[hsn915]

It's way too complicated. But if this is all you have ever seen and if you've been designing such systems for a decade, this seems like normal to you.

Here's an alternative stack that can handle over 99% of websites:

- Self contained executable

- One-file database

- Cache is memory

- Text search is a library function

- Indexing is a library function

- Serving http is a library function

Such a stack can handle > 20k concurent connections (per second). The code doesn't need to be "optimized"; just non-pessimized.

You can scale "vertically" for a very long time, specially in 2020 and beyond, where you have machines with over 64 CPU cores. That's almost a cluster of 64 machines, except in one single machine.

If you _must_ scale horizontally for some reason - maybe you are Twitter/Facebook/Google - then you can still retain the basic architecture of a single executable but allow it to have "peers" and allow data to be sharded/replicated across peers.

Again all the coordination you need is just a library that you embed in the program; not some external monster tool like k8s.

[kgeist]

There are several reliability issues:

  1) a single panic/exception/segfault in the executable brings down the whole website and so it will be unavailable until the executable restarts

  2) entropy *always* increases (RAM usage, memory corruption, hardware issues, OS misconfiguration etc.) so eventually the application will break and stop serving traffic until it's repaired/restarted (which can take time if it's a hardware issue)

  3) deployments are tricky if there's nothing before the executable (stop, update, restart => downtime)

  4) if cache is in-process, on a restart it will have to be repopulated from scratch, leading to temporary slowdowns (+ and maybe a thundering herd problem) which will happen *every time* you deploy an update

I think much of it is ignoreable if the site is just a personal blog or a static site. But if the site is a real time "web application" which people rely on for work, then you still need:

  1) some kind of containerization, to deal with inevitable entropy (when a container is restarted, everything is back to the initial clean state)

  2) at least two instances of the application: one instance crashes => the second one picks up traffic; or during rolling updates: while one instance is being killed and replaced with a new version, traffic is routed to another instance

  3) persistent data (and sometimes caches) need to be replicated (and backed up) -- we've had many hardware issues corrupting DBs

  4) automatic failover to a different machine in case the machine is dead beyond repair

>not some external monster tool like k8s

What can you use instead of k8s for this kind of scenario? (an ultra reliable setup which doesn't need a whole cluster)

[dgroshev]

It seems to me that people tend to vastly overestimate their uptime requirements. "Real time 'web application'" used by hundreds of millions of people can be down for hours and yet succeed wildly, just look at Twitter, both its old failwhale and new post-Musk fragile state. Complexity, on the other hand, and thus lower iteration speed and higher fixed costs can kill a business much easier than a few seconds of downtime here and there.

You don't need an "ultra reliable setup" or even a "cluster". You can have one nginx as a load balancer pointing at your unicorn/gunicorn/go thing, it's very unlikely to ever go down. You can run a cronjob with pgdump and rsync, in an off chance your server dies irrecoverably corrupting the DB (which is really unlikely for Postgres), chances are your business will survive fifteen minutes old database.

Most "realtime web applications" are not aerospace, even though we like to pretend that's what we work on. It's an interesting confluence of engineering hubris and managerial FOMO that got us here.

[srehell]

> It seems to me that people tend to vastly overestimate their uptime requirements. "Real time 'web application'" used by hundreds of millions of people can be down for hours and yet succeed wildly

That may be true for social media apps where the Terms of Service don't include any SLAs/SLOs, but if you're a SaaS company of any kind, the agreements with clients often include uptime requirements. Their engineers will often consider some form of "x number of nines" industry standard.

[kgeist]

In the projects I work on, things go down all the time, for various reasons (hardware issues, networking problems, cascading programming errors). It's the various additional measures we have put in place which prevent us from having frequent outages... Before the current system was adopted, poor stability of our platform was one of the main complaints.

I agree that for many projects it may be an overkill.

[dgroshev]

Networking issues and even hardware issues are very unlikely if you can fit everything into one box, and you can get a lot in one box nowadays (TB+ RAM, 128+ core servers are now commodity). MTBF on servers is on the order of years, so hardware failure is genuinely rare until you get too many servers into one distributed system. And even then, two identical boxes (instead of binpacking into a cluster, increasing failure probability) go a very long way.

