Things like throwing another node into the cluster, or rolling updates are free, which you would otherwise need to develop yourself. All of that is totally doable, of course, but I like being able to lean on tooling that is not custom, when possible.

When your infrastructure does need to become more complicated, you're already ready for it. Even if I were only serving a single language, starting with a K8s stack makes a lot of sense, to me, from a tooling perspective. Yeah normal VMs might be simpler, conceptually, but I don't consider K8s terribly complicated from a user perspective, when you're staying around the lanes they intend you to stay in. Part of this may also be my having worked with pretty poor ops teams in the past, but I think K8s gives you a really good baseline that gives pretty good defaults about a lot of your infrastructure, without a lot of investment on your part.

That said, if you're managing it on a bare metal server, then VMs may be much easier for you. K8s The Hard Way and similar guides go into how that would work, but managing high availability etcd servers and the like is a bit outside my comfort zone. YMMV.

Most monoliths software companies build aren't actually monoliths, conceptually. Let's say you integrate with the Facebook API to pull some user data. Facebook is, within the conceptual model of your application, a service. Hell, you even have to worry "a little bit" about maintaining it; provisioning and rotating API keys, possibly paying for it, keeping up to date on deprecations, writing code to wire it up, worrying about network faults and uptime... That sounds like a service to me; we're three steps short of a true in-house service, as you don't have to worry about writing its code and actually running it, but conceptually its strikingly similar.

Facebook is a bad example here. Let's talk Authentication. Its a natural "first demonolithized service" that many companies will reach to build. Auth0, Okta, etc will sell you a SaaS product, or you can build your own with many freely available libraries. Conceptually they fill the same role in your application.

Let's say you use Postgres. That's pretty much a service in your application. A-ha; that's a cool monolith you've got there, already communicating over a network ain't it. Got a redis cache? Elasticsearch? Nginx proxy? Load balancer? Central logging and monitoring? Uh oh, this isn't really looking like a monolith anymore is it? You wanted it to be a monolith, but you've already got a few networked services. Whoops.

"Service-oriented" isn't first-and-foremost a way of building your application. It's a way of thinking about your architecture. It means things like decoupling, gracefully handling network failures, scaling out instead of up, etc. All of these concepts apply whether you're building a dozen services or you're buying a dozen services.

Monolithic architectures are old news because of this recognition; no one builds monoliths anymore. It's arguable if anyone ever did, truly. We all depend on networked services, many that other people provide. The sooner you think in terms of networked services, the sooner your application will be more reliable and offer a superior experience to customers.

And then, it's a natural step to building some in-house. I am staunchly in the camp of "'monolith' first, with the intention of going into services" because it forces you to start thinking about these big networking problems early. You can't avoid it.

The key issue here is that k8s was written with very large goals in mind. That a small business can easily spin it up quickly and run a few microservices or even a monolith + some workers is just coincidental. It is NOT the design goal. And the result of that is that a lot of the tooling and writing around k8s reflects that. A lot of the advice around practices like observability and service meshes comes from people who've worked in the top 1% (or less) of companies in terms of computing complexity. What I'm personally seeing is that this advice is starting to trickle down into the mainstream as gospel. Which strangely makes sense. No one else has the ability to preach with such assurance because not many people in small companies have actually been in the scenarios of the big guns. The only problem is that it's gospel without considering context.

So at what point does k8s make sense? Only when you have answers to the following:

* Getting started is easy, maintaining and keeping up with the going ons is a full time job - Do you have at least 1 engineer at least that you can spare to work on maintaining k8s as their primary job? It doesn't mean full time. But if they have to drop everything else to go work on k8s and investigate strange I/O performance issues, are you ready to allow that?

* The k8s eco system is like the JS framework ecosystem right now - There are no set ways of doing anything. You want to do CICD? Should you use helm charts? Helm charts inherited from a chart folder? Or are you fine using the PATCH API/kubectl patch commands to upgrade deployments. Who's going to maintain the pipeline? Who's going to write the custom code for your github deployments or your brigade scripts or your custom in house tool? Who's going to think about securing this stuff and the UX around it. That's just CICD mind you. We aren't anywhere close to the weeds about deciding if you want to use ingresses vs Load balancers and how you are going to run into service provider limits on certain resources. Are you ready to have at minimum 1 developer working on this stuff and taking time to talk to the team about it?

