Are Cloud servers less likely to fail than dedicated?

@amityweb:

Something I have always been unsure about is exactly what hardware "cloud" servers provide. e.g. do I have a slice on a dedicated box, or is it some magical distribution over several machines?

It can be both. In linode's case, it's the former. In others (eg Amazon ECS), it's the latter.

@amityweb:

Am I right or wrong, is my Linode hard drive as likely to fail Or ds any VPS and dedicated machine?

It depends on the hardware, but it's probably similar likelyhood.

@amityweb:

So in this event I need to do a restore. Does it use mirrored RAID so actually is better than my dedicated without RAID? In which case I assume I would restore when both hard drives fail the same time? I assume Linode fix one of the hard drives in a mirrored RAID if it fails and I wont need to do anything?

Yes, linode uses RAID 10, so if a single drive fails, you won't even know about it. linode would replace the drive and everything hums along normally. If enough/the right drives fail in short enough order, the RAID set could be lost. Not likely, but it can happen. So, you still need backups, but the situation is certainly better than a dedicated server without RAID.

Cloud doesn't mean running one operating system over multiple physical machines. Cloud just means don't worry about the details, they are not your problem.

Linode runs many virtual machines on each physical host. They mirror the disks so it's highly unlikely you will lose any data. When disks break they replace them. On good hardware you can just pull out the broken disk, plug in the new disk, and the server will mirror the data back without the operating system even noticing anything happened.

A Linode machine is certainly better than a dedicated server with an unmirrored disk but disks are such unreliable things that there is always some risk of data loss. You should always keep backups just in case. And test the backups too.

@glg:

@amityweb:

Something I have always been unsure about is exactly what hardware "cloud" servers provide. e.g. do I have a slice on a dedicated box, or is it some magical distribution over several machines?

It can be both. In linode's case, it's the former. In others (eg Amazon ECS), it's the latter.

Amazon doesn't offer one operating system running across multiple physical machines, neither Linux nor Windows are capable of that. It offers one operating system or more per physical machine with the flexibility to easily create very many virtual machines on the same or other physical hosts.

Adding more points of failure NEVER makes a system more reliable.

The "cloud" is nothing more then marketing hype. Somehow the marketing fops thought it would be an easier sell to put it "in the cloud" then telling people - "yes, all of your stuff will be running on someone elses equipment on someone elses network and completely under someone elses control".

And big surprise - not, people are dumb enough to buy into it.

Right OK thanks a lot. So I guess there isnt anything special about it being a "cloud" service? Just scaleable (although the restart puts me off upgrading). I guess they are like any VPS/dedicated but just using RAID mirrors so if I had that on my other VPS/Dedicated any issues I have with hard drive failures are just as likely regardless (but probably unlikely any issues would impact on me and my sites if using RAID so maybe not to worry about that so much)

To muddy the discussion by introducing a precise definition, cloud computing has five essential characteristics (per NIST Special Publication 800-145):

• On-demand self service: You can add/remove services without human involvement

• Broad network access: You can access it over a broad network (e.g. the Internet)

• Resource pooling: Physical/virtual resources are divvied up among multiple customers

• Rapid elasticity: You can quickly add/remove capacity to meet demand

• Measured service: You pay for what you use

The traditional VPS model meets these requirements, and some of the more sophisticated dedicated server hosts probably do too. So, there's nothing particularly special about "cloud computing," other than its embrace of rapid change. There is no requirement for persistent storage… indeed, Amazon EC2 does not have persistent storage without EBS and/or S3. And if you only have a handful of server instances, you won't necessarily reap the benefits of cloud computing. Adding more servers is the "cloud way" to scale up, not resizing your existing ones. Also, if one of your cloud servers dies for whatever reason (hardware failure, software crash, etc), you deploy a new one and life goes on. The reliability of each component is less important than the reliability of the whole.

So, you probably want a VPS, not a self-healing cluster of cloud servers. That's fine.