It's a vicious circle. We build distributed multi-node systems, overlay software-configured networks, self-healing clusters, separate distributed control planes, split everything into microservices, but it all makes systems more fragile unless enough effort is spent on supporting all that infrastructure. Google might not have a choice to scale vertically, but the overwhelming majority of companies do. Hell, even StackOverflow still scales vertically after all these years! I know startups with no customers who use more servers than StackOverflow does.

[hsn915]

Re: Crashes.

If there's a bug that brings the server down, it will happen in all instances and repeatedly no matter how many times you restart. Specially when the users keep repeating the action that triggered the crash.

Re: Entropy. Entropy increases with complex setup. The whole point of not having a complex setup is to reduce entropy and make the system as whole more predictable.

Re: caches. There are two types of caches: indicies that are persisted with the database, and LRU caches in memory. LRU caches are always built on demand so this is not even a problem.

Plus modern CPUs are incredibly fast and can process several GBs of data per second. Even in the worst cases, you should be able to rebuild all your caches in a second.

[kgeist]

>If there's a bug that brings the server down, it will happen in all instances and repeatedly no matter how many times you restart.

Not necessarily so. Many bugs are pretty rare bugs which are triggered only under specific conditions (a user, or the system, must do X, Y, Z at the right moment). So it doesn't happen all the time. But when it happens, the whole server crashes or starts behaving in a funky way and other users are affected. Sure you may say if it's a rare bug, then users will be rarely affected. But we don't have a single bug like that, there's always N such bugs lurking around (we never know how many of them in a large application); multiply it by N bugs and you have server crashes for different reasons quite often, making your paying customers dissatisfied. It also assumes you can fix such a bug immediately while it's not always true, there's often Heisenbugs it takes weeks to root out and fix, while your customers are affected (sure the application will restart but ALL users (not just the one who triggered the bug) can loose work, get random errors when the app is not available -- not a good experience). So having several app instances for backup allows to soften such blows, because there will always be at least one app instance which is available.

>Entropy increases with complex setup. The whole point of not having a complex setup is to reduce entropy and make the system as whole more predictable

I agree that entropy increases with complex setup, but there's also base entropy which accumulates simply because of time (which I think is more dangerous). Like make a sufficient number of changes to the setup of your application (which you often need if you release often) and eventually someone or something somewhere will make a mistake or expose a bug somewhere, and you will need to repair it and you won't be able do it easily because your setup is not containerized which would allow to return to the clean state quite easily with no effort. We've had issues like that with our non-containerized deployments and it's a very complex and error-prone undertaking to do it flawlessly (no downtime or regressions) compared to containerized deployments.

>Plus modern CPUs are incredibly fast and can process several GBs of data per second. Even in the worst cases, you should be able to rebuild all your caches in a second

Hm, usually caches are placed in front of disk-based DB's to speed up I/O, i.e. it's not a matter of slow CPU's, it's a matter of slow I/O. Rebuilding everything which is in the caches from DB sources is not super fast.

[the8472]

> and you will need to repair it and you won't be able do it easily because your setup is not containerized which would allow to return to the clean state quite easily with no effort.

Automated deployment including server bringup is orthogonal to using containers or hot failover. For example at $WORK we're deploying Unreal applications to bare metal windows machines without using containers because windows containers aren't as frictionless as linux ones and the required GPU access complicates things further.

[prmoustache]

Note that you can totally have more than one instance of the same app/binary running on the same machine. You don't even need containers for that.

[zerd]

But then you need some kind of load balancer, which hsn915 said was "too complicated".

[the8472]

Upfront customer requirements often say they want >99.5% uptime (which allows for 3.5h downtime a month anyway) or some such. In practice B2B customers often don't care much if hour-long downtimes happen every week during off-hours. Sometimes they're even ok when it gets taken down over a whole weekend. Things serving the general public have different requirements but even they have their activity dips during the late night where business impact of maintenance is much lower.

[satvikpendem]

> 2) entropy always* increases (RAM usage, memory corruption, hardware issues, OS misconfiguration etc.) so eventually the application will break and stop serving traffic until it's repaired/restarted (which can take time if it's a hardware issue)*

This is not what entropy means. Even if you constrain it to hardware, there is no reason to think that this will happen eventually, unless your timeline is significantly long.