* Speaking about the team, k8s and Docker in general is a shift in thinking - This might sound surprising but the fact that Jessie Frazelle (y'all should all follow her btw) is occasionally seen reiterating the point that containers are NOT VM's is a decent indicator that people don't understand k8s or Docker at a conceptual level. When you adopt k8s, you are going to pass that complexity to your developers at some point. Either that or your dev ops team takes on that full complexity and that's a fair amount to abstract away from the developers which will likely increase the work load of devops and/or their team size. Are you prepared for either path?

* Oh also, what do your development environments start to look like? This is partly related to microservices but are you dockerizing your applications to work on the local dev environment? Who's responsible for that transition? As much as one tries to resist it, once you are on k8s you'll want to take advantage of it. Someone will build a small thing as a microservice or a worker that the monolith or other services depend on. How are you going to set that up locally? And again, who's going to help the devs accumulate that knowledge while they are busy trying to build the product. (Please don't put your hopes on devs wanting to learn that after hours. That's just cruel).

I can't write everything else I have in mind on this topic. It'd go on for a long long time. But the common theme here is that the choice around adopting k8s is generally put on a table of technical pros and cons. I'd argue that there's a significant hidden cost of human impact as well. Not all these decisions are upfront but it is the pain that you will adopt and have to decide on at some point.

Again, at what point does k8s make sense? Like I said, you ideally should be paining before you start to consider k8s because for nearly every feature of k8s, there is a well documented, well established, well secured parallel that already exists in the myriad of service providers. It's a matter of taking careful stock of how much upfront pain you are trading away for pain that you WILL accumulate later.

PS - If anyone claims that adopting a newer technology is going to make things outright less painful , that's a good sign of immaturity. I've been there and I picture myself smashing my head into a table every now and then when I think of how immature I used to be. Apologies to people I've worked with at past jobs.

PPS - From the k8s site, "Designed on the same principles that allows Google to run billions of containers a week, Kubernetes can scale without increasing your ops team." <-- is the kind of claim that we need to take flamethrowers to. On paper, 1 dev with the kubectl+kops CLI can scale services to run with 1000's of nodes and millions of containers. But realistically, you don't get there without having incurred significantly more complex use cases. So no, nothing scales independently.

Also, migrating to microservices for existing services might not be worth it, especially if you don't operate at a massive scale.

Keep it simple stupid is still a solid design decision, despite all the microservice/container hype.

Most bussinesses only need a couple of servers that provide the service, spread redundantly with a HA capability.

1) For three whole days, it was questionable whether or not a user would be able to launch a node pool (according to the official blog statement). It was also questionable whether a user would be able to launch a simple compute instance (according to statements here on HN).

2) This issue was global in scope, affecting all of Google's regions. Therefore, in consideration of item 1 above, it was questionable/unpredictable whether or not a user could launch a node pool or even a simple node anywhere in GCP at all.

3) The sum total of information about this incident can be found as a few one or two sentence blurbs on Google's blog. No explanation nor outline of scope for affected regions and services has been provided.

4) Some users here are reporting that other GCP services not mentioned by Google's blog are experiencing problems.

5) Some users here are reporting that they have received no response from GCP support, even over a time span of 40+ hours since the support request was submitted.

6) Google says they'll provide some information when the next business day rolls around, roughly 4 days after the start of the problem.

I really do want to make sure I'm understanding this situation. Please do correct me if I got something wrong in this summary.

- Someone at Google right now, probably.

It looks like the UI issue was actually fixed, and that we just didn't update the status dashboard correctly. But we're double checking that and looking into some of the additional things you all have reported here.

Why do you guys suffer global outages? This is your 2nd major global outage in less than 5 years. I’m sorry to say this, but it is the equivalent of going bankrupt from a trust perspective. I need to see some blog posts about how you guys are rethinking whatever design can lead to this - twice - or you are never getting a cent of money under my control. You have the most feature rich cloud (particularly your networking products), but down time like this is unacceptable.

No one would ever ask why you chose AWS. The old “no one ever got fired for buying IBM”.

Even if you chose Azure because you’re a Microsoft shop, no one would question your choice of MS. Besides, MS is known for thier enterprise support.