Back in the day, it was not unheard of for a few friends to chip in on co-locating a decent server in a datacenter somewhere. This meant they could afford nicer hardware and better connectivity than any of them could individually, in exchange for not having exclusive use of the hardware. This arrangement evolved over the years to become the VPS model: multiple tenants on one physical machine, with better hardware than any of them could afford individually. This, in my opinion, is the major benefit of VPSes over dedicated servers. A well-connected and reliably-powered multi-core rackmount server with four high-speed disks in a RAID 10 configuration is not cheap, and it is probably much more power than you need.

So, with dedicated servers, you quickly whittle that down to a poorly-connected, unreliably-powered single-core desktop machine with a single disk drive to save money. With VPSes, you still have the reliable, redundant, and crazy powerful server, but you have dozens of tenants sharing it. You also have a layer of abstraction that's difficult to do with a dedicated server: self-provisioning, all the OS reinstalls and reboots you could want, ssh-based console, automatic online off-server backups, and you never have to deal with "remote hands." The tradeoff is that my $20/mo Linode has less disk space and RAM than my cellphone… but my cellphone doesn't have fifty people sharing it.

..

Also, as a standard disclaimer: RAID IS NOT A BACKUP. :-)

To confuse the situation slightly further, there are some advantages to running in a VPS farm, if that farm is designed to handle things. For example, a VPS farm backed by SAN storage (as opposed to local disk) can be more resilliant - if the server you're node runs on dies then the instance can immediately be brought up on another machine because the storage isn't local to the machine. (Linode can't do this; their storage is local). It's even possible (eg ESX with vmotion) to migrate a running instance from one node to another with no downtime.

Obviously this isn't 100% reliable; then SAN could still explode, for example :-) But it does show how a VPS farm can be better than physical machines. Of course you need to contact your VPS provider (or potential provider) to learn how they do things to determine if you get any of these advantages.

@sednet:

@glg:

@amityweb:

Something I have always been unsure about is exactly what hardware "cloud" servers provide. e.g. do I have a slice on a dedicated box, or is it some magical distribution over several machines?

It can be both. In linode's case, it's the former. In others (eg Amazon ECS), it's the latter.

Amazon doesn't offer one operating system running across multiple physical machines, neither Linux nor Windows are capable of that. It offers one operating system or more per physical machine with the flexibility to easily create very many virtual machines on the same or other physical hosts.

That's not what I said. Yes, each VM must run on a single physical machine, but it needn't be tied to a single machine like linode does. One can use a clustering solution that has pools of VM's running on pools of physical machines, including (as sweh mentions) vmotion with live migration to a different physical machine in the cluster.

It's worth noting that when shared storage systems (SANs, Amazon EBS, etc) fail, it's really really bad. A sizable portion of Netflix and Reddit outages can be traced to such a storage failure…

Given how rarely I'd benefit from live migration of my Linodes (approx. once every few years), it'd be like amputation to solve a headache.

@glg:

One can use a clustering solution that has pools of VM's running on pools of physical machines, including (as sweh mentions) vmotion with live migration to a different physical machine in the cluster.

Amazon does nothing of the sort (vMotion would imply VMware, and Amazon runs only Xen and an alternative for specific scenarios; they also do not do any kind of hot migration).

The reason Amazon instances are able to move about at whim is due to most instances being EBS-rooted, which can follow you between host machines. Indeed, when an instance is stopped, it is very likely rehomed as the "ephemeral" storage – which I understand to be onboard, like Linode's primary storage -- is immediately wiped on stop.

That only happens on instance stop, however, and notably not reboot.

@jed:

@glg:

One can use a clustering solution that has pools of VM's running on pools of physical machines, including (as sweh mentions) vmotion with live migration to a different physical machine in the cluster.

Amazon does nothing of the sort (vMotion would imply VMware, and Amazon runs only Xen and an alternative for specific scenarios; they also do not do any kind of hot migration).

I never said they did hot migration, only that they don't dedicate your VM to a specific physical host like linode does.