getting "Delivered", but that doesn't match reality. We've been testing with our own accounts -- ones that receive copies of all automated emails, as well as our personal accounts.

The Heroku MG logs says "Delivered" for all the emails.... but using 4 different addresses across 4 different carriers confirms: a VAST number of emails (since noon today) were not actually delivered. The only change in configuration: MG's downtime. I seriously hope all of these "Delivered" emails are re-sent. If someone from MG could weigh in, that would be fantastic! (We have a ticket filed, but email also in profile).

I've asked our support team to contact you.

Mad props for the prompt replies even during crazy times. I'm routinely impressed with MG.

Mailgunner here: usually we just redirect traffic from one environment to another, but this time we are having unexpected networking issues that are preventing us from that. We are still debugging the issue, stay tuned on status.mailgun.com

I don't get it. If a server restarts, don't you just get a different one? Why is that disruptive?

At work every machine reboots at least every month. Everything is designed to cope with that reality.

Did you bother to read either of the other comments in this thread before posting? Not only are the reboots of every system happening simultaneously and taking up to hours to complete, the specific issue the post links to is affected by networking issues.

Yes I read both of those rather uninformative posts and the OP. What fraction of servers are going down at once? What fraction of VMs are not scheduled?

All next-gen (i.e. created within the last 20 months) servers are being rebooted at an undeterminable time within a 3-4 hour window, the window specific to the datacenter they're in. No next-gen servers are not being rebooted.

100% churn in three hours is a bit aggressive. I'm not a rackspace customer but I'm wondering why they don't (if they don't) offer service level guarantees for eviction rate and time spent not running.

This is a little different from the usual "virtual server goes down" scenario. This is presumably to patch the new undisclosed Xen bug so the actual physical machines running the virtual servers need to go down.

> At work every machine reboots at least every month. Everything is designed to cope with that reality.

Count your blessings. Not all of us work with software so well designed. For us, at least, a server going down can be anywhere from an "eh" to a desperate need for attention, depending on the exact VM. Some technologies just don't take it that well. (Including many that are on HN.)

Further, every server went down in a very tight window. If you weren't replicated across regions, for all you knew, every VM might be rebooting at the same time. I know you should be replicated across regions, but again, count your blessings.

It's disruptive if you have a long running tasks and you make sure your planned reboots are in between runs - but when rackspace reboots you with barely any notice that's a big problem.

And not every organization has the budget for multiple machines, many run on just a single machine. Yes, that means accepting downtime may happen - but that doesn't mean being happy about it.

Ah, Rackspace with your 'fanatical' support!

I suppose it takes a fanatic to justify working week scheduled down time.

Anyone recommend a good dedicated server Rackspace competitor?

Hmm... is it just a coincidence that our school services portal had a semi-scheduled downtime this sunday from 6am-6pm, but we weren't notified of this until late friday... Also there is never maintenance for this and when there is the official time is 2-5am on friday....

This is when I enjoy our new Google Compute Engine instances with transparent maintenance (thanks to live migration).

https://cloud.google.com/compute/docs/zones#maintenance

To the admin who had to hit <Enter> on the script that kicked off these reboots.... I imagine this guy turning and looking the other way before clicking.

What poor execution on this...Xen updates can't be that rare that this is their first rollout.

Given that both Amazon and Rackspace (and presumably others) Hit The Big Red Button on this one, I'm inclined to believe the security hole is bad enough that it was worth the panic.

I understand reboots are necessary, but why not migrate the instances from a node to a different node, reboot the node now that there are no instances running on it, and do this in a rolling manner?

Because it's unknown which servers and when - they are bouncing physical servers... We avoided any major downtime by being multiple dc's

I understand that Rackspace is rebooting physical servers. They also have the knowledge to know what VM's are running on said machines, and they also have the ability to migrate the VM's from one compute node to another compute node.

If they have network block storage you would likely be correct, but afaik they also give you emphemeral storage that is local to the hypervisor which is not network block storage and therefore makes it very very hard and very very slow to migrate you around automatically (ever tried transferring a 20GB+ file around, ya, it takes forever...)

20 GB file over 10 Gbit/sec would take 16 seconds.

I just recently migrated a 180 GB instance from one KVM compute node to another in 3 minutes.

Depending on the nature of the patch, they might not be able to migrate from unpatched to patched hosts. KVM upgrades can also be fun that way.

The VMs are _probably using local disk storage, thus live migrations would be non-trivial to achieve and may be impossible given their current setup.

From memory[1], you had to be migrating between the same architecture, and if they're anything like the smaller providers they're running less bigger nodes, and thus it's actually a lot quicker to just push the power button.

[1] I worked with Xen years ago, my information is no doubt out of date.

The Rackspace status page: https://status.rackspace.com/

And nothing of value was lost.

Reply on news.ycombinator.com