Within 3 days another power outage at Linode (Fremont) | Hacker News Reader

Within 3 days another power outage at Linode (Fremont)

status.linode.com

41 pointsjrnkntl62 comments

I don't understand what they did the last days if this is another power outage.

In their posting about the outage on the 20th they said: "At this point all we know is a severe lightning storm in the area caused a power outage and redundant UPS systems failed." http://status.linode.com/2010/11/possible-power-outage-in-fr...

Redundant UPS systems failed? And now it fails again? What kind of data center are they running in Fremont?

I worked in a datacenter (serving many companies you've heard of) during a catastrophic power failure that lasted almost 24 hours. It's kind of like a plane crash - it's never one thing failing that causes the problem (that's what redundancy is for) - it's that perfect chain of events, multiple 'once in a lifetime' failures that causes it.

For example, power outage occurs at the same time the UPS batteries are being changed. Bypass fails and Diesel generators fail to kick in. Circuit breakers blow everywhere making it extremely difficult to get the generators back on line. This all happens in the middle of the night in a winter storm (or 'lightning storm) which causes a further delay in response time.

Been there, done that.

Edit: Also, bureaucracy, lack of documentation, and a manager CF added several hours to the outage. Sometimes you just have to STFU and let the geeks fix the problem.

Except, this isn't a once in a lifetime occurrence. As of today it's twice in a lifetime. That's indicates to me they didn't just hit that perfect chain of events, it tells me they have problems in their architecture or implementation that need to be addressed and tested before the next crisis occurs.

I worked in a datacenter as well. The whole point is that you test those scenarios when it's not an emergency so you can be confident everything will operate smoothly.

Don't discount the chance that this was caused by two totally different 'once in a lifetime' chain of events... i've seen it happen.

However, as you point out it is far more likely that there are architecture or implementation issues that caused the repeated outages.

There are so many reasons that I don't want to have servers in California... high costs, dense population vulnerability to earthquakes and a hot terrorism target. Massive bandwidth but massive utilization, somewhat shaky power grid, heck even the state has financial issues. I picked Atlanta for my Linode server which is better (it's still a little close to the coast for me) and I host my WordPress sites at a company in Dallas, TX. Did you know texas has their own power grid - separate from the rest of the USA?

What you've described here would make an AWESOME movie. Granted, I'm not sure how one would work "Datacenter" into the title and still sell many tickets.

Cory Doctorow (of Boing Boing fame) wrote a sci-fi novelette about a datacenter during a crisis: http://craphound.com/overclocked/Cory_Doctorow_-_Overclocked...

I personally didn't find this piece particularly great, but there's an entry in the genre for you.

What makes it even more exciting is that some of the companies were a bank and a hospital. Lesson learned - a bank gets far more anxious than a hospital when their servers go down.

I'm pretty sure there were Zombies roaming the aisles of servers where the power was out, that might help the box office sales.

> multiple 'once in a lifetime' failures that causes it.

That reminds me of hearing mention of "multiple six-sigma events" I think it was, when things were crashing in 2008: the only sane conclusion is that the models are broken.

Linode doesn't run datacenters, they colo with other providers (Hurricane Electric in Fremont).

HE doesn't have the best reputation for resilient datacenter services. However, redundant power is a very complex problem and prone to failure if you can't afford to do it right... which you can't if you're selling colocation for as cheap as HE does.

I really, really wish Linode would launch a sort of premium offering in better datacenters.

Incidentally, in some ways it seems kind of odd to run data centers in places like Fremont, where both land and people are expensive. Google and Facebook, for instance, have demonstrated that it's quite possible to build datacenters in rural Oregon, where land is a lot cheaper, and presumably labor costs are as well, even though my guess is that in someplace like Prineville, most of their people are imports...

Power is expensive as well (it's usually the big expense), but there are a couple of reasons you might opt for more expensive areas for datacenters, particularly if you're selling hosting.

1) Many companies want facilities that are physically accessible to them, it's very difficult to sell to these companies if they'll have to travel 900 miles to do any work and don't trust the provider to do the work for them, or simply don't want to pay for it. Given the target market for hosting, Fremont's a pretty good option.

2) Depending on the service, a datacenter is only as useful as the transit providers you can access. We do all our hosting at a Chicago facility where we can get a cross connect to any bandwidth provider you can think of. Our services are also targeted at end users, so we're better off being centrally located and using providers with extensive peering agreements (less latency, woo!). If we had serious computation requirements rather than delivery requirements, one of those spiff Icelandic datacenters might be a good choice. I'm unsure of what Google and others stick in their rural datacenters, but I doubt it's anything that slows down your experience.

Power is not the big expense and neither is land. See http://perspectives.mvdirona.com/2010/10/31/DatacenterNetwor... for a breakdown of expenses.

I may be missing something, but that post doesn't address the expense of power, it talks about the relative consumption of networking equipment compared to the rest of the facility. Power is expensive, and James Hamilton generally spends lots of time talking about it, his Velocity keynote (http://www.youtube.com/watch?v=kHW-ayt_Urk) about datacenter innovations focused primarily on power and cooling (heat being a biproduct of power consumption).

