> This frequently trips people up. Cloud instances (typically) have far lower availability than physical servers. They also have no (cheap) persistence across VM evictions.

This may be true for AWS but (in my experience since 2014) it is not true for Google compute engine. On GCE, VM instances automatically migrate when the host is rebooted, and VM availability and uptime is impressive. Also, on GCE, the default recommended disks are all persistent, since they are block devices mounted over the network; this wasn't the case years ago, but is now, and it is very impressive.

Disclaimer: my startup is in the Google startup program.

Yep. GCE VM migration and persistent storage is really amazing, and getting better with every release, however there are still limitations and caveats that can get you in trouble (especially if you have very tight latency requirements.)

Glad you like it! Good luck with your startup.

Disclaimer: I work on infrastructure systems at Google :-)

> Cloud instances (typically) have far lower availability than physical servers.

This is common belief that is just plain wrong.

Servers can stay online for years without issues, same things on AWS and physical servers.

> They also have no (cheap) persistence across VM evictions.

I don't even understand that this sentence mean. You can start by explaining what is "VM eviction"?

Whatever it is. Physical servers have no soft of failover or persistence for when one server suddenly dies.

People started migrating from physical servers to virtualized more than 10 years ago (i.e. VmWare). AWS/GCE are not fundamentally different.

> This is common belief that is just plain wrong.

Sorry, you'll need to back that up.

Evictions are when your VM is booted off a physical machine to a different one. This usually happens when something in the the underlying infrastructure is undergoing maintenance (e.g., networking, power, kernel, physical machine, etc.)

VMs have to deal with servers dying plus evictions plus control and management plane failures, which significantly reduces the availability of a single instance. So strictly speaking, cloud instances typically have lower availability than physical machines.

> Physical servers have no soft of failover or persistence for when one server suddenly dies.

Servers suddenly dying is not the failure mode that requires persistence -- it's evictions. This is why cloud services tend to prefer external distributed storage systems to local disk, and a multitude of ways to quickly and automatically reinstall machines on instance turnup.

BTW, GCE recently started supporting VM migration (which you can opt into), where they attempt to migrate your instance to another physical machine without losing any state. There are still limitations to this, but for a lot of simple cases can work well.

It doesn't matter whether the application is abstracted on bare metal, on VmWare or on AWS.

There's still physical hardware down the chain that calls for maintenance. When the physical hardware require maintenance everything on top is fucked or need to be moved.

For planned maintenance:

- If you're bare metal. You cannot migrate to a working server.

- If you're virtualized, you have to stop the service and restart on another host.

- If you're virtualized on GCE or vSphere datacenter, they'll migrate the VM live while it's running.

GCE has supported live migrations for multiple years. VmWare supports that for about 10 years (yep, no kidding ^^)

> VMs have to deal with servers dying plus evictions plus control and management plane failures, which significantly reduces the availability of a single instance. So strictly speaking, cloud instances typically have lower availability than physical machines.

Bare metal servers have to deal with hardware dying plus evictions plus control and management plane failures, which significantly reduces the availability of a single server.

So strictly speaking, cloud instances and physical servers are the same thing. They both live on physical hardware.

What do you mean recently? We've been doing this for about three years, and it was part of declaring Compute Engine as generally available.

As for "limitations", I mean we can't save you if the NIC or ToR fails, but we lose VMs really rarely. Feel free to ping me internally, and I'll point you at our dashboards.

Disclosure: I work on Google Cloud.

In fairness, I could launch EC2 instances as a whole about 10 years ago so the whole of GCP is recent. (Good job on it though, quality is a differentiator)

Agreed! My issue was which word "recently" bound to (GCE or "the universe"). I interpreted it as GCE.

Thanks, yes I know :-) I consider three years to be pretty recent (from the perspective of people managing their own servers.)

> Servers can stay online for years without issues, same things on AWS and physical servers

Why they don't back it up SLA promise then? Having a VM does not guarantee increased uptime. On premises virtualization can guarantee it,cloud cannot.

"Servers can stay online for years without issues, same things on AWS and physical servers."

This is mostly luck of the draw, I typically have seen one VM failure out of 50 every 5-6 months on AWS. Ymmv.

Of course with EBS that usually just means "restart it".