Amazon EBS Failure Brings Down Reddit, Imgur, Others

Several readers have sent word of a significant Amazon EBS outage. Quoting: "Amazon Web Services has confirmed that its Elastic Block Storage (EBS) service is experiencing degraded service, leading sites across the Internet to experience downtime, including Reddit, Imgur and many others. AWS confirmed on its status page at 2:11 p.m. ET that it is experiencing 'degraded performance for a small number of EBS volumes.' It says the issue is restricted to a single Availability Zone within the US-East-1 Region, which is in Northern Virginia. AWS later reported that its Relational Database Service (Amazon RDS) and its Elastic Beanstalk application plaform also experienced failures on Monday afternoon."

Comments

I hope this doesn't affect Facebook.

[Anonymous Coward]

Or else my afternoon is going to totally suck.

Re:I hope this doesn't affect Facebook.

[sortius_nod]

I'm just glad I moved my hosting away from AWS. It seems they've had a few problems lately in their datacentres. Local Aussie hosting seems to have better bandwidth anyway.

Re:I hope this doesn't affect Facebook.

[darguskelen]

TBH self hosted sites are at the mercy of your ISP or Data center.

Productivity up

[Phisbut]

Productivity reached a record high this afternoon.

Re:Productivity up

[fustakrakich]

Should we expect a baby boom in nine months?

Re:Productivity up

[Anonymous Coward]

More blind and hairy hand people, probably.

Re:Productivity up

[MichaelSmith]

Not from reddit users (or slashdotters for that matter).

But But But

[Anonymous Coward]

It's the cloud! It's like never like down, and webscale!

Interestingly enough...

[Anonymous Coward]

Since no one can go on reddit, they will come back to /. only to find out why reddit is down!

Re:Interestingly enough...

[maxwell demon]

Hey, confirming is Netcraft's job!

Re:Interestingly enough...

[KodaK]

All of those things were done here before they were done at reddit. You might want to get a new prescription for your rose colored glasses.

Other Victims

[Revotron]

Coursera is also down as a result.

define "leading" ...

[magarity]

/. is working just fine.

Are those karma points in the mail?

Oblig

[sortius_nod]

It's as if millions of geek voices cried out in terror & were suddenly silenced.

Re:Oblig

[sortius_nod]

Holy shit, when did memes get banned from the internet?

Re:Oblig

[jeffmeden]

Holy shit, when did memes get banned from the internet?

reddit is down, he is expecting to see nothing but NEW shitty in-jokes and hasty photoshops as he takes refuge from the storm... your attempt to re-use old humor would normally earn you a downvote but he cant find the thumb buttons on this jalopy of a website.

Re:Oblig

[Bigby]

If a geek cries out in terror and there's not site to read it on, do they really cry out in terror?

Single AZ my butt

[Anonymous Coward]

We are seeing EBS problems across multiple AZs with our services, as are many others. Amazon is downplaying the issue.

See HN for ongoing discussion as well: http://news.ycombinator.com/

Same region as the storm in June

[bill_mcgonigle]

Bad luck if you're hosted in the US-East-1 Region, I guess.

Heh, I should really start advertising the LVS clusters I tend to as 'private clouds with better uptime than Amazon'.

Re:Same region as the storm in June

[malakai]

According to amazon, it's not an outage, it's a "performance disruption". My guess is, this will negate costly concessions based on SLA's.

Re:Same region as the storm in June

[RulerOf]

Real bad luck.

Desk phones and SIP clients out for 2.5 hours for me. Calls rolled over at the provider level like they were supposed to though. Didn't think I'd have to put that to the test so soon.

The server qualifies for the free tier, and that's probably why it just went straight unresponsive for two hours. Maybe I should upgrade to a slightly larger paid/reserved instance and..... Wait, I smell conspiracy.

Low Availability?

[mkosmo]

I have to admit, due to this outage I just logged in to Slashdot for the first time in a year. We're experiencing our own outages at work, unrelated to AWS, but I'd hate to be an AWS admin during one of these major outages. This makes me wonder why Reddit, Imgur, etc., don't have presences in multiple availability zones to prevent this kind of outage.

