People will remark about how this is a waste of time, others will say it is absolutely necessary, even more will laud it just for the fun of doing it. I'm in the middle camp. I wish software/systems engineers would spend more time optomising for size and performance.
Docker adds other value to the lifecycle of your deployment. An "optimization" where you're removing value is just a compromise. Otherwise we'd all run our static sites on UEFI.
i've been meaning to ask you this for a decade. whatever happened to when you wrote a blog with insanely irritating serifs that connected certain letters together? what was the rationale behind that? never seen it since
I'm insanely impressed by APE and redbean by the way, blows OP out of the water!
Oh you mean the blog with the long s? I was reading a lot of books at the time that were written before 1800 and I found it so fascinating how different typography was back then. I found a font I could pay for called Quant that did a really good job reproducing archaic ligatures and the long s, so I used it on a blog for a short period of time. Sadly it got negative feedback. So lately I've been focusing on https://justine.lol/ which uses Roboto. I'm glad to hear you're enjoying it!
ahhh ligatures, that's the word I was looking for here. Yeah it was kind of irritating but good blog content so I just read it anyways. it was kind of hard not to read it as someone lisping everything though.
This is a really good point, and something I think a lot of people forget. It's true, the most secure web app is one written with no code/no OS/does nothing.
Adding value is a compromise of some increased security risk - and it's our job to mitigate that as much as possible by writing quality software.
You can have multiple instances of the server running on the machine without interfering with each other.
You can limit file system access for the server to only a certain folder.
You can similarly limit port access and manage conflicts (e.g. multiple servers can think they are listening on a certain open port but those are mapped to something else on the host).
If you have multiple machines with different operating systems or even architecture you can deploy your server as a container more easily on them without needing to rebuild or test for each one.
You can have the same environment running locally while development or on CI servers without complicated setups.
The system can scale out a lot more easily to hundreds/thousands of machines if you decide to use something like Kubernetes.
You have to login to some docker repository anyways and know the series of commands to actually run it. Cloning a repo and running a shell script is probably a lot easier and faster than that.
What kind of work are you doing that requires really fast auto scaling? Is a few minutes to spin up a new instance really that cumbersome? Can you not signal for it to spin up a new instance a tiny bit earlier than when it's needed when you see traffic increases?
> You have to login to some docker repository anyways and know the series of commands to actually run it. Cloning a repo and running a shell script is probably a lot easier and faster than that.
In isolation, yes. But if, for instance, you're already running a container orchestration tool with hundreds of containers, and have CI/CD pipelines already set up to do all of that, it's easier just to tack on another container.
Ok when you say a few off the top of my head it implies that there are a bunch and these are like some super obvious ones, but it sounds like this is actually only useful if you have a bunch of infrastructure set up to serve sites for projects and customers that need containerization and then you just throw this simple little static site docker instance in there because when you're maintaining a lot of docker instances it is just simpler to do?
Which seems like sort of an edge case for value adding, and makes me feel like it really doesn't add any value to do this unless you already are doing it for everything, and thus you really wouldn't be throwing out any value by just serving the static site without the docker overhead.
Adding to some of the other responses, one reason I chose to deploy a SPA I'm working on as a Docker image is atomicity - if I want to deploy a newer version I simply switch out the tag in my container orchestrator's config (Nomad in this case, but the same principles apply to k8s and friends) and it's guaranteed that the new deployment will be pristine, without the risk of leftover files from a rsync or similar - and if I need to roll back I do the exact same.
There’s value in that, but you don’t need Docker with its related debugging and maintenance overhead to get it. NixOS, among other tools, will do the same thing while constructing a “flat” operating system image.
Anything else, though? There’s got to be more to it than that, or it wouldn’t be as popular as it is.
I'm super curious to know what the value to people who see that happens to be. It's serving static websites, why do I need to wrap THAT of all things in a container?
> why do I need to wrap THAT of all things in a container?
