I've seen so many projects bending over backwards to avoid k8s and pay large cloud bills to avoid it at all costs. (ECS and app services are hopelessly expensive and bothersome)
K8s is really good, pretty easy to maintain, but a bit hard to understand. Mostly because distributed, zero down time systems are a bit hard to get by nature. But if you have someone that wants to take it on I've managed k8s clusters, solo, without incident, while doing lots of other stuff too (working with larger teams now though). Not to mention there's a lot of competence out there that can take over if I'd move on. Most of the deep complexity comes with more advance use cases, that wont show up for smaller deployments.
That said, no h8 towards going your own way! If your a solo developer (or small team) for a smallish project, don't feel the absolute need. If you get to the point you need it you should be earning enough to start paying someone to help ya get your app to a distributed system like k8s.
I think it's good to invest the time in understanding k8s though as a professional. Even if you won't directly run it it teaches you a lot about how to think about distributed, zero downtime systems. And what requirements that puts on an app.
1. Retrying a non-idempotent request on a failure type that does indicate that no action was taken is not necessarily safe.
2. It’s possible and actually fairly common to design a backend that can do a clean shutdown: it stops accepting new requests, completes old requests, then exits. I sincerely hope that Docker’s tooling is good enough for a service to unregister itself before it actually stops accepting requests, but I’m not actually very familiar with using Docker to manage HTTP routing. (I use a home grown tool that is far simpler.)
https://notnotp.com/notes/what-job-interviews-taught-me-abou...
I still think almost all of my complaints were valid: - not just container orchestration, more like a whole AWS solution (its own routing)
- has single-points of failure that would be avoided if you just used aws
- unnecessary complexity for almost all use cases
- ecosystem easily allows you to 2x-8x that complexity (helm+argocd+istio+carpenter+dozens more=hundreds of new failure modes)
- fundamentally moves many teams to a "I don't understand why my service is crashing, so let's just bring up more nodes every time rather than ever learning to debug it" mentality.
- All the feature teams that are supposed to "own their own kubernetes implementation for their own services" never do.
Of course a lot of that frustration is more at working at a place with a poorly-operated SOA where it takes a half-dozen services to send an email to a client or send a text or something silly. It sure is a waste of runway to obsess over a SOA when you aren't profitable yet.
But at this point it's sort of a sunk cost because it's become the industry standard. And AI can help with 90% of the complexity, which is it's own yellow-flag, but here we are.
The reason my position has changed is because:
1. The tooling has gotten better for setting up and managing K8.
2. In two of the last 3 jobs where we opted for a simplified alternative to k8, we came to regret that decision within a couple of years of that decision being made. If you’re core architecture is changing on a timescale of months (not years) then you picked the wrong foundations to build from.
That all said, I still think there is a pragmatic decision that needs to be made. And if I were in the author of this articles position I probably wouldn’t have picked k8s for this task either, despite what I said above. But, and as I said in my comment dismissing this article, they are dealing with low traffic and none of the problems that lend themselves to the benefits of k8. So my criticism of this article is that it’s misleading because their problem is easy but they’re writing as if they’re having to deal With problems of scale when they’re actually not.
But yeah, pretty cool DNS resolving features in HAProxy, that's nifty
That isn’t a lot. You could easily run that from one host. The reason people reach for Kubernetes (and similar) is because they need to scale past that single host dependency.
It's great.
They’re multi-region, but that doesn’t mean they’re running across multiple hosts in each region.
Docker compose doesn’t support pooling multiple hosts, so if they are running multiple hosts per region then there’s a lot more complexity to their setup than they’re documenting in that blog. Even if that complexity is human toil managing each host as a separate entity.
I have some stuff on single-node k3s. Because it's standard so I don't have to care.
To be fair: using Kubernetes anyways builds the skill just in case you become one of the 0.1% who actually need it down the line.
K3S takes about 5 minutes to setup the first time and you instantly have an entire universe of standardized operational tooling. I wouldn't touch docker compose with a 20 foot pole for production work.
Setting up K8s isn't rocket science, but maintaining it are offputting, to say the least.
Which is exactly what is happening with us, too bad we didn't choose K8S from the get-go and stuck with a "simpler" tool (gaining very little in the process).
This shittake was probably valid 10y ago, I would have agreed with you back then
> The entire infra the vast majority of Kubernetes users have could run on a single bare metal machine
Where are you pulling this out of? A large number of k8s users don't need it, but the alternative you have sounds hyperbolic.
The readme covers connection draining with Traefik which should solve one of the issues the author mentions
These are things I'm trying to figure out at work using Podman. Would love to hear about any experience in these areas.