Re:Low Availability?

[Anonymous Coward]

>Reddit, Imgur, etc., don't have presences in multiple availability zones to prevent this kind of outage

They do. It's a multi-AZ outage, despite what Amazon is saying.

Re:Low Availability?

[segedunum]

We're experiencing our own outages at work, unrelated to AWS, but I'd hate to be an AWS admin during one of these major outages.

I used to be an admin working on AWS through some of these outages, and it's not pleasant let me tell you. The amount of redundancy you need to get through this makes putting stuff in the cloud prohibitively expensive and things are basically out of your hands. When you run your own servers you know how long it will take to replace a piece of hardware or take emergency measures to keep things running. At least you know you have control over the process. Amazon? They recover what they can of your EBS disks in a few days without telling you anything and in the case of the European outage they actually screwed the EBS snapshots with a recovery job they ran. Thankfully I ran backups every night that took all data off Amazon's system. All I didn't know was when I could be back up and running.

Using AWS for throwaway computing where you just want some computing power for a few weeks of the year? Yes, fine. Permanently running stuff in it? Nope.

Re:Low Availability?

[segedunum]

They do. It's a multi-AZ outage, despite what Amazon is saying.

Amazon's multiple availability zones stuff is total bullshit. It has become painfully apparent during every single one of these outages that the so-called availability zones are not separate because an EBS problem propagates everywhere. No one can actually work the availability zones out either because what Amazon cunningly does is call zones by different letters for different customers, so availability zone 'a' for one might be availability zone 'c' for another so no one can actually compare. That fact alone sent my bullshit meter off the scale. It just seems excessively evasive and sneaky for my taste.

If you want redundancy you are going to have to go to completely geographically separate zones. Keeping those zones in sync is prohibitively expensive for the vast majority. Either that or you have a backup cloud provider, but again you have to be so paranoid and trust Amazon so little that you have to be able to have your data out and off Amazon's infrastructure at least nightly at a moment's notice. Sorry, but that just doesn't work.

Re:Low Availability?

[segedunum]

Remember that availability zone 'a' might be 'd' for others. Amazon does not let you work out what availability zones everyone really has.

Re:Low Availability?

[eln]

That's old Web-2.0 thinking. We're in the era of the cloud now, and the cloud is magic. Trust the cloud.

Re:Low Availability?

[segedunum]

....and in the case of the European outage they actually screwed the EBS snapshots with a recovery job they ran. Thankfully I ran backups every night that took all data off Amazon's system. All I didn't know was when I could be back up and running.

I felt this was worth emphasising. These are EBS snapshots, not just the EBS disks - the ones supposedly stored in S3 and immune to corruption. Your backups, in other words. If you use RDS you rely on these completely for backup.

AWS is OK to get yourself up and running without paying huge amounts up front for hardware, but be aware that you just simply cannot trust this infrastructure.

Re:Low Availability?

[i_hate_robots]

Multi AZ IS "completely geographically separate zones" and yes, you can specifically define which ones. Amazon is very clear that US East 1a,b,c,d are all the same physical data center. However, West is not. It's in Oregon (as opposed to VA for East) I've seen no evidence that true Multi AZ instances (as described by Amazon) are down. If you've got some though, I would be interested to see it because I would be pretty concerned.

Re:Low Availability?

[hawguy]

Seems to me that the answer is just to host things yourself, instead of relying on another company's infrastructure.

How do you host anything without relying on another company's infrastructure? Do you purchase right-of-way's between your site and all of your customers and string your own fiber? Do you run your own power plant? Do you build your own UPS, right down to the batteries so you don't need to trust a UPS vendor? Do you build and service your own CRAC's?

It's impossible for any company to *not* rely on another company's infrastructure even if just for internet connectivity, the only question is where to draw the line - do you really want to rack and stack your own servers? Do you trust a vendor to do periodic preventative maintenance on your generators, or do you use your own staff? Do you certify your own staff to service your fire suppression system, or do you contract out to a vendor? Do you want to own your own network equipment and do your own network admin? Do you want to swap out servers and disk drives when they fail? Do you keep staff electricians on-hand to take care of electrical issues? Do you want to run a 24x7 NOC to monitor and maintain your datacenter?