If you can't see a reason why, then you probably don't need to. You probably have different needs than other people.
Many people use Docker not because of what they're doing inside of the container, but because it is convenient for tangential activities. Like lifecycle management, automation, portability, scheduling, etc.
I have several static sites in Docker containers in production. We also have dozens of other microservices in containers. We could do everything the same way, or we can one-off an entirely separate architecture for our static sites. The former makes more sense for us.
Because you want a reproducible environment/runtime for that static server. Nix/NixOS takes it a step further, in that it provides not only a reproducible runtime environment, but a reproducible dev and build environment as well.
I actually found myself needing something like this a couple weeks ago. I use a self-hosted platform (cloudron.io) that allows for custom apps. I wanted to host a static blog on that server. Some people are happy to accept "bloat" if it does, in fact, make life easier in some way.
If you literally ONLY ever need to run a single static website, then yeah, containers might not be helpful to you.
But once you start wanting to run a significant number of things, or a significant number of instances of a thing, it becomes more helpful to have a all-purpose tool designed to manage images & run instances of them. Having a common operational pattern for all your systems is a nice, overt, clean, common practice everyone can adapt & gain expertise in. Rather than each project or company defining it's own deployment/operationization/management patterns & implementations.
The cost of containers is also essentially near zero (alas somewhat less true with regards to local FS performance, but basically equal for many volume mounts). They come with great features like snapshots & the ability to make images off images- CoW style capabilities, the ability to mix together different volumes- there's some really great operational tools in containers too.
Some people just don't have real needs. For everyone else...
OpenBSD for the past 10 years or so has been really good to me and my clients, and it just keeps on getting better while linux keeps on getting worse. It's kind of a nobrainer these days.
Hell if you just need to serve static HTTP it even has its own built in webserver now:
You forgot the step where you had to provision that server to run the software and maintain all the systems security updates on the live running server, and that server requires all the same maintenance, with or without docker. And if you fuck it up, better call the wife and cancell Sunday plans because you forgot how it all gets installed and ......yeah, just use docker :p
Security updates are only needed on the OS level if you're running Docker on bare metal or a VPS. If you're running Docker in a managed container or managed Kubernetes service such as ECS/EKS, you only need to update the Docker image itself, which is as simple as updating your pip/npm/maven/cargo/gem/whatever dependencies.
I see two main places where Docker provides a lot of value: in a large corp where you have massive numbers of developers running diverse services on shared infrastructure, and in a tiny org where you don't have anyone who is responsible for maintaining your infrastructure full time. The former benefits from a standardized deployment unit that works easily with any language/stack. The latter benefits from being able to piggy-back off a cloud provider that handles the physical and OS infrastructure for you.
I am arguing against Docker for maintainability reasons, not CPU cycles.
Relying on hidden complexity makes for a hard path ahead. You become bound by Docker's decisions to change in the future.
For example, SSLPing's reliance on a lot of complex software (among which NodeJS and Docker) got it to close, and it got on the front page of HN recently.
Docker Swarm isn't Docker; it's an orchestration service on top of Docker, that happens to have originated with the same organization that Docker does - hence the name. A few years ago Swarm looked like it might be competitive in the container orchestration space, but then Kubernetes got easy to use even for the small deployments Swarm also targeted, and Swarm has withered since. It wouldn't be impossible or probably even all that difficult to switch over to k8s if that were the only blocker, but as the sunset post notes and you here ignore, that wasn't the only or the worst problem facing SSLping - as, again, the sunset post notes to your apparent disinterest, it's been losing money for quite a while before it fell apart.
(Has it occurred to you that it losing money for a while might have contributed to its eventual unmaintainability, as the dev opted sensibly to work on more sustainably remunerative projects? If so, why ignore it? If not, why not?)
Similarly for the Node aspect - that's very much a corner use case related to this specific application (normally SSLv3 support is something you actively don't want!), and not something that can fairly be generalized into an indictment of Node overall. Not that it's a surprise to see anyone unjustly indict Node on the basis of an unsupportable generalization from a corner case! But that it's unsurprising constitutes no excuse.