[yobbo]

Also, there are typically multiple processes. A panic stops only one process.

[porker]

> - Text search is a library function

What text search will provide me with the same features as Elasticsearch? Index time analysis, stemming, synonyms; search time expansions, prefix matching, filtering and (as a separate feature) type ahead autocomplete?

I would love to never touch another Elasticsearch cluster so this is a genuine question.

[hsn915]

What about any of this prohibits it from being a library?

https://lucene.apache.org/core/

This is the Java library that ES is based on. Without even having to look at it I can make the following judgement:

It should be easy to port to any language.

It's open source, and it's Java. Java has no special features that makes it impossible or particularly difficult to replicate this functionality in any other compiled language, like C, Rust, Go, or any other language that is not 100x wasteful of system resources.

[porker]

> This is the Java library that ES is based on.

Based on, but Elasticsearch is not just a server wrapped around the library. Features ES has are not in Lucene, otherwise anyone could release a competitor by wrapping the library.

> It should be easy to port to any language.

You win the "Most Hacker News comment of March 2023" award. This thread is talking about less effort, and you bring up porting Lucene to another programming language.

[lmz]

I thought it was already ported to other languages eg. https://clucene.sourceforge.net/

Not sure about feature parity though.

[prmoustache]

> Based on, but Elasticsearch is not just a server wrapped around the library. Features ES has are not in Lucene, otherwise anyone could release a competitor by wrapping the library.

Those competitors exist.

[tm-guimaraes]

Go is not less wasteful than java, both are garbage collected and their memory pressure depend highly on the given workload, and the runtime of the program. But java allow more GC tuning and even different GCs for different use cases (ie: shenadoah and ZGC favor very low latency workloads, while the default G1GC favors throughout (not that simple, but you get the point))

Regardless, Java/Go tier of performance is good enough for this kind of thing.

[hsn915]

I was referring to Ruby/Python when I said 100x wasteful languages.

[re-thc]

Problem is it doesn't support HA. You're stuck on that single server model. Upgrades always = downtime = painful. You're also missing things like self-healing and your Lucense index can corrupt.

Real world experience says better to move away from it e.g. lots of self-hosted Atlassian instances over the years. Lucene was a major pain point.

[snikolaev]

Manticoresearch provides mosts of the listed features.

[porker]

Thanks for the reminder. Manticoresearch is an alternative I haven't tried yet. I tried the hip alternatives (Melisearch, Typesense) in autumn 2022 and both were severely lacking for CRM workloads compared with ES.

[hardware2win]

>- One-file database

SQLite?

If yes, then I dont really believe that you can have 20k concurrent users where significant part goes to db, not cache.

But Ive been messing with just 1 vCore, so.

[hsn915]

You can always put an LRU cache between you and SQLite.

I personally moved from SQLite to a B-Tree based key-value store, and most requests can be serviced in ~500us (that is microseconds). I don't mean a request to the database: I mean a request from the client side that queries the database and fetches multiple objects from it. In my SQLite setup, the same kind of query would take 10ms (that is 20x the time) even _with_ accelerator structures that minimize hits to the DB.

But you can always scale up vertically. You can pay $240/mo for 8 vCPUs with 32GB of RAM. Much cheaper than you would pay for an elastic cloud cluster.

[osigurdson]

>> ~500us (that is microseconds).

500us is slow. This kind of performance does not remotely obsolete an LRU cache (main memory access is ~5000X faster).

500us is essentially intra-datacenter latency. Obviously your data is in memory on the B-Tree server as there is no room in this budget for disk IO. Postgres will perform just as well if data is in memory hitting a hash index (even B-Tree probably). I don't think the B-Tree key-value store you mention is adding much. Use Redis or even just Postgres.

[malborodog]

When you say text indexing and serving http are library functions, what do you mean? Also, is the language here go or what? Since you said python is too slow and then necessitates all the infra to manage it.

[hsn915]

Go or any language that actually gets compiled down to machine code to get executed directly on the hardware, and where libraries are compiled into the final product.

When I say something is a library function, I mean you just compile it into your code. In your code, you just call the function.

This is in contrast to the current defactor practice of making an http request to ask another program (potentially on a different machine) to do the work.

[malborodog]

Beautiful. Got it; thank you.