While a large company may be able to keep many of these tasks in-house, many small companies can't afford the staff it would take to control all of their infrastructure.

Re:Low Availability?

[hawguy]

Multi AZ IS "completely geographically separate zones" and yes, you can specifically define which ones.

Amazon is very clear that US East 1a,b,c,d are all the same physical data center. However, West is not. It's in Oregon (as opposed to VA for East)

I've seen no evidence that true Multi AZ instances (as described by Amazon) are down. If you've got some though, I would be interested to see it because I would be pretty concerned.

Availability Zones are not geographically separate - regions are:

http://aws.amazon.com/ec2/#features

Availability Zones are distinct locations that are engineered to be insulated from failures in other Availability Zones and provide inexpensive, low latency network connectivity to other Availability Zones in the same Region. By launching instances in separate Availability Zones, you can protect your applications from failure of a single location. Regions consist of one or more Availability Zones, are geographically dispersed, and will be in separate geographic areas or countries

Re:Low Availability?

[segedunum]

Multi AZ IS "completely geographically separate zones" and yes...

Availability zones are not geographically separate nor is there any evidence that they are geographically or even logically separate from the nature of every major EBS outage there has been.

Amazon is very clear that US East 1a,b,c,d are all the same physical data center. However, West is not. It's in Oregon (as opposed to VA for East)

a, b, c and d are availability zones. US East, West etc. are different regions. I'm afraid you're not understanding just what is meant by availability zones or just muddying the waters.

I've seen no evidence that true Multi AZ instances (as described by Amazon) are down. If you've got some though, I would be interested to see it because I would be pretty concerned.

As I've said above, Amazon makes it as difficult as possible to verify availability zone failures because AZ 'a' for one customer might be 'c' for another and 'b' for another, so you can't verify anything with others. However, it becomes very clear when you get on Amazon's forums and look at major sites that have implemented in multiple zones from their perspective that they are down and have EBS problems in different zones they have. You don't get much more evidence than that.

If you're not concerned when looking at that then I smell some apologism I'm afraid.

Re:Low Availability?

[aaarrrgggh]

No, most companies can't take full control of their infrastructure. But, they can diversify across two providers, and try to ensure that there is no major work in common change control windows. In a perfect world, you would have hosted services that support 100% of your peak needs, plus a hot disaster recovery site in your own facility that can handle your full average load.

Unfortunately, I am sure that there is some sorry company out there that says "Let's PM all of our generators at the same time this weekend!"

Re:Low Availability?

[eWarz]

I'm afraid to say, you guys are doing it wrong. Currently building an eCommerce platform that scales across any server, even if said servers are across multiple providers. Oh and it'll only cost us about a hundred bucks a month. The cloud isn't about throwaway computing, the cloud is about scalable applications. If you use EC2 for static hosting you are doing it wrong.

Re:Low Availability?

[petsounds]

Caution: Magic Cloud may suddenly accelerate to dangerous speeds.
Do not taunt Magic Cloud.
Warning: Failure to believe in Magic Cloud may result in a targeted nuclear strike in your availability zone.
Magic Cloud should not be used if you are feeling angry.
Never ask Magic Cloud to play a game.

Magic Cloud: satisfaction guaranteed!*

(*) Except for satisfaction-free areas. Please consult your Service Level Agreement for more information.

Bright and Sunny Skies Today!

[IonOtter]

Do you still think that putting your digital life in the "cloud", without any ability to fall back on a physical hard drive or device, is a good idea?

Re:Bright and Sunny Skies Today!

[Anonymous Coward]

Because physical servers don't ever fail?

Re:Bright and Sunny Skies Today!

[gstoddart]

Do you still think that putting your digital life in the "cloud", without any ability to fall back on a physical hard drive or device, is a good idea?

My first thoughts as well.