Other than that you seem here to rely on truisms, with no apparent effort to demonstrate how they apply in the context of the argument at which you gesture. And even the truisms are misapplied! Avoiding dependencies for the sake of avoiding dependencies produces unmaintainable software because you and your team are responsible for every aspect of everything, and that can work for a team of Adderall-chomping geniuses, but also only works for a team of Adderall-chomping geniuses. Good for you if you can arrange that, but it's quite absurd to imagine that generalizes or scales.
Just because a project has a larger budget, doesn't mean any of it should be spent on Docker, Docker Swarm, or Kubernetes or whatever other managers of Docker that you can mention here.
Fact is, for 3 servers, it would be hard to convince me of any use of Docker compared to the aforementioned deployment shell script + Debian unattended-upgrades.
What problem does Kubernetes address here? So what if it is "easy to use"? I prefer "not needed at all".
> but also only works for a team of Adderall-chomping geniuses
Of course, not everything should be implemented by yourself. Maybe this project wouldn't have been possible at all without offloading some complexity (like the convenient NodeJS packages).
But in particular Docker and its ecosystem are only worth it when you have an amount of machines that make it worth it - when things become difficult to manage with a simple shell script everyone understands: when you have a lot of heterogeneous servers, or you want to deploy to the Cloud (aka Someone Else's Computers) and you have no SSH access.
> truisms
I don't have any experience with Kubernetes nor Docker Swarm. The reason is that the truisms have saved me from it. If you don't talk me into learning Kubernetes, I won't, unless a customer demands it explicitly.
> Has it occurred to you that it losing money for a while might have contributed to its eventual unmaintainability
It absolutely has. Maybe if the service hadn't used Docker Swarm or Docker at all, it would have lasted longer, since updating Docker would not have broken everything, since this was named a factor in the closure. And therefore, the time and money would have gone further.
And maybe if my grandmother had wheels she'd be a bicycle. But - over a so far 20-year career of which "devops" work has been sometimes a fulltime job, and always a significant fraction, since back when we called that a "sysadmin" - I've developed sufficiently intimate familiarity with both sorts of deployment workflows that, when I say I'll always choose Docker where feasible even for single-machine targets, that's a judgment based on experience that in your most recent comment here you explicitly disclaim:
> I don't have experience with Kubernetes nor Docker Swarm. The reason is that the truisms have saved me from it.
Have they, though? It seems to me they may have "saved" you from an opportunity to significantly simplify your life as a sysadmin. Sure, your deployment shell scripts are "simple" - what, a hundred lines? A couple hundred? You have to deal with different repos for different distros, I expect, adding repositories for deps that aren't in the distro repo, any number of weird edge cases - I started writing scripts like that in 2004, I have a pretty good sense of what "simple" means in the context where you're using it.
Meanwhile, my "simple" deployment scripts average about one line. Sure, sometimes I also have to write a Dockerfile if there isn't an image in the registry that exactly suits my use case. That's a couple dozen lines a few times a year, and I only have to think about dependencies when it comes time to audit and maybe update them. And sure, it took me a couple months of intensive study to get up to speed on Docker - in exchange for which, the time I now spend thinking about deployments is a more or less infinitesimal part of the time I spend on the projects where I use Docker.
Kubernetes took a little longer, and manifests take a little more work, but the same pattern holds. And in both cases, it's not only my experience on which I have to rely - I've worked most of the last decade in organizations with dozens of engineers working on shared codebases, and the pattern holds for everyone.
I don't know, I suppose. Maybe there's another way for twenty or so people to support a billion or so in ARR, shipping new features all the while, without most months breaking a sweat. If so, I'd love to know about it. In the meantime, I'll keep using those same tools for my single-target, single-container or single-pod stuff, because they're really not that hard to learn, and quite easy to use once you know how. And too, maybe it's worth your while to learn just a little bit about these tools you so volubly dislike - if nothing else, in so doing you may find yourself better able to inform your objections.
