I know that Lemmy is open source and it can only get better from here on out, but I do wonder if any experts can weigh in whether the foundation is well written? Or are we building on top of 4 years worth of tech debt?
I know that Lemmy is open source and it can only get better from here on out, but I do wonder if any experts can weigh in whether the foundation is well written? Or are we building on top of 4 years worth of tech debt?
It’s decent, but it isn’t scalable, at least not yet.
Right now the entire Lemmy backend is one big “monolith”. One app does everything from logins and signups to posting and commenting. This makes it a little harder to scale, and I wouldn’t be surprised to see it split out into multiple micro services sooner rather than later so some of the bigger instances can scale better.
I’d love to know where the higher level dev stuff is being discussed and if they’ve made a decision on why or why not microservices.
There’s no reason that a monolith can’t scale. In fact you scale a monolith the same way you scale micro services.
The real reason to use micro services is because you can have individual teams own a small set of services. Lemmy isn’t built by a huge corporation though so that doesn’t really make sense.
I disagree that it being a monolith is immediately a problem, but also
This is just not true. With microservices, it is easy to scale out individual services to multiple instances as demand requires them. Hosting a fleet of entire Lemmy instances is far more expensive than just small slices of it that may require the additional processing power.
What microservices would you split Lemmy into? The database, image hosting and the UI are already separate.
Well, I’m going to start by repeating that I don’t necessarily agree that it being monolithic is necessarily a problem right now.
The immediate thought in my mind would be all of the federation logic. That’s where all of the instances seem to be lagging behind, and it seems the common fix is “just increase the workers to one billion”. Apparently that does something meaningful, but the developer in me wants to know how a few cores can put so many workers to use.
Spinning federation off into a microservice means you could deploy it on something like Cloud Run or AWS ECS, and have it autoscale as the workload demands it. Seems like a pretty prime candidate to me.
To me this sounds like a code / DB problem more so than a monolith vs microservice issue. You can totally run only the worker part of a monolith inside AWS ECS and have it autoscale, this is not specific to microservices.
I’d split it out into a few systems
The ultimate goal, and I don’t know how possible it is with rust, would be to have a way to run those as individual services or as one part of a larger monolith. Smaller instances would be able to run it as one binary, while larger instances like Lemmy.world or Lemmy.ml can run each part independently. That would allow easier scaling on large instances while (hopefully) leaving it just as simple to deploy on a small scale too.
There’s no reason to split it into 100 different services, but 4-5 mid sized ones might help.
Lemmy’s backend is native code, not run in a virtual machine or interpreted from text like most backends (Java and PHP are still so, so popular, as is Python). You’re not going to pay much extra for an extra megabyte or 10 of RAM being used per instance for the extra code sitting idle. It certainly shouldn’t use much processing power when not in use.
I’d be less concerned with memory (of which Lemmy seems to use very little), and much more concerned with CPU core count. I touched on it in my other comment, but I don’t understand how a few cores is supposed to handle the ridiculous number of federation workers people are setting their instances to.
Does code that exists but isn’t being executed often actually impact CPU usage that much?
Linux kernel has like 30 million lines of code and you probably compile a fifth of it into your actual kernel binary by default. That’s still several million lines, but it doesn’t use much CPU at all. Rather, eliminating all the excess code keeps your size down so you can load it from disk in 0.0001 seconds instead of 0.0002.
Now if code that doesn’t need to run often, runs more often because we scaled the monolith horizontally - that IS a problem, but it’s not a problem inherent to the monolith design pattern, but rather a specific instance of bad design.
Again, my knowledge of the Lemmy codebase is very small, and we could possibly host the monolith in microservices style. The point I am making is this (when it comes to scaling monolith vs microservices):
If the federation logic were split out, we could configure it to run on super tiny docker instances on Google Cloud or AWS. Any time we needed it to, it would autoscale to handle the traffic. The configuration for these dockers could be super minimal memory, no storage, and multiple weak CPU cores. This would be super affordable while still being able to handle as much traffic from federation as we ask it to. One of the cool things with Google Cloud Run is that it handles load balancing between docker containers for you (just point the federation traffic at the necessary URL)
IF Lemmy has things like background services, scheduled tasks, etc, this would significantly muddy the water (we would need each service to be able to handle being run on a multitude of instances, or we would need to be able to disable each one instance by instance). And if we just scaled by spinning up more instances of Lemmy, we would also need to ensure that only federation traffic is heading to the weaker instances that we spun up for such purpose, or we would need to ensure that each spun up instance has enough resources to handle federation traffic along with the main application.
I feel like I need to state once more: I don’t necessarily think Lemmy needs to move to microservices. Only that scaling monolith vs microservices is not necessarily the same.