A friend was recently telling me about an issue they were having at work ... they host stuff for other people, and have very high-availability SLAs. Unfortunately, the support they have from some of their own internal people is "weekdays 9-5". So when an outage happened, they were dead in the water, because their own people basically said "sorry, we don't do after hours support".

Your SLA is only as good as your weakest link. Granted, some of these sites may not have SLAs, but if you have an external vendor providing some of this stuff, and their service levels suck, then your service level can't be any better.

For me, I can't see why companies would be willing to do this kind of thing. The risks are just too high.

Re:Bright and Sunny Skies Today!

[Anonymous Coward]

For me, I can't see why companies would be willing to do this kind of thing. The risks are just too high.

That's because you don't have an MBA.

Re:Bright and Sunny Skies Today!

[rrohbeck]

No but you can make them reliable if needed.
In the cloud you're at the mercy of the beancounters at Amazon & co.

Re:Bright and Sunny Skies Today!

[hawguy]

Your SLA is only as good as your weakest link. Granted, some of these sites may not have SLAs, but if you have an external vendor providing some of this stuff, and their service levels suck, then your service level can't be any better.

For me, I can't see why companies would be willing to do this kind of thing. The risks are just too high.

Because many companies are not willing to spend what it takes to get availability greater than what they can get at Amazon - especially if they take advantage of multi-AZ or multi-region redundancy.

Sure, having a physical server at the office that you know you can fix by buying parts at the local computer store sounds attractive. Until the day you find that your building has burnt to the ground. Or a truck knocked over the utility pole providing network and electricity to your building. Or you discover that when you looked at the flood maps to make sure you weren't in a flood zone, the maps didn't account for a water main breaking and flooding the basement where your telecom equipment is... or the clogged roof drains that let 20,000 gallons of water to build up on the roof during a rainstorm until the roof collapsed and flooded your datacenter. Or the earthquake (or hurricane or tornado or flood or whatever) that takes down your site for days or weeks or even months, and your employees are more concerned with surviving than trying to get your critical systems back online.

Meeting an SLA for your own facility only works when that facility is running, and often the company that rents office space has little control over the facility.

My company has a number critical services running in one Amazon region with replication to a second region for failover. The second region costs very little, just a single instance to hold data replicated from the primary instance, then if we need to spin up the servers in the secondary region, it takes about 10 minutes to push the data from the local copy to the other servers once we start them up.

We could automate the whole process, but Amazon problems are rare enough that it hasn't been worth it.

We do have a couple servers in us-east-1a but so far those servers appear to be fine, although the AWS management interface has not been working for managing servers in that region/AZ. If we ran servers out of our local office instead of Amazon, we would have had at least 2 instances of complete downtime in the past year - one 3 hour internet outage, and a 48 hour power failure on a weekend when a transformer blew and the power company didn't have an available spare and had to truck it in from out of area.

Re:Bright and Sunny Skies Today!

[MoNsTeR]

If you think the risks of running in the cloud are less than the risks of running in a traditional data center, you're very much mistaken.

If one AWS AZ goes down I can bring up servers in a second one. If one AWS region goes down I can bring up servers in a second one. In fact to hedge against these risks I *already have* servers in multiple zones and regions.

Sure you can do that with traditional data centers. Just host your stuff across more than one, right? Do you have any concept of what that COSTS? Especially if you, say, want to add servers in multiple data centers, or move servers from one to another. Plus now you have multiple vendors, contacts, SLAs, and so son, and so forth. And heaven help you if you ever want to *decrease* your capacity. Have fun selling those servers on ebay.

Reddit and friends are suffering downtime from a single AZ outage because their architectures have single points of failure. Don't build your systems that way! If you have single points of failure it doesn't matter whether you're hosted in the cloud, in a commercial data center, or in your own data center. Conversely if your architecture is good and doesn't have single points of failure, the hosting question comes down to this: what do you specialize in as a business? If that list doesn't include "running a data center", don't run your own data center. If it doesn't include "maintaining a shit load of hardware", then don't host in a commercial data center either, run in the cloud. I think you will find that this latter category includes 99.99% of businesses.

Re:Bright and Sunny Skies Today!