All that said, and faint praise indeed at this point, but on this one point we're in accord:
> If you don't talk me into learning Kubernetes, I won't, unless a customer demands it explicitly.
I did initially learn Docker and k8s because a customer demanded it - more to the point, I went to work places that used them, and because the pay was much better there I considered the effort initially worth my while. That's paid off enormously for me, because the skills are much in demand; past a certain point, it's so much easier to scale with k8s especially that you're leaving money on the table if you don't use it - we'd have needed 200 people, not 20, to support that revenue in an older style, and even then we'd have struggled.
I still think it's likely worth your while to take the trouble, for the same reasons I find it to have been worth mine. But extrinsic motivation can be a powerful factor for sure. I suppose, if anything, I'd exhort you at least not to actively flee these technologies that you know next to nothing about.
Sure, you might investigate them and find you still dislike them - but, one engineer to another, I hope you'll consider the possibility that you might investigate them and find that you don't.
> what, a hundred lines? A couple hundred? You have to deal with different repos for different distros, I expect, adding repositories for deps that aren't in the distro repo, any number of weird edge cases
Well, here is where I must thank you. Thank you for replying to me, and giving me a real reason to look at this ecosystem.
My personal deployment script is really just one SCP command - it copies the new version of my statically-built blog to my server. The web server comes with the OS, and that's all I need.
But when I read "hundred lines? A couple hundred?" I realized my company has a script fitting that bill. There might be an opportunity for improving it. While I am still somewhat skeptical, because using Kubernetes instead of that script might still not be worth it long-term for a 7-person company (of which 3 devs and one sysadmin), I will check out its capabilities.
> in so doing you may find yourself better able to inform your objections.
Thank you for the patience to follow up, in spite of my arrogance. I might just come up with an improvement somewhere. Certainly we're a long way from a billion in ARR - I am thankful for your valuable time, and wish you continued and further success!
Don't worry about the arrogance - I was much the same myself, once upon a time. :) It's worth taking thought to ensure you don't let it run too far away with you, of course, but I'd be a fool to judge harshly in others a taste of which I once drank deep myself. And hey, what the hell - did anyone ever change the world without being at least a little arrogant? There are other ways to dare what seems impossible to achieve, I suppose, but maybe none so effective.
One thing, I'd suggest looking to Docker before Kubernetes unless you already know you need multi-node (ie, multi-machine) deployments, and maybe as a pilot project even if you do. Kubernetes builds upon many of the same concepts (and some of the same infrastructure) as Docker, so if you get to grips with Docker alone at first, you'll likely have a much easier and less frustrating time later on than if you come to Kubernetes cold. (And when that time does come, definitely start with k3s if you're doing your own ops/admin work - it's explicitly designed to work on a smaller scale than other k8s distributions, but it also pretty much works out of the box with little admin overhead. As with starting on Docker alone vs k8s, it's all about managing your frustration budget so you can focus on learning at a near-optimal rate.
But hey, thanks for the well-wishes, and likewise taking the time in this thread! It's been of real benefit to me as well. If we're to be wholly honest, as an IC in my own right I've never been above mid-second quartile at absolute best, and at my age I'll never be better than I am today - or was yesterday. But that also means I'm at a point in my career where I best spend my time helping other engineers develop; if I can master the skill of making my accumulated decades of experience and knowledge, and whatever little wisdom may be found in that, of use to engineers who still have the time and drive to make the most of it in ways I maybe failed to do - well, I'll take it, you know? It's not the kind of work I came into this field to do, I suppose, but I've done enough of it by now to know both that I can do it, and that it is very much worth doing.