I suppose that’s completely fair - async workers tend to fare well as standalone services and are often split off even in monoliths. But I guess what I’m saying is that splitting it might not actually win you THAT much compared to just scaling the whole thing. Not until we’re talking like 100 runners to 1 API instance or something. It gives you a bit of additional flexibility, but won’t necessarily be a huge difference in total resource cost, is what I’m saying. But it is still a good idea because it results in cleaner code and, as outlined before, tinier docker images.
Also the thing about Google Cloud Run is that it’s probably not a good idea for many instance owners. Autoscaling can lead to unexpected costs if set up by an amateur. But that’s an unrelated can of worms.
You can easily scale a monolith. You typically horizontally replicate any web server (monolith or not) to handle whatever traffic you’re getting. It shouldn’t really matter what type of traffic it is. Plenty of the world’s biggest websites run monoliths in production. You know how people used to say “rails doesn’t scale”? Well they were wrong because Rails monoliths are behind some huge companies like GitHub and Shopify.
The lemmy backend is also quite lightweight and parallel so it’s cheap and effective to replicate.
In my professional experience microservices are usually a dumpster fire from both the dev perspective and an ops perspective (I’m a Site Reliability Engineer).
This isn’t really contradictory to what I said. I only wished to express that the two don’t scale exactly the same (in terms of execution)
You definitely can’t scale a monolith the same way you can scale a micro service
You can easily scale a monolith. You typically horizontally replicate any web server (monolith or not) to handle whatever traffic you’re getting. It shouldn’t really matter what type of traffic it is. Plenty of the world’s biggest websites run monoliths in production. You know how people used to say “rails doesn’t scale”? Well they were wrong because Rails monoliths are behind some huge companies like GitHub and Shopify.
The lemmy backend is also quite lightweight and parallel so it’s cheap and effective to replicate.
In my professional experience microservices are usually a dumpster fire from both the dev perspective and an ops perspective (I’m a Site Reliability Engineer).
I can’t say I disagree… Poorly implemented microservice architecture is the bane of my existence. Well implemented, though, and it makes my job so much easier.
Granted, my SRE team has all public facing production infrastructure built using an IAC process, if something causes too much trouble, it’s easier to quarantine and rebuild the offending node(s), and can be complete in under 10 minutes.
The biggest problem is far too many developers ignore the best practices and just shift existing code into smaller services. That will never give you either performance or stability benefits. Honestly, it will probably make any issues worse. Microservice architecture is a huge shift in thinking. The services need to be fairly independent of each other to really make any gains. To get to that point will always take a whole lot of work. That being said, there is nothing inherently wrong with some monoliths, but the benefits of splitting out as much of the higher traffic and resource intensive work should never be overlooked.
Microservices aren’t a silver bullet. There’s likely quite a lot that can be done until we need to split some parts out, and once that happens I expect that federation would be the thing to split out as that’s one of the more “active” parts of the app compared to logins and whatnot.
Definitely not a silver bullet, but should stop the app from locking up when one thing gets overloaded. I’m sure they have their reasons for how it’s designed now and I’m probably missing something that would explain it all.
I’m still not familiar enough with how federation works to speak to how easy that would be. Unfortunately this has happened all as I’ve started moving and I haven’t gotten a chance to dive into code like id want to.
It’s also not the only solution for high-availability system. Multiple monoliths with load-balancing can be used as well.
Also, a lot of people are self-hosting. In this case, microservice won’t give them any scaling benefit.
The problem with scaling monoliths is you are scaling everything, including the pieces that have lower usage. The huge benefit you get from going to micoservices is you only have to scale the pieces that need to be scaled. This allow for horizontal scaling to use less compute resources. It also allow for these compute resources to be spread out as well.
A lot of the headaches can be removed by having an effective CI/CD strategy that is completely reusable with minimal effort.
The last headache would be observability. There you’re stuck either living with the nightmare of firefighting problems with 100 services in possibly 10 locations, rolling your own platform using FOSS tools or spending a whole lot of money on something like honeycomb, datadog or new relic.
But I’m an SRE, I live my life for scalablability and DevOps processes. I know I’m biased.
I find that often the overhead from microservices is worse than any savings from dropping a megabyte worth of unused machine instructions from a binary.
When your microservices need to talk to each other (and I’m not sure how many services you could split out of Lemmy without them needing to talk to each other), you’re doing a bunch of HTTP requests that are way slower than just calling another function in your monolith.
I see this at work every day. We run a distributed monolith because someone thought microservices would be a good idea, but we can’t actually separate everything, so it’s usual for an incoming API call to make 2-3 more calls internally. It can get so, so slow.
Overly chatty micoservices are definitely an issue.