[hawguy]

You're talking like hosting your own servers on premises or being in the cloud are your only choices. You could also rent space in a high quality data center and replicate you data out to another high quality datacenter where you also rent space in a different geographic location. Then, when your primary data center goes down, you switch over to the other one. Or run off both at the same time if your architecture allows you do do that. That basically covers you in most instances. If both your rented datacenters go out at the same time, and they are in different locations, there's probably much bigger things to worry about. Or you didn't pick very good datacenters in the first place.

Isn't that the same as putting your servers into multiple Amazon regions? You're still putting your destiny in your hands of the datacenter.

multi AZ?

[i_hate_robots]

An honest question, why don't these large, big-name sites utilize the Multi Availability Zone failover that Amazon offers? It seems these AWS outages make for good headlines, but shouldn't any large site be co-located in multiple physical locations to ensure uptime? If they WERE using Multi AZ, or there is some other technical reason why it wouldn't help, I'm really curious to know why...

Re:multi AZ?

[segedunum]

Do you have any evidence of this? Because I haven't seen any. And it sounds tin-foil-hat.

Sites who implement multiple across multiple zones are down and the forums are full of customers who complain about EBS slowdowns and problems regardless of the availability zones they personally use. You're an apologist if you haven't grokked this yet.

Actually, I run a load-balanced, redundant site on AWS. I ask the question because Multi-AZ (as defined by AWS) means geographically different...

This is total rubbish. Availability zones are not geographically separate, and don't give me that 'as defined by AWS' crap to give yourself a back door (they don't, anyway). Expanding to multiple regions which is the only thing you can do is not the same thing.

as in US West (in Oregon) vs US East (in Virginia) - NOT just the difference between US-East-1a,b,c,d (which Amazon makes very clear are in the same data center). That's why it's odd that Virginia's issues would affect Oregon (or any of the other AZs)

No, Amazon is very, very clear on what an availability zone actually is. Stop trying to make AZs out to be separate regions to get yourself out of this. They are not.

Try being helpful next time and answering the genuine question instead of smarting off because you can't get on reddit.

I'm afraid you don't run any geographically separate system that spans multiple regions because it is prohibitively expensive to do so. You don't maintain AMIs and backups in different regions and you don't pay for the extremely large amount of bandwidth you need to keep those regions mirrored and synchronised.

Sorry, but you aren't doing what you say you're doing and you don't know what the difference between availability zones and regions actually are, which was central to the question you asked. You were called out on it.

Re:multi AZ?

[c0lo]

If they WERE using Multi AZ, or there is some other technical reason why it wouldn't help, I'm really curious to know why...

Here's your answer: cascading failures.

In short, the cascading failures don't happen because one local failure cause the entire capacity of the network to be exceeded... you see, it is not a case of every node connected to every node (O(N^2) connections), thus a failure only need to overload the capacity of the nodes connected to the failing one...

But the cloud is so much better to use!

[BetaDays]

But the cloud is so much better to use!

As Usual...

[broginator]

There's an oblig xkcd: http://xkcd.com/908/ Guess someone tripped over the wire.

No Fancy Uptime Numbers for them

[NinjaTekNeeks]

Looks like there won't be any fancy reports about the "cloud" having spectacular up times, with over an hour passed they can no longer claim more than 3 nines uptime.

Re:No Fancy Uptime Numbers for them

[Revotron]

Well... as it's currently referred to, the "cloud" is a singular entity. So, as long as there's one single server running as part of that infrastructure, you could weasel your way around any downtime and reassure the ignorant masses that "the cloud" is is still up, even if the only remaining piece is a Raspberry Pi running over a cable modem in some guy's basement.

Hey, look everybody, the cloud is still up! You can't do near as much as you usually can, but it's up! 100% uptime! Woo!

wow, mainframe problems in the cloud

[Dan667]

If only there were some lessons learned over decades and decades of mainframe use that that could be applied to the cloud.

Re:I don't...

[Antipater]

Hell even if the internet would go down I still have a phone book and a land line.

Hey, me too! It's always good to have things lying around to club people with when civilization ends.