So, in that spirit and quite seriously meant - I might be off work sick this afternoon, one peril I'm finding attends ever more frequently upon advancing age, but evidently that's no barrier to improving the core skill that I intend to build the rest of my career around. Thank you for taking the time and trouble to help make that possible today, and here's likewise hoping you find all the success you desire!
That's true, but in my experience there is nothing mutually exclusive in systems being simple and systems running Docker.
Granted, you do need to learn how Docker works, and be ready to help others do likewise if you're onboarding folks with little or no prior experience of Docker to a team where Docker is used. That's certainly a tradeoff you face with Docker - just as with literally every other shared tool, platform, codebase, language, or technological application of any kind. The question that wants asking is whether, in exchange for that increased effort of pedagogy, you get something that makes the increased effort worthwhile.
I think in a lot of cases you do, and my experience has borne that out; software in containers isn't materially more difficult to maintain than software outside it if you know what you're doing, and in many cases it's much easier.
I get that not everyone is going to agree with me here, nor do I demand everyone should. But it would be nice if someone wanted to take the time to argue the other side of my claim, rather than merely insisting upon it with no more evident basis than arbitrarily selected first principles given no further consideration in the context of what I continue to hope may develop into a discussion.
Whatever set-up your application needs is a still necessary step in the process. But now you've not only added more software in docker with its a docker registry, and Docker's state on top of the application's state, you've also introduced multiple virtual filesystems and a layer of mapping between those and locations on the host, mappings between the container's ports and the host's ports. There is no longer a single truth about the host system. The application may see one thing and you, the owner, another. If the application says "I wrote it to /foo/bar", you may look in "/foo/bar" and find that /foo doesn't even exist.
All of that is indirection and new ways things can be that did not exist if you just ran your code natively. What is complexity if not additional layers of indirection and the increase of ways things can be?
Okay, and in exchange for that, I've gained single-command deployments of containers that already include all the dependencies their applications require, and at most I only have to think about that when I'm writing a deployment script or doing an update audit.
It's rare that I need to find out de novo where a given path in a container is mapped on the host. When I do need to do that, I can usually check a deployment script, or failing that inspect the container directly and see what volume mounts it has.
I don't need to worry about finding paths very often - much less frequently than I need to think about deployments, which at absolute minimum is once per project.
So, sure, by using Docker I've introduced a little new complexity, that's true. But you overlook that this choice does not exist in a vacuum, and that that added complexity is more than offset by the reduction of complexity in tasks I face much more often than the one you describe.
And that's just me! These days I have a whole team of engineers on whose behalf, as a tech lead, I share responsibility for maintaining and improving developer experience. Do you think I'd do them more of a favor by demanding they all comprehend a hundred-line sui generis shell script for deployments, or by saying "here's a single command that works in exactly the same way everyone you'll work with in the next ten years does it, and if it breaks there's fifty people here who all know how to help you fix it"?
To host something as a docker container I need 2 things: to know how to host docker, and a docker image. In fact, not even an image, just a dockerfile/docker-composer.yaml in my source code. If I need to host 1000 apps as a docker containers, I need 1000 dockerfiles and still to know (and remember) 1 thing: how to host docker. That's 1 piece of knowledge I need to keep in my head, and 1000 I keep on a hard-drive, most of the time not even caring what's the instruction inside of them.
If I need to host 1000 apps without dockerfiles, I need to keep 1000 pieces of knowledge in my head. thttpd here, nginx to java server there, very simple and obvious postgres+redis+elastic+elixir stack for another app… Yeah, sounds fun.
I think the real value is just focusing on the absolute minimum necessary software in a production docker/container image. It's a good practice for security with less surface area for attackers to target.
The difference between a systems engineer and a software engineer is that to a systems engineer a half functioning 5MB docker image is okay but to a software engineer a fully functional 5GB Node image is fine.
People will remark about how this is a waste of time, others will say it is absolutely necessary, even more will laud it just for the fun of doing it. I'm in the middle camp. I wish software/systems engineers would spend more time optomising for size and performance.