This is pretty small scale, Kubernetes comes in when you've got a larger workload.
“We don’t know how to scale Traefik so we went with haproxy”
Well doh. Haproxy is designed for this. You can make haproxy serve copious amount of traffic on a single arm core and a little bit of ram. Imagine what you can do with a few replicas on your large clusters.
This has nothing to do with the choice of CI/CD or docker versus kubernetes.
Over the past decade, I'm seeing k8s used everywhere for everything, companies setting up clusters to run literally one simple app with couple of hundred requests per hour.
I think this line summarizes better than anything. Perfect example of how move fast and break things begins gloriously at first then inevitably, the breaking you thought you were doing hasn't even started and you find out what that part means.
I've always been the one saying "this is going to be a problem in a couple months" then I get shot down for "being negative." Then in a couple of of months when it fails I start getting aggression thrown at me "oh i know you want to say I told you so" and such even though I've never said such a thing when something fails. No. I would just like you to hear out my thoughts even when they may not be what you want to hear. We are all working towards the same goal.
the site is down for me.
Like, sorry, no, not to a point. Yes, if you have a small app without a lot of scale, and it doesn't need to be uber reliable and have very little if almost 0 downtime, okay, sure. Most use cases are like that! This is correct but applying it as a generality is just plain wrong and displays the type of arrogance people accuse kubernetes users of having.
What happens if a container in the VM goes down or the app inside of it crashes, how do you recover? Now you need some self-recovery mechanism via systemD or whatever, which will grow in complexity and fickleness over time. Congrats, you are now doing your own version of kubernetes.
What happens when you need to upgrade/restart your VM? Ok, make a standby VM as backup that will mostly sit idle, or require a full-app redeploy any time you need to do anything to the first VM. Now you need to design a blue/green mechanism between them, and probably some networking layer work. Congrats, you are now doing your own version of kubernetes.
What happens, if running in cloud, you have a regional outage or degradation? Stand up another VM in another region and manage the networking layer between them. Or, if running locally, your ISP has an outage because of a backhoe or something. Ok, we'll rent rack space in another data center as backup. Own all the mechanisms between cutting between those two now. Congrats, you are now doing your own version of kubernetes.
What happens if your app gets huge volume during peak times, and very little volume during non-peak, and you find yourself overprovisioning to the point your CFO/CTO freaks out about the bill? Well, we'll make our own dynamic scaling mechanism. Congrats, you are now doing your own version of kubernetes.
What happens when your app traffic gets so large you start running into OS limitations, like file descriptor limits? Start trying some of the aforementioned solutions. Congrats, you are now doing your own version of kubernetes.
What happens if you need service discovery, monitoring, or ensure network isolation between various services? Different VM's + your own hacked together service mesh, or wire something in the VM. Congrats, you are now doing your own version of kubernetes.
What happens when you need to guarantee secret isolation between containers? Congrats, you are now doing your own version of kubernetes.
Let's say you don't actually need any of this or think you never will. Fine! That's valid. But what you don't want, is to suddenly hit some scale and any of these things (I could list way more but I feel I am belaboring the point), migrating off these setups can become a year+ project, if not way longer. I know because I have had to do this twice now. I cannot possibly overstate how painful it is.
So, people usually just go with kubernetes because 1) it is operationally not that hard to deal with compared to the things I just mentioned, and has a massive ecosystem and 2) the risk of the VM + container spiraling into complexity is perceived as way more than going more complex at the start.
You just used a load-balancer (an AWS one preferably) with a few machines behind it, optionally an auto-scale group. AWS has multi-region support built in.
Re: file descriptors -- that's not something handled by kubernetes, if anything you just have more layers now.
You don't need service discovery, you never did, Host your services at a private zone {service-name}.{env}.company.com zones which points to the loadbalancer.
You need monitoring and should use an observability solution for that, has 0 to do with kubernetes.
You're just taking a bunch of random entirely solved problems and for some reason suggesting kubernetes is somehow helpful for them.
You do need service discovery at some point at scale if you are using any kind of distributed system. that is quite a strong claim. what do you have to support that claim?
You sound like you’ve never really seen scale and I’m not sure how to respond in a way that is polite. It is heavily dependent on your use case and industry. if you can get by with docker containers, good for you! My only point is you are probably reinventing the wheel here.
Bro I've worked at meta, I've worked at places that handled 1billion requests a day with a team of 20 people. I have a lot of questions myself about your experience that might be impolite as well. The fact that you linked kubernetes and file descriptors makes me think you really don't actually have the baseline fundamentals at all here.
meta only uses inhouse tooling as far as I’m aware, so what exactly is your expertise you’re claiming here?