Changing your mindset to a microservice oriented architecture is not an easy feat, it’s something that took a lot of time for me to fully grasp (back in my architect/developer) days. Yes, you gain overhead that will need to be compensated for. But when do the benefits outweigh the disadvantages?
Here are some questions to ask during design: How much of this chattiness is because you are tightly coupling these services? Hom much should a microservice be talking between each other? Can you implement an event bus to handle that chatter between services?
Designing an application using microserves but just replicating the monolith application will give you scalability, but will not give resilience. What can you do to overcome that single point of failure? First, no more synchronous calls to these APIs, toss an event over then fence and move on. Degrade your application if the failure is something you can’t overcome, but don’t just stop the application because one API is no longer responsive.
Do you need everything to be a microservice? Probably not. The first thing you look at when moving from a monolith to microservice architecture is what makes the most sense to be moved. How much work can be offloaded to background jobs (using something like sidekiq)?
How do you handle installs? How many packages do we now have to create for this application to work?
There are a lot of questions that have to be answered before moving toward a microservice architecture. On top of that, there is a complete mindset change as to how the application works that needs to be accomplished. If you design your microservice application with a monolith application mindset, you’ll never realize any of the gains by making the move
I’m not into this topic, but i do want to share a talk by Joe Armstrong (creator of erlang).
https://www.youtube.com/watch?v=cNICGEwmXLU
The whole chanel is full of stuff like this.
Scaling monoliths still works fine though. Microservices are first and foremost an answer to an organizational problem, not a technical one. There is a very high chance that if you are doing microservices with less than 20 people, or let’s say even 50 people, you are doing it wrong.
Microservices introduce a ton of overhead in engineering effort required, which needs to be balanced with the benefit they provide.
Scaling shouldn’t be the first and only reason for doing microservices.
deleted by creator
The admin of lemm.ee is currently scaling things horizontally - it is possible.
Do you have more details about this? Lemmy and lemmy-ui are stateless and can be easily scaled horizontally, but pictrs is not stateless and use a filesystem-based database (sled) with lock and it can only be run as a single replica or it crash (if the replica run in the same host and can’t acquire the lock) or will have severe data inconsistency issue (if the replica run in separate host with separate sled database file).
I’m a software engineer myself, but not familiar with your field. How would your practice be applied to self-hosting? I’m assuming a bunch of people with their home servers wouldn’t want to just run OpenShift.
Personally, I wouldn’t touch OpenShift. As someone that has a kubernetes cluster hosted at my house on a mixture of RPis, a nas and in VMs, I’m not one to to say what anyone else would do :).
But, that can be overcome, it’s all about designing you application for multiple different installs. You don’t have to have all your services running fully separately. You can containerize each service and deploy to an orchastration engine such as kubernetes or docker swarm, or you can have the multiple endpoints on a single machine with an install package that keeps them together. It’s all about architecting toward resiliency, not toward a pattern for no other reason.
Also, Google has some very good books on SRE for free.
I’ve been wanting to check out Kubernetes! Thank you!
Microk8s or k3s are the implementations I’d start with. Easy install and gets you started with something without a lot of the work.
Microservices can oftentimes cause more performance issue than they solve, as soon as they need to start talking to each other. Here’s someone with more experience than me explaining how it often goes wrong.
There’s nothing stopping you from putting a load balancer and running multiple instances of a monolith connected to one database. Then the database will also become a bottleneck, but that would still happen with microservices.
Exactly, and nothing prevent a monolith from doing vertical slicing at the database level as long as the monolith is not crossing its boundaries. This is the only scaling part that is inherent to microservices. If the issue is the horizontal scaling, microservices doesn’t solve anything in this case.
Also specifically on what I understand of the Fediverse, you want something easy to host and monitor since a lot of people will roll out their own instances which are known issues when running microservices.
This is a discussion I’m also interested in. Migrating a monolith to microservices is a big decision that can have serious performance, maintainability and development impact.
Microservices can be very complex and hard to maintain compared to a monolith. Just the deployment and monitoring could turn into a hassle for instance maintainers. Ease of deployment and maintenance is a big deal in a federated environment. Add too much complexity and people won’t want to be part of it.
I’ve seen some teams do hybrids. Like allowing the codebase to be a single artifact or allowing it to be broken by functionalities. That way people can deploy it the easy way or the performant way, as their needs change.
That’s what I’m thinking. Microservices could be a huge pain in the ass, but a hybrid approach would make things much better. Smaller instances wouldn’t be a problem, but the larger instances would be able to separate out components.
To keep it possible to run monolithicly would probably need a lot of work, but it’s possible to do and would probably be the best approach.
Gitlab is a great example of a piece of software that has multiple deployment strategies like that.