Sending Excess Load To the Cloud?

TristanBrotherton writes "Cloud computing seems to be a good choice for startups like ours, looking to scale easily with users. (We're providing a series of Web services, assets, and Web applications to users of our mobile client.) There are the obvious choices of Google, Amazon, and smaller shops like EngineYard. The biggest issue we have in choosing cloud computing to run our applications is trust in their robustness. If the provider goes down, we suffer. In traditional hosting environments we mitigate this with multiple sites / vendors. It's not really feasible to host on multiple compute services, so I wondered if a better option might be to set up a small (perhaps two servers) origin infrastructure in a traditional manner at a datacenter, running our applications, but then send excess load, or in the event of our origin servers failing, all load, to compute services. This would give us the best of both worlds. Has anyone done this, or had experience in designing Web applications to scale seamlessly across both environments? Is there particular load-balancing hardware we can use to do this?"

Comments

Will "the cloud" be there when you need it?

[Animats]

Unless your "cloud" provider offers a service level guarantee with teeth, is contractually obligated to continue to provide the service for some period of time, and has sound financial fundamentals, this is risky.

I think we'll see a big shutdown of money-losing web services over the next year.

Depends on what you are doing

[mcrbids]

If you are running a web-based, hosted financial application, outsourcing "to the cloud" is a non-starter. If you are hosting pictures of kitty cats, the cloud can be an excellent resource.

Throw a server up, upload some files, start up a PG or MySQL database, and integrity is easy. But as soon as you introduce the 2nd system, integrity issues start jumping out of the woodwork. It gets worse with each additional node. Redundancy isn't just fancy-sounding, it's damned hard to do right, and as soon as you introduce it, you have to accept an elevated error rate because the number of things that can go wrong go UP, even as the number of catastrophic system failures drop.

For a great example of redundancy in action, take a look in the mirror. You have individual cells dying by the millions every minute. Your memory is fuzzy at best, your pattern-recognition in your brain frequently sees things that aren't there, and you make stupid mistakes every single day. And that's fine, because the overall system is pretty damned redundant and resilient. A mash of protein goo and calcium deposits able to sustain one of the most complex information systems around, reliably, 24x7, for an average of 70 years or so apiece.

Good luck getting any kind of hosting platform to maintain that kind of uptime, no matter the expense! But in biology, minor errors are so commonplace that they are hard to catalogue, let alone count.

So pick your battle, and realize that high-performance, high-redundancy clustering is very, very difficult to do well.

In the meantime, spend money on good quality hardware, and use top-notch colo hosting. The cost of doing it right is actually significantly lower than doing it "on the cheap" so spend money where it counts (good quality infrastructure) and save where it matters. (EG: public opinion) It's almost odd - if you look for the very, very best colo, regardless of cost, you'll find that their monetary cost is probably one of the lower ones around. (head scratcher) I've found this to be rather consistent with several reviews under my belt.

Also, I find it best to use whitebox systems with midrange hardware. These are quality, high-performance hardware developed with everything but the name brand. In my case, I've standardized on 1U multicore X86/64 systems with hot-swap, high performance 15k SCSI drives put out by Tyan and SuperMicro. There are a large number of dealers of such systems, my current favorite is Aberdeen Inc. They can sell you an amazing amount of performance and reliability for around $2500.

This is the stuff that Sun will sell you for $8,000/pop. They will stand up to day-in, day-out heavy use for years, with hundreds or thousands of users every day, millions of website hits per day, etc. They are high performance. This is quality hardware. And with the money you save, you can have an immediate hot backup for less than the cost of the "premium support" of the big guys, and more redundancy in the meantime.

My $0.02. Since it's free advice, you're free to use it as you see fit!

Re:The C word

[MrNaz]

It's a term invented by idiot managers who saw all those diagrams where the wider internet is represented by a picture of a cloud and were too stupid to grasp the concept of a representative diagram, so they took the picture of the cloud to be literal, and now there is an entire generation of managers who have an image of electrons flying around the sky. They confusion they suffer is only exacerbated when there's a thunderstorm and they hear the word "torrent" to describe the rain, thinking that the storm is the result of those damned P2P users.