The systems in your 2022 world are hard to test and maintain because they are bad, and the tools we built to test and maintain them are largely built on the same foundational ideas and technologies, so they are even worse.
We're going to have to rip everything back down to the foundation in order to make progress beyond finger-pointing (X is DevOps but Y is software engineering and Z is IT admin).
No, not really. Nowadays things are way way better than how they were a decade ago. Nowadays you can get pipelines that build and test on multiple target platforms with a dozen or so lines of code, and things run so well that they became reliable infrastructure integrated with your prod system.
Back then you had none of this, and had to pay the salary for a "build engineer" who worked like a wizard stirring a cauldron just to get a single production build out of the door.
We need better tools if we want better things.
There's so much more going on. I don't know if the complexity is all required but we're doing more now. There's just more going on. That's how it is. We need the abstractions. Maybe not all of them. But we need more abstractions and tools running now. There's so much to manage.
I don't see us reducing this complexity that much other than consolidation and some simplification to clean things as they get sorted further. But there is still a need for the layers. It's not 2005. Things are more complex.
The nearest analogy is would you like to code everything in line numbered BASIC or would you like to do it using some modern equivalent? Even with the additional layers there has been real progress in the tools. Improvements across a multitude of metrics. It's not all complexity for the sake of complexity.
Smile. It's another stage of progress. The old mess will fall away. At the next stage what we see as an improvement now will be the next stage's mess. There's a whole slew of problems that will need even newer tools.
We haven't even scratched the surface of the automation we'll require in ten years time.
Really not sure what you're commenting about here
It doesn't matter to me who turned the wrench, but in my career I've encountered these types where it's only good enough if it's theirs. And IMHO their output is generally quite atrocious. Root cause: Psychological issues.
What’s missing a lot is empathy, and that takes slowing down and observing.
Just to share the terminology, under underlying stuff I mean the linux kernel, the JVM, etc, while higher level stuff could be certain build tools (khm, looking at you npm), and of course end user tools that are unfortunately way too buggy. Like I have a hard time listing end user applications where I didn’t notice a bug yet.
Give me a tight, highly coherent API, with a culture of avoiding feature factory work. At least get "Make it work, make it right, make it fast" to not skip "make it right" anymore. That's sort of the core illness in software today.
Open source doesn't have to be pushed by business concerns. There's a degree of regulatory capture, but the biggest problem is that we just don't know any better. We repeat what we see, and make sure our own pain points are handled. I've seen this play out in API design where an awful API is introduced, and each person who fixes it only fixes 20% of the pain, and so it's 10 years before we get from a mediocre library to one you could actually call 'good', because nobody made 'good', they made better than better than better than ridiculous.
In my past at $medalyaorwherever it was a fucking nightmare everyday, the Architect wouldn't let the PR close.
I tell my boss you want 2-3 people like me on the project. If everyone was like me we'd make each other miserable. I have worked on trying to frame these things in a positive manner. Unfortunately the guy who would have been the best mentor in that regard went into semi-retirement.
I don't know about OP, but I personally feel like software like GCC/clang/MSVC and Debian/Ubuntu and Docker and SSH and Git are pretty great, and were never better.
Heck, the whole dotnet ecosystem is turning some significant problems in developer experience into at most minor nuisances.
Even Firefox+Chrome are stellar, and extremely solid as-is.
Where exactly is this bad software OP is talking about?
If you are using one of the major cloud providers and instead of just using a native IaC scripts, you have decided that it would be better if the developers only had access to your custom kubernetes operators to manage their infrastructure then you are the problem.
If you are using a deployment pipeline that developers have zero involvement in, and they just "import your Jenkins scripts" then you are the problem.
If you are using major cloud provider and rather than just using their managed kubernetes, you have decided to deploy your own, then the chances are that you are the problem.
Rarely have these approaches ever been stable, in my 20+ year experience they have never been stable. And they are the reason that "real DevOps" came along because developers could do a better job with less downtime if people actually engaged them.
Asking prospective "DevOps", "software engineer" and "IT admin" candidates to program in C is an effective way to weed out those who are not truly capable of writing "good" software. It is easy to find mistakes and expose the incompetent. That may offend the incompetent who believe they are "good" programmers. Thus there are new programming languages created every year, more people using them to write "bad" software, and more people who feel emboldened to attack C as the source of problems, instead of "bad" programmers.