From a developer/architect standpoint, I’ve been focused the last year on learning everything I could about AWS and chose a company that fully embraced it. AWS experience is much more marketable than GCP. It’s more popular than Azure too, but there are plenty of MS shops around that are using Azure.

>We will provide more information by Monday, 2018-11-12 11:00 US/Pacific.

Wait, did the people tasked with fixing this just take the weekend off?

What would a small business do as a contingency plan?

One thing I do care about though, is root cause analysis. I love reading a good RCA, it restores my faith in the company and makes me trust them more.

(I'm not affect by the GKE outage so opinions may differ right now!)

The timeline of this disruption matches when we started experiencing cloud build errors.

https://status.cloud.google.com/incident/container-engine/18...

Yet, somehow every major cloud provider experiences global outages.

That old AWS S3 outage in us-east-1 was an interesting one; when it went down, many services which rely on S3 also went down, in other regions beside us-east-1 because they were using us-east-1 buckets. I have a feeling this is more common than you'd think; globally-redundant services which rely on some single point of geographical failure for some small part.

Always saying resource not available. My account is a pretty new account.

In contrast, one of my friend is having a pretty old account which is very active. He has no such issue.

So I think due to this issue, Google has enabled some resource limitation for new accounts.

But they should properly communicate this issue.

An instance in us-central1-a has refused to start since last Thursday or Friday.

I created a new instance in us-west2-c, which worked briefly but began to fail midday Friday, and kept failing through the weekend.

On Saturday I created yet another clone in northamerica-northeast1-b. That worked Saturday and Sunday, but this morning, it is failing to start. Fortunately my us-west2-c instance has begun to work again, but I'm having doubts about continuing to use GCE as we scale up.

And yet, the status page says all services are available.

Is the typical of others' experiences?

Perhaps if you explained it on a whiteboard...

From my personal experience - i think all big cloud providers first two level support staff is no good if it isn't an obvious dumb one on your part. I always prefer to forgo support and try to go through every bit of their documentation to figure it out on our own. This helps to save huge amount of time. But if you have developer support - it can help to expedite things little faster though.

That's my favorite.

As another comment pointed out, what's the point of having so many zones and redundancy around the globe if such global failure can still happen? I thought the "cloud" was supposed to make this kind of failure impossible

I created a new instance in us-west2-c, which worked briefly but began to fail midday Friday, and kept failing through the weekend.

On Saturday I created yet another clone in northamerica-northeast1-b. That worked Saturday and Sunday, but this morning, it is failing to start. Fortunately my us-west2-c instance has begun to work again, but I'm having doubts about continuing to use GCE as we scale up.

And yet, the status page says all services are available.

The problem is that running a global SDN like this means if you do something wrong, you can have outages that impact multiple regions simultaneously.

This is why AWS has strict regional isolation and will never create cross-region dependencies (outside of some truly global services like IAM and Route 53 that have sufficient redundancy that they should (hopefully) never go down).

Disclaimer: I work for AWS, but my opinions are my own.

Disclaimer: google employee in ads, who worked on many many fires throughout the years, but talking from my personal perspective and not from my employer. I am sure we are striving to have 0, but realistically, i have seen many that says things happen. Learn, and improve.

Most of what you can read of Google's approach will teach you their ideal computing environment is a single planetary resource, pushing any natural segmentation and partitioning out of view.

It's the opposite really: the expectation that service providers have no unexpected downtime is unrealistic, and it's strange this idea persists.

5. Years.

Nothing to see here, move along.

- Security posture. Project Zero is class leading, and there's absolutely a "fear-based" component there, with the open question of when Project Zero discovers a new exploit, who will they share it with before going public? The upcoming Security Command Center product looks miles ahead of the disparate and poorly integrated solutions AWS or Azure offers.

- Cost. Apples to apples, GCP is cheaper than any other cloud platform. Combine that with easy-to-use models like preemptible instances which can reduce costs further; deploying a similar strategy to AWS takes substantially more engineering effort.

- Class leading software talent. Google is proven to be on the forefront of new CS research, then pivoting that into products that software companies depend on; you can look all the way back to BigQuery, their AI work, or more recently in Spanner or Kubernetes.