Slide 3 says that power, power distribution, and cooling account for only 31% of the monthly cost. So while power related expenses aren't trivial they're significantly less than the cost of servers. I think land would be counted under other infrastructure so it's hardly worth considering.

Locations are selected often due to proximity of a target market for reduced response latency. Also getting transit Tier I transit/backbone providers to lay a network in smaller markets is quite a hard sell. Google and Facebook benefit from being able to commit to large monthly bandwidth usage.

It is probably going to be almost impossible to have a "carrier neutral" hotel in rural Oregon but you might get a single ATT or Level3 or if you're really sucking it, a Cogent line.

HE Fremont is specifically great for reaching markets in APAC and HE itself is pushing IPv6 quite hard and so is one of the few providers that make it widely available.

> It is probably going to be almost impossible to have a "carrier neutral" hotel in rural Oregon

Fair enough, but there's plenty of middle ground between Prineville and the Bay Area. The latter is expensive - it makes sense to have the really high-end things there, like Google R&D, not commodities like data centers and factories.

IMO. I'm certainly not an expert in that sort of thing though.

Incidentally, as a native of Oregon, I still think it's pretty funny that Facebook is building a center in Prineville, heretofore best known as the home of Les Schwab "Free Beef!" Tires.

Definitely I agree. But some clients will pay more for proximity such as monitoring or DNS providers. Chicago(350 Cermak) and NY are an example because of the financial trading markets. Clients pay "through the nose" for a colo spot there because every millisecond counts. If XYZ Trading Company sets up in the suburbs of Chicago, that 5milliseconds makes a huge difference.

But you are definitely right that most people don't need that kind of silliness.

I came from North Carolina where Google and Apple saw the lower power/employee costs as a reason to open datacenters. In the western part of the state is a beautiful city called Asheville which was never known as a network hub. But because of the geographic location halfway between Atlanta and DC a company(uberbandwidth.com/netriplex) decided to create a datacenter there. Lower costs...but the only way IN and OUT of their network was a backhaul through Atlanta or through DC. If one of those backhaul lines go down they've basically lost their single selling point.

Sure, I understand about the financial companies, but, going back to HE, that's definitely not the business they're in: they'd be in SF proper if they wanted to be close to that world.

I wasn't explicitly referring to Linode but to the people running the datacenter (which is, in this case, HE).

I wouldn't so much wish for a premium offering, just better datacenters in general for Linode's racks.

Or you can keep a backup linode in a different datacenter. Or better yet, have redundancy with a completely separate provider, so you can deal with a complete failure of either one provider.

Wow, a "severe lightning storm" in the bay area? My friend mentioned that on facebook too. Not an every-day occurrence.

In SF proper, we saw the 'lightning', but not the 'severe'.

It could have been worse in Fremont, I suppose, or that might just be a bit of puffery being used to explain the outage.

Nope, I was in Fremont, and saw maybe 1 flash all night. Heard no thunder.

you're expecting competence from HE?

i colo'd at HE until last year when they ran ~400V through my racks which had 110V circuits, several PDUs and servers were damaged. that was the cherry on top of the sundae though, there had been 2 power outages at that point and 2 other outages after that point.

Every Datacenter I've been in for more than a couple years, for the last 15 years has had a power outage (or two). AT&T, Qwest, Exodus, AIS (San Diego), Layer 42 (San Jose), Media Temple (Los Angeles). It's like cable-cuts on your circuit - they happen so reliably that you put these events into your business plan. If you need network redundancy, you always have two (diverse) circuits. If you need data center redundancy you always have two (diverse) data centers. These events happen so reliably that the surprise is when they _don't_ happen, not when they _do_ happen.

Plan for a minimum of one power outage every two-three years and you won't be disappointed.

I feel for the HE guys - back to back power outages has got to be killing them right now.

I host directly with HE in Fremont1 and just experienced the power outage. I've had equipment there for 4+ years and this is the first power issues I've had with them. HE isn't perfect, but up until now I've been perfectly happy there. Yes, HE runs a fairly relaxed data center there for better or worse. HE hasn't communicated about this outage--which I find very disappointing. I would guess this power outage was a result of attempting to fix whatever broke Saturday.

Here are my own availability stats for Linode this year. We use Panopta to monitor each data center. All are within the SLA:

Dallas - 99.951% Newark - 99.969% London - 99.986% Fremont - 99.989% Atlanta - 99.995%

Thanks for sharing. We're you able to identify with Linode the outages for each datacenter or do the stats include any false positives?

Linode's SLA is monthly, so they are under for the month in Fremont. Though they will, of course, give credit.