C is truth serum. It is difficult for people programming in C to pretend they are "good" programmmers. They will be found out. It is funny how people on HN attack the language for exposing so many "bad" programmers. Under this theory, programmers are absolved of all responsibility for their mistakes. The language is at fault. "Good" programmers do not blame languages for their own mistakes.
IMHO, there is sometimes a benefit to languages that make it more difficult, not easier, to create and sustain Rube Goldberg complexity. Not to mention languages that compile to smaller, faster programs. "Programmer productivity" is ambiguous. For example, it could be a positive for a programmer trying to justify the expense their salary presents to an employer or it could be a negative to people who are forced to use an ever-increasing quantity of "bad" software. It could be a positive to a programmer who wants to keep implementing new "features" or it could be a negative to a software user who dislikes "feature creep".
The dichotomy of good versus bad software is such a subjective topic that the term "we" really needs to be defined. Different groups of people have different interests and therefore different opinions.
> The language is at fault. "Good" programmers do not blame languages for their mistakes.
I also know that my time to get something functional in higher level languages is often 10x less, and my probability of having very subtle bugs to hunt is much lower.
There's a super-weird tradeoff here. There's all kinds of modern, high-level techniques that improve programmer productivity and reliability of bigger systems.. They convert people who can't be productive C programmers into doing OK work.
But they're also slow and a bit opaque in how they work. And it's really easy to run out of performance and have to do exotic things to get them to scale, in which case you have to build bigger systems still and cede a lot of those advantages.
Whereas, if you write to the metal, you can get a lot of performance out of a single large computer.
[1]: http://dtrace.org/blogs/bmc/2018/09/28/the-relative-performa...
Yes, Rust may have somewhat better programmer productivity than C, but it's not really at a massively higher level of abstraction.
Come on, C is not an exceptional language from any point of view. It just pretends that everything is a fast PDP-11, and thanks to the insane amount of work and mental gymnastics done by the compiler, it will output good code. It has shitty abstractability that hinders performance if anything (just look at the small string optimizations done by C++) and hurts maintainability and productivity. At least it should have a proper macro system, but preprocessor macros are just a disgusting hack.
It is if not outright a “bad language”, it should definitely be left behind outside of maintaining the litany of programs written in it, and one should start new programs in either managed languages (if absolute control over the hardware is not required), or at least in C++ or Rust.
The question is whether C is useful. For example, in building things from the ground up. Or vetting programmers.
Another way to vet programmers is to observe how many negatives they try to cram into their sentences. For example, "Not being... doesn't... non-popular."
If they are unable to make clear, concise, positive statements the same deficiency is likely to carry over to writing programs.
Everyone wants to bank on the software craze, and so everyone develops pointless middleware solutions.
Also, looking at principles over tooling, DevOps has a lot to offer that trivially make sense.
Personally, I’ve worked through having this mindset myself. I “grew” up with the C and tots descendant family of languages. I wrote shellcode in C and ASM as an exploit dev. I later learned python and now ruby, arguably some of the most abstracted languages.
At each iteration I openly opined at how “C would let me do X without the handcuffs I put myself in by using Language Y”. It’s really just senseless complaining. There is a reason why someone chose Ruby over PHP. PHP over pascal, Pascal over Ada and so on. The point is, to scoff at the product as a bystander is just immature. Throwing shade on an entire industry based off some weird pet purist outlook is just immature :/.
Of course, the irony here is that this type of mindset is no different from the apprentice carpenter who scoffs at every new house he walks through that he didn’t build. It’s no different from the wine snob who cringes at what his mom puts out when company is over. It’s no different from the tens of submissions HN accumulated in a week with something of the tune “I rewrote X in Rust”.
The fact is that businesses either are or are converting to cloud-first because why incur the risks of maintaining the one special-snowflake server that runs your business when you can turn a crank and spin up as many instances as you need -- theoretically very many as you take on more customers and transactions and need to handle the load? And ypu can just restart them if they go down?
Stop thinking technology and start thinking BUSINESS. The cloud may be more janky and complicated than what you're used to but it serves the business's needs better.
I'm pretty certain you can do the same with software written in C (or any other language) without resorting to AWS/Azure/GCP/etc.