I didn’t claim kubernetes handles file descriptors, lol. I was saying single vm solutions hit these OS limits faster than you’d think. Try understanding the argument you’re responding to. You haven’t responded with anything of substance, so I can safely ignore you, especially given your comment history seems to have an overwhelming bias against kubernetes, which is weird. Have a good day.
>What happens if a container in the VM goes down or the app inside of it crashes, how do you recover?
Docker will restart container automatically. You don't have to do anything. Docker-compose will restart after VM restart. You don't have to do anything. If a VM goes down - I do have a HA (another VM at another provider) and DNS load balancing.
>Now you need some self-recovery mechanism via systemD or whatever, which will grow in complexity and fickleness over time. Congrats, you are now doing your own version of kubernetes.
While I don't like systemd, it does this automatically, while, it's not really used here.
> What happens when you need to upgrade/restart your VM? Ok, make a standby VM as backup that will mostly sit idle, or require a full-app redeploy any time you need to do anything to the first VM. Now you need to design a blue/green mechanism between them, and probably some networking layer work. Congrats, you are now doing your own version of kubernetes.
This has been pretty much answered already but, upgrades does not affect containers (unless docker engine upgrade). Restarts - docker will handle these automatically - nothing to do here.
> What happens, if running in cloud, you have a regional outage or degradation? Stand up another VM in another region and manage the networking layer between them. Or, if running locally, your ISP has an outage because of a backhoe or something. Ok, we'll rent rack space in another data center as backup. Own all the mechanisms between cutting between those two now. Congrats, you are now doing your own version of kubernetes.
This actually handles way better w/o managed kubernetes, as it's usually a single region and your cluster and workloads would simply be completely down, while mine would work, because of provider redundancy.
> What happens if your app gets huge volume during peak times, and very little volume during non-peak, and you find yourself overprovisioning to the point your CFO/CTO freaks out about the bill? Well, we'll make our own dynamic scaling mechanism. Congrats, you are now doing your own version of kubernetes.
Kubernetes with autoscaling wins hands down here, but, it's not automatic, nor hassle free. You are also assuming overprovisiong which is usually not the case for traffic spikes.
> What happens when your app traffic gets so large you start running into OS limitations, like file descriptor limits? Start trying some of the aforementioned solutions. Congrats, you are now doing your own version of kubernetes.
This also affects k8s, exactly the same way.
> What happens if you need service discovery, monitoring, or ensure network isolation between various services? Different VM's + your own hacked together service mesh, or wire something in the VM. Congrats, you are now doing your own version of kubernetes.
I do have service discovery and network isolation built into docker, thanks.
> What happens when you need to guarantee secret isolation between containers? Congrats, you are now doing your own version of kubernetes.
Believe it or not, it's the default with docker.
> Let's say you don't actually need any of this or think you never will. Fine! That's valid. But what you don't want, is to suddenly hit some scale and any of these things (I could list way more but I feel I am belaboring the point), migrating off these setups can become a year+ project, if not way longer. I know because I have had to do this twice now. I cannot possibly understate how painful it is.
All my workloads are containerized and I can just move them to a k8s cluster whenever I want, if needed.
2) the risk of the VM + container spiraling into complexity is perceived. as way more than going more complex at the start.
The risk of your k8s ecosystem spiraling into operators madness and argoapps over helmfiles all while trying to accommodate for ci/cd and costs offing the chart is - IMHO - way higher.
Rolling your own zero-downtime deployments is as about as a good idea as rolling your own security... it's not a good idea.
I run our small system on a single EC2 instance with K3s. It runs a half-dozen or so services and does it quite well. I don't think it is particularly complex or over engineered. I like how easy it is to maintain the configuration in a helm package and quickly deploy it in different environments.
There is a learning curve for doing K8s well, but that's true for any non-trivial system.
sounds like ghetto engineering
But if someone wrote a blog to brag about not using k8s, they can't stop people from wanting to compare their work against k8s. If there's any arrogance in the air, it feels stronger on the other side.
EKS is literally one click of a button away and you dont need to handroll this.
even if you dont know AWS console nor terraform, claude code with aws mcp can do that for you
Out came Docker, dnsmasq, miles of duct tape and a whole lot of swearing. Just to come nowhere close to reinventing something better folks were doing years prior.
Just because you can (or think you can) doesn't mean you should. I sure do hope no one is maintaining that NIH monstrosity now!
(on iOS with dark mode enabled system wide)
I use it to keep from getting flashbanged by my monitor. In this case it also fixes the above site, however some websites need the color filter mode changed to work better, so realistically I’m not ending up with less fixing of websites, just easier fixing.