I found this a little disturbing (having just put a new app on a node there - fortunately pre-production), but then I remembered that the last two places I worked paid the premium to be hosted at 365 Main in SF, with its flywheels and diesel generators etc, and that didn't turn out to be magic:

http://www.365main.com/status_update.html

The moral of the story for me is;

* These things are complicated

* Failures will happen

* You have to be prepared to deal with them

I wouldn't host anything in Hurricane Electric @ Fremont for several reasons:

Back in 2008 an HE based colo "McColo" were shut down because they were hosting a HUGE amount of botnet controllers, spamming operations, and similar shady operations. When they were shut off some security firms saw a 50% drop in spam going through their firewalls.

However it only happened after immense pressure on HE and other providers involved by Google, Washington Post and all sorts of other players.

HE would have been aware because large IP blocks were being blacklisted (I heard at one point all of HE Freemont's IPs were blocked by some of the more extreme SBL lists) but they turned a blind eye and/or claimed Ts & Cs were not being infringed.

More here http://news.cnet.com/8301-1009_3-10095730-83.html

I found that highly irresponsible, both in terms of the detriment to their other colo customers who were sharing the BGP-level bandwidth but also from a wider 'being a good actor' perspective.

Given that they are also on a fault line and that good connectivity to Europe is more important to me than Asia, I would prefer to host on East Coast.

I'm actually in Linode's New Jersey and London Colos, and they are both excellent.

I'm not sure that is a bad thing, actually. Although I do not like spam and botnets, at least they are willing to stand up for their customers and require due process. Others will cave on bogus DMCA notices and assume that the customer is guilty.

I would actually call pulling the plug without due process irresponsible.

Also, I question your claim about it happening only after "immense pressure." Your own link praises them for their response, and some googling suggests similar wording in all coverage I can find.

Folks in their irc channel are saying that the problem is with Hurricane Electric.

I bought my first Linode on the morning of the first power outage. Does this kind of thing happen regularly there? I'm having regrets...

This is the exception for sure. Word is that the problem is with the datacenter their boxes are in (Hurricane Electric). Of course, ultimately this could cause customer loss for them, so it becomes their issue.

I've had Linodes in Newark, NJ for almost a year now and have not had any issues there at all

I signed up in September after a year of horrible problems at vps.net and I've had no issues so far, my personal rule (assuming this isn't mission critical servers) is to wait 1 month before I make a judgement, having a problem on the first day doesn't necessarily define the ongoing experience.

I host there since jan/09 and never had problems in Atlanta.

It's been ages since i can recall problems in their New Jersey data center.

This week is the first time in the year that I've owned mine that it has been down.

Never had any problems in london.

Resolved after 44 minutes of downtime (according to http://wasitup.com) for my server.

Posting a reference to the previous outage thread here as it's the only place I could track down their support IRC info. (Might need to dig deeper in the Linode support docs)...

http://news.ycombinator.com/item?id=1926368

Actual IRC info...

Server: irc.oftc.net Channel: #linode

These recent outages have been giving me a lot of stress lately. I manage 7 Linodes at the Fremont data center. Fortunately these outages have been at non-peak times, but if this happens during a peak time. I will need to rethink my server architecture.

Out of curiosity, I am wondering why someone would choose Linode over AWS?

Nuanced, but different products.

AWS is a cloud product, with pros and cons -- instances can die at any time and data (RAM and disk) won't persist, network is slightly strange with their NAT IPs but in return you get a setup that lets you connect with big storage (S3), potentially large clusters (the new GPU core product) etc.

Linode is a VPS which is just presenting you with an abstracted server. If the instance or the hypervisor gets bounced anything on disk is preserved. Networking is normal (standard IP address) and everything runs as if it is a bare metal server (more true for XEN based VPS's like Linode rather than SolusVM)

Performance and value.

I live down the street from this datacenter and my home experienced this same power outage. Makes me wonder how they handle power failures. My guess is that they don't.

Yikes, makes me happy I chose atlanta for my project. I kind of figured it would be a good idea to have my server a long way away from California, for many reasons.

I would really like a way to move my linode out of the Freemont data center. My linode in Newark, New Jersey has been perfect for three years.

You can, just open a support ticket requesting a migration. They'll queue it up for you, then you click a button to migrate at your leisure.

I noticed two network outages in the Newark data center over the last year. I moved to the London facility in May and have had 100% uptime there since.

I was under the impression you can migrate Linodes between data centres?

Linode's SLA: http://www.linode.com/faq.cfm#what-is-your-sla

" 99.9% uptime, or your lost time is refunded back to your account" - unless my math is wrong that's 7.2 hours a month (for a 30 day month). Most SLA's are terrifying if you do the math.

Your math is wrong--there are about 730 hours in a month so 99.9% update means it would need to be up 729.27 hours or no more than ~43 minutes of downtime. However the proportional refund here is so small to barely be noticeable. Pocket change.

Agreed, also the refund of lost time, which is rather terrifying too, e.g. less than $1 for a whole day outage if you are using the $20 monthly plan.

It's 0.72 hours a month.

Right. This is a nice little problem that can be easily checked by Google:

Search: 0.1% of 1 month in hours

Result: 0.1% of (1 month) = 0.730484398 hours

that sounds much better.

0.72 hours (45 minutes) per month.

to say this is annoying is an understatement

Reply on news.ycombinator.com