You can do the same cranking on your own hardware, except you keep the hardware after the peak has passed and you need to plan for the peak.
Or, you can simply get a DO droplet, and crank out as many new instances as you need, and turn them off when you don't, with software written in C (or anything else).
There's no need to go all-in on the AWS/Azure/GCP solution, with the ability to orchestrate instances as cattle, when you have, at peak, 3 instances[1].
[1]If you wrote your software in C, or Rust, or Go instead of Python, or Ruby, you could potentially get away with fewer instances than you need, thereby reducing to treating the system as a system of pets because there are so few of them.
1. Never going to happen unless everything collapses, which is extremely unlikely.
2. We are progressing. AR, VR, GPUs with raytracing, complex web applications and global services, global cloud, etc.
(Long, but worth it, assuming you haven't seen it already)
People are always suprised when they see a minimal webserver instance on a ten year old debian [0] handling tens of thousands of requests, without issue. It might go down once a decade because the disk failed, but it cost 1200$ to run over that decade. I don't think that's the perfect way, but modern infrastructures love to include a lot of complexity when it's not needed. The problem is that including something is very easy and the cost is only payed once it breaks. Also, hardware is really cheap (you might not think so, but if you compare it to western IT salaries, it is).
[0] I know outdated instances with no update strategy are not really a benchmark, but I encourage you to go out and ask people what their container base image upgrade strategy is - the situation really did not change that much.
I think that's the key. Most of us aren't Google/Facebook/whatever. Yet lots of people want to mimic those. A simple stack with a relation database on a single machine still goes a long way. In fact, it goes a lot longer way today than it did 20 years ago. It's ok to scale when you need to, but I think premature scaling is the wrong way to go, especially since most things never get to FB/Google/etc size.
Every time this argument pops up, I end up wondering if it’s really about architectural complexity, or if it’s about sticking with the old and known.
Now to try and answer your question, I think a relational database is a lot more complex than a simple key value store, but if it's a good implementation, it will hide this complexity from the user, and using the database will be rather simple. It can be as simple as a key value store since you can just use a relational database like one. Note that this hiding of complexity is by design, it is one of the main goals of the relational model, it's not accidental.
Example Go's mgo driver by Gustavo Niemeyer was simple and effective, but abandoned. The official drivers is unnecessarily complicated. And Go's need to have "context" everywhere adds to this, but MongoDB is not complex. Idk I'm not some super genius and picking up MongoDB was really easy. Querying (aka aggregation pipelines), you have to think of that as "pipes in bash". It's in the words aggregation pipelines. find | groupby | sort | filter | select. Something like that. It's not SQL it's different, but not complex, sorry.
Like, if you were never exposed to rxjs and are now trying to build things with it doing
$stream.pipe(
switchmap(),
filter(),
etc(),
)
stream.map().filter().etc()
Now write DBA scale ISO SQL 2016 / Transact SQL / PL/SQL in that interesting Mongo flavoured language, including database engine debugger integration, and JIT compilation.
a raspberry pi can pretty much serve 100k req. / second. An average dev laptop should be able to handle 1M req. / second without much issues
Depends on what you do. It's quite easy to hit these numbers with a low number of static pages, but apps that can easily work with static pages usually don't have DevOps people. That's why I went with thousands, to have a reasonable number for an app that actually does some complex processing.
I appreciate the sentiment, but my intuition says that number is too high. What would that look like? I can't picture nginx or haproxy serving up a static 'hello world' response at that volume on a Pi3.
And if you do anything with a slow-ish 3rd party API you're probably going to hit the TCP socket limit before you can respond to 100k/s.
Legitimately curious about whether my intuition is wrong. I don't have an unused Pi handy to test it myself.
Edit: It's possible you're citing https://levelup.gitconnected.com/serving-100k-requests-secon... ? Pretty interesting article in any case, especially this claim:
> This test pitted a native C++ server against a scripted Node.js one, so of course the outcome was given. Well, not really. If I run the same test using µWebSockets.js for Node.js, the numbers are a stable 75k req/sec (for cleartext). That’s still 8.5x that of Node.js with Fastify and just shy of 75% of µWebSockets itself.
Edit 2: Doing a bit more research and finding some other benchmarks online, it seems like 100k with µWebSockets is plausible. I recant my skepticism.