- GKE. Its miles ahead of the competition. If you're on Kubernetes and its not on GKE, then you've got legacy reasons for being where you're at.

Plenty of great reasons. Reliability is just one factor in the equation, and GCP definitely isn't that far behind AWS. We have really short memories as humans, but too soon we seem to forget Azure's global outage just a couple months ago due to a weather issue at one datacenter, or AWS's massive us-east-1 S3 outage caused by a human incorrectly entering a command. Shit happens, and it's alright. As humans, we're all learning, and as long as we learn from this and we get better then that's what matters.

GCP can be a fair bit cheaper than AWS and Azure for certain workloads. Raw compute/memory is about the same. Storage can make a big difference. GCP persistent SSD costs a bit more than AWS GP2 with much better performance and way cheaper than IO2. Local SSD is also way, way cheaper than I2 instances.

Most folks deploying distributed data stores that need guaranteed performance are using local disk, so this can be a really big deal.

Going multi region on AWS should be safe enough.

If a multi region, multi service meltdown happens on AWS, it will feel like most of the internet has gone down to a lot of users. Being such a catastrophic failure, I bet the service will be restored pretty fast, not in 3 days.

You could go multi cloud though. But when half of the internet struggles to work correctly, I’d not feel too bad about my small business’ downtime.

Then start looking at points of failure and sort them based on severity and probability. Is your own software deployment going to generate more downtime per year than a regional aws outage?

There are formal academic ways to determine what your overall availability is, but don't have those on hand. Suffice to say, it takes significant research, planning, execution, and testing to ensure a target availability. (See Netflix https://medium.com/netflix-techblog/the-netflix-simian-army-... ) if someone says they have 99.9% or better up time, they had better have proof in my mind (or a fat SLA violation payout)

People outsource to cloud providers not because they are cheap, but because managing infra in house is hard. Also move fast and break things.

Read AWS docs about availability, there are availability zones in a region, spread across those to minimize impact. Then test when something goes down. Fix/repeat.

Ultimately it’s a risk/return decision.

“Is going exclusively with AWS/azure/GCP etc a better decision in reliability, financial and mantainability terms than complicating the design to improve resiliency? And will this more complex solution actually improve reliability?”

If AWS ever screws up, you will be able to continue running the business even if it might take weeks to start over.

For live redundancy, you should have a secondary datacenter on another provider, but realistically it's hard to do and most business never achieve that. Instead, just stick with AWS and if there is a problem the strategy is to sip coffee while waiting for them to resolve it. Much better this way than you having to fix it yourself.

Depends on your definition of small. If it's small enough not to have a dedicated infrastructure team designing multicloud solution, then the contingency plan may be: switch DNS to a static site saying "we're down until AWS fixes the issue, check back later".

Otherwise it depends on your specific scenario, your support contracts, and lots of other things. You need to decide what matters, how much the mitigation costs vs downtime, and go from there.

I wish I was only being tongue-in-cheek.

Terraform using AMIs plus chef recipes that work in the cloud and bare metal. Dont use AWS specific services.

This would allow you to spin over to another cloud provider , vsphere or bare metal with minimal work

Some services in AWS are also not available in others (I'm quite familiar with AWS Data Pipeline not being available outside the "core" regions like us-east-1, eu-west-1) and having services in one region make usage of resources in another region is a huge change when most developers outside ones with technology literate customers are under the gun to push features out fast over sound design. The matrix of services and configurations necessary to mix and match regions and availability zones is non-trivial if you make extensive usage of AWS services above the IAAS layer.

Also, cross-region VPC peering has a TON of limitations that rather annoying depending upon how well your network has been architected (by default in most companies outside enterprises with a deep bench of network engineers, this would be rated at "complete crap barely better than a typical home wifi network"). Heck, even though I'm non-dumb at networks I have to keep reminding myself of various cross-region VPC limitations when working with refactoring cross-region VPCs like where you can reference security groups, how to propagate Route 53 records, etc.

It might be, until they jack up the prices 15X with limited notice (looking at you, Google maps [1]). No deprecation needed, just force users off the platform unless they're willing to pay a massive premium.

[1] https://www.google.com/search?q=google+maps+price+increase

If they ever do deprecate something people have built on though they're gonna get absolutely crucified. That's probably better protection than any terms of service.