I'm not sure you have much experience working on websites. Multiple deployments are required for performance reasons, which is the main driving force for paying hefty bills for CDNs, and their recent investment in edge computing.
You deploy to multiple deployments in multiple regions in a good part to drive down those 300ms requests to sub 50ms. Otherwise your user gets fed up and drops you in favor of a competitor.
But if you work on a service who has only a dozen users in the same region and can wait in turns to see who can make a request, go right ahead with your production raspberry pi setup.
one raspberry pi per continent :p
e.g. I'm in a small rural village in southern france right now and I get 2GB down / 1GB up more or less with a 9ms speedtest ping (which is slow, it's often 2ms) - anyone from europe would get < 50ms ping notwithstanding the latency between them and their isp
by the time my website has, say, > 500k req/second I can definitely invest in, idk, some used 2015 i7 laptop to give myself a x10. For reference, 1M requests per second is Instagram in 2015, one of the most used websites in the world - by the time you're at that scale you've been bought by Facebook anyways.
I bought artisanal knitted socks for Christmas. If their website is faster than their knitting that’s good enough. That website is part of the 99%.
I ultimately agree with you, but in between a raspberry pi and “hefty bills CDNs” is something like a half rack at a Colo, a f5, and a couple PowerEdges.
(Half kidding)
Not every page loads 50 different resources. Not every page serves people globally. Not every page has competitors.
I might say the same about you. Be respectful.
How big are these requests? If you're returning more than 1kB including all the headers, you're not beating 100kreq/s on a standard consumer-level 1Gbit link.
It wouldn't be possible to get that request rate outside of a microbenchmark, just because of the network speed the pi supports.
- 2.5G: https://www.jeffgeerling.com/blog/2020/testing-25-gbps-ether...
- 10G (in practice 3.6G max which is more than what most can expect from their ISP): https://www.hackster.io/news/10-gbps-ethernet-on-a-raspberry...
At my current company we're at cusp when we need to address scalability to able to grow as a business, and the business is growing. It took several years to get to this point, but here we are.
I prefer simplicity as much as the next guy, but at some moment it becomes simpler overall to use cloud services controlled by k8s than to maintain large amounts Ansible scripts and Jenkins jobs.
With that, by the way, comes an inevitable untangling of various parts in the software, and simplification of it. To be fast you need the processing path to be simple.
What functionality? Automated unit tests to ensure you're not shipping code that is so broken that it even breaks basic invariants? Automated integration tests that ensure you're not shipping code that fails to call other components/services? UI tests to ensure that your latest commit didn't made your prod service unusable for the end user? Performance tests that help you determine if your latest change doesn't suddenly exhaust your computational resources in peak demand?
I think the increased complexity in our architectures is correlated with the DevOps role coming to the picture, but I'm not sure there's a cause link there. My recollection from having lived through the initial years of DevOps (I began working professionally in 2000) is that an important goal was to reduce the terrible friction between dev and ops (the latter including sysadmins, DBAs, etc). Whatever the extra complexity we have now, I would not want to go back to the days when I carried a pager and responded to find out the problem was caused by an application I had no insight into, and there was no corresponding on call person on the side of the dev team. Another important goal was to manage infrastructure a bit closer to how we develop software, including the big step of having the configuration (or the code that generates it) in SCM. Another thing I don't miss is logging into a server to troubleshoot something and finding /etc/my.cnf, /etc/my.cnf.<some date>, /etc/my.cnf.test, etc. It's much easier to just run tig on either the file or the ansible/chef/whatever that generates it IMHO.
Unit tests weren't even common practice back then. Yeah, most places had tests but there was no common language to describe them.
And as much as git can be a big complex pain. . . merging was a BIG thing back then too. I seldom deal with long lived branches and the nightmarish merges they often needed.
Also, to all the young folks who "want to work with physical servers" again. Have fun with the next set of device driver patches you need to roll out.
I heart my containerized IaC buzzword laden cloud existence. #FuckThatNoiseOps
2022 Your app takes some user data, stores it in a database, and presents a filtered view of it to the user on request.
Since then I've working on things that use several servers, things that use no servers with edge functions instead (which have servers, but still), and that use autoscaling across lots of servers.
No doubt some businesses need autoscaling but most don't. It definitely shouldn't be something you reach for before you actually anticipate a need for it.
Everyone talks FAANG scale on day 1. It's common and absurd.
Yeah, because obviously nobody handled any payment in 2005 … The same can be said for everything in your list.
And more importantly, it's unlikely that your business is more complex than what it used to be in 2005. And if you're spending more resource to deliver the same business value, you're wasting them. (That's why PHP is still ubiquitous btw, it sucks by most standards, but it's good enough to do business and that's what matters).
For example, why do you "queue tasks for later"? Do you have exceptionally punishing latency and throughput requirements that can only be met at a reasonable cost by not answering synchronously, or it's because your database doesn't do transactions?
Similarly, what do you do with "queues, caches, and so on"? Meet further extreme performance and availability requirements, or attempt to mask with additional complications the poor performance of inefficient complicated components?
In 2005, but also in 2000, web applications had already progressed past "just serving a web page", mostly without complicated architectures and therefore without the accompanying tools.
I think tool improvements made cargo-culting actually advanced state of the art software architectures and processes easy and affordable, creating a long term feedback loop between unnecessary demand for complexity (amateurs dreaming of "scaling up") and unnecessary development and marketing of advanced (but not necessarily good) complexity-management tools, often by the same amateurs.
Load balancers and HA setups were already in use. LVS existed. VMs were popular. "CI/CD" existed, without that name.
> 2005 your average box ...
Speak for yourself.
Sure we built a bunch of it ourselves, and we had 4 sysadmins, but please don't tell me that this wasn't already going on over 20 years ago.
By 2004 we were doing perforce and CI. No Continuous Deployment because we were building on-prem software, but by 2008 I was doing svn and Continuous Delivery (similar problem, not hosting).
We would have been if it had been appropriate.
To be fair though, I think a lot of my coworkers suspect i’m talking out of my ass because they’ve been doing this stuff for 5-8 years and think everyone else has too. As tempting as it sometimes is to launch into a lecture., it wouldn’t really help. Everybody knows what they know, and they’re going to know it until they watch it catch on fire and then they will either want to hear about alternatives or they’ll share ideas for fixing it, not realizing they’re repeating something you told them two years ago. Hey that’s a great idea, wish I’d thought of it.
> 2 sysadmins were absolutely not managing “large scale deployments with multiple data centers”
Please don't make things up because I was there.
> But today it’s actually super plausible that 2 or 3 Devops with modern tooling can manage all of that.
Not really.
The problem is that we’re being sold a lie and it’s hard to swallow.
“Cloud will save you on staffing costs” - no, it just means you have specialists working on proprietary solutions, just like IBM mainframes or storage appliances or load-balancers/traffic-shapers of yore.
“This technology will make rollouts easier” - until it breaks down and is not easy to put together again, you’re at the mercy of your upstream and you had better hope you keep a really solid update cadence and not introduce something you can’t consume later.
“Layers on layers means things are composable” - sometimes, but you need more people to know more things to put it together properly. Running everything from QuickStarts is doomed to fail, but I see so much of it.
Our config management tools back in the day were crummy as hell, cfengine being the only notable one (which was awful), our monitoring systems required a lot more handholding and truthfully: people were not happy to talk through requirements; which was probably why “devs run production” was appealing.
these things have gotten better, but nearly everything else has definitely gotten worse.
If you want to rebut this you need to demonstrate the kind of capabilities we have now thanks to this complexity that we could not have before.
> you handle payments
No you don't. Most people handle payments by integrating with a 3rd party service, most likely stripe or paypal.
Your system today doesn’t even handle left pad, outsourcing even that to a non deterministic tree of geo-distributed modules.
Every complex tool can be configured to work in a simpler manner, and all configuration repositories directing these tools can be organized much better. It's akin to a codebase after all. Many code quality processes and best practices to apply to these, but sysadmins don't like to think like developers, and things get hairy (I'm a sysadmin who does development, I see both sides of the fence).
The sad part is, these allegedly more robust platforms do not provide the transparency neither to developers, nor to sysadmins, nor to users. In the name of faster development, logging/debugging gets awry, management becomes more complex and users can't find the features they have used to have.
Why? We decoupled systems and decided to add this most used feature at a later date, because it's not something making money. Now my account page even doesn't show my e-mail address, I can't change my subscription renewal date, or see my past purchases, why? They're at distant tables or these features are managed by other picoservices which are not worthy to extend or add more communication channels.
Result? Allegedly mature web applications with weird latencies and awkwardly barren pages, with no user preferences or details.
Adding layers and layers of complexity to patch/hide shortcomings of your architecture doesn't warrant a more complex toolset to manage it.
Everything you described was 2005.
* 2005: ship it if it builds.
* 2022: run unit testing, deployment to beta, run integration testing, run UI testing, deployment to preprod, run internationalization testing, run regional/marketplace-specific testing, run consistency checks, run performance tests, deployment to prod.
I distinctly remember buying things on Amazon in 2005. Actually come to think of it, I did everything you list in 2005, at multiple companies.
This just means “use APIs”, adds {base_url, secret_key} pairs to a config file. Where are orders of magnitude devops-wise?
Generally you need solid APM and logging to support the operations of this.
Then there's queueing tasks for later and making sure you never lose messages between systems. All sorts of fun happens when you hit a capacity limit in your queueing system due to unintended/unnoticed design flaw and all of a sudden system performance starts to tank as things aren't getting processed due to too much stuff in the queue that isn't draining at a sufficient rate. You need to figure out is the queue actually draining or is it still increasing in size? Are all the worker nodes still running or did one (or several) of them hang and didn't die but simply isn't processing any jobs out of the queue but still consuming a spot in your auto-scaling group blocking a new healthy instance from coming online? Is it a problem in your software or is it actually a blip in operations from your cloud-provider?
It's way, way, way more complex than simply "chuck some API keys in a file and she'll be right!". That's hobby project level stuff, not realistic ops of a production system.
I’d fight at your side few years ago, but since starting to work in this field and asking around I realized nobody cares about that ideal way, cause it’s so expensive and hard to maintain, and the cost of change is prohibitive. It even left few scars on my mental health, exactly because I was too anxious to go this way after rotating in environments where devops is praised as something inevitable to survive. If you squint at us at the right angle you still can see devops, but I think we are just expressing two opposite sets of ideas, each applicable in its own type of a business env. You may still call that a hobby-level, and I’d even agree (cause it essentially is) if that hobby didn’t bring as much as to sustain and profit the company at the levels many companies would wish they were at. If it works and brings revenue who cares how it’s categorized.
2005: two system administrators wrote all automation using a handful of Bash, Perl and Python scripts AND LEAVE THE COMPANY a couple of years laters
20xx: new hired system administrators continue to rewrite scrips. no shared knowledge because scortched-earth policy in effect
2022: HN: DevOps is a failure -- you just need two system administrators...
Security adds complexity
The very opposite.
> Security adds complexity
If anything, complexity is the enemy of security
And can create more than 2 orders of magnitude in terms of economic value (i.e. profit) as a result.
It did, just not in the way you're describing it. This is how it's manifested itself:
> Apple, Amazon, Alphabet, Microsoft and Facebook all account for just under 20 percent of the market value for the entire S&P 500. With a collective value of nearly $5 trillion, these top tech companies easily dwarf other entire industries in the index, with companies like Berkshire Hathaway and JPMorgan Chase falling well short. Currently, the total valuation of the S&P 500 is almost $27 trillion.
[0] - https://www.statista.com/chart/20794/tech-companies-highly-v...
Then you clearly misread my initial point. economic value != overall economy growth
This!
The primary problems with DevOps are...
1. It is still in its infancy therefore it's changing quickly.
2. Bad (or no) documentation is much more painful than before. A single family house without blueprints can, usually, be adequately serviced; whereas, a 27 story office building cannot.
That code in 2005 was better in every way.
Context: just learning loopback API framework. What a steaming pile of overengineered garbage.
2005 your average box served some php and static assets, connecting to some generic relational database. Reading logs means grepping files over ssh.
2022 your architecture runs in the cloud, has multiple flavors of databases, queues, caches, and so on. You have at least two orders of magnitude more complexity because you aren’t just serving a web page anymore - you handle payments, integrate with other services, queue tasks for later, and so on. Your automation may be an order of magnitude more complex than 2005, but it enables two orders of magnitude more functionality.