Typescript (nodejs) backends are not performant, and your ultra modern stack is going to just cause you issues going forward if you need to scale. Prisma is really bad too, slow queries and no flexibility. If you ever need to do any sort of complex query you will just have to write sql anyway. And typescript is only sort of static typing.
These technologies are great for prototyping and building a v1 release to see if what you're trying to achieve is actually possible, but you will regret it later on.
The reason I know this, I work at a startup where we literally had the same backend stack and its been nothing but preformance issue after preformance issue. And it all needs to be replaced. We would have been better off building everything with go/rust in the first place. Or even java.
1. I agree, Prisma is not a great ORM. Drizzle is a better choice. It's close to the metal and when the abstraction inevitably leaks, it leaks towards the user using raw SQL
2. Modern JS is extremely performant. Look at any of the benchmarks for the new JS runtimes that have come out in the past 6 months (e.g. Bun / WinterJS / etc...). It approaches Go / Java in terms of performance.
3. Even the traditional NodeJS runtime has been optimized out the ass by Google. For example: JSON parsing has highly performant SIMD instructions under the hood. When a trillion dollar company puts billions of dollars behind a technology it will get fast.
4. There is no possible way that building your CRUD backend in Golang / Rust is a "faster" solution than just using React Server Components.
5. The vast majority of startups are IO-bound, not CPU-bound - so "fast" languages like Go / Rust won't be as relevant.
The benefits of Go (for most companies) only apply once your company hits an inflection point and starts to scale and starts to see the throughput that can really take advantage of a lower-level language.
If you're building a high-throughput infra company or something, then the things I've mentioned are less relevant.
>> 2. Modern JS is extremely performant. Look at any of the benchmarks for the new JS runtimes that have come out in the past 6 months (e.g. Bun / WinterJS / etc...). It approaches Go / Java in terms of performance.
I like JS but lets not blow smoke up anyones ass here. Your not picking node or its faster safer cousin bun for server speed. You're picking it because you can keep your engineering overhead to a smaller number. Your picking it because you need to generate pages out of your SPA like app.
> 4. There is no possible way that building your CRUD backend in Golang / Rust is a "faster" solution than just using React Server Components.
Spend a month with SQLC and writing blood and guts API's in go. Your gonna realize that the slight slow down in finishing the feature code is more than made up for in finishing testing, ease of deployment and a million other things that you skip by NOT using a JS backend.
>> The benefits of Go (for most companies) only apply once your company hits an inflection point and starts to scale and starts to see the throughput that can really take advantage of a lower-level language.
This is some holdover thinking from PHP vs JAVA days that isnt true when comparing node/ruby to go/rust ... Im going to leave python out because you may have OTHER reasons (existing bindings to C ...) to head in that direction.
>> If you're building a high-throughput infra company or something, then the things I've mentioned are less relevant.
Maybe this is true. But I worry about the long term stability and paintability of anything written in JS. So much flavor of the month and then poof, off to the next thing, it IS a concern.
I only use TS on the frontend, so less opinionated about some of this.
I’d love to see some of the codebases where people complain about performance so I could profile them myself. Would put money on being able to improve the situation by orders of magnitude without switching stack.
We use Python on the backend of our web app for realtime image recognition and it’s fine because we’ve been thoughtful about data structures and algorithms.
There’s numpy in the middle, so bits are outside python. The vast majority of the code is Python though and the performance gains come from being strategic about the architecture.
My point is that for most usecases you can go a lot further by looking at the code that’s running rather than the language that’s running it.
I think python have the best libraries you can buy, and that makes python a very strong choice if you do not suffer the NIH syndrom or do not need extremely fast code (python is fast enough now).
One big issue i have though: the lack of easy multithreading reduce your possibilities, or at least limit your creativity: you will often choose to use async/await (and sometimes use signals) rather than use producer-consumer designs, which are often the optimal solutions.
Node.js backend performance is comparable to Java. Perhaps you had bad developers - trust me, if they screwed it up so bad they wouldn't be able to make your Java backend any faster, and probably couldn't get Go or Rust to work at all.
Go is perhaps the simplest language to learn there is. It's almost impossible to not get it to work because it has so few things you can actually do.
And nodejs only preforms well in hello world benchmarks, real world applications are nothing like that. Once you start having to manipulate large arrays or do any large amount of math nodejs preformance goes into the dumpster.
> Once you start having to manipulate large arrays or do any large amount of math nodejs preformance goes into the dumpster.
We’re talking about web applications, no? You probably shouldn’t be manipulating large arrays or doing large amounts of math directly in your web application server. That should be isolated in some type of service or worker, which could be written in another language. Or maybe there’s a NumPy-like package for Node.js, I haven’t looked.
The question then is where do you draw that line of using another language? It probably depends on your application, but I think Node.js is perfectly suitable for typical web applications backends.
Can you give an example? I just think that no matter what programming language or framework you're using, if you are querying for giant arrays through an ORM and passing them through an API result to a web app, it's going to be slow.
There are other kinds of solutions for this problem like breaking up the data into chunks and only returning the necessary data. Maybe it's a DB optimization where you can add indexes. Or caching the result of your ORM query.
I've never seen any web app written with any tech that was snappy while making requests for large amounts of data and waiting for it to come back in one big honkin array.
> Web apps deal with fairly big data structures, if nothing due to orms.
Perhaps due to using ORMs unnecessarily and/or inefficiently and failing to drop down to SQL when needed.
It's hard to think about an example, this is very application specific. But there are background jobs and that's where a big array could be manipulated
Probably still need to stream or chunk the data instead of dealing with giant arrays to get decent performance. That's not a language/platform issue.
I recently had to process and aggregate metrics for 5 million rows of user data (a few GBs) on my MacBook with Node.js. By streaming / iterating over the items without loading them all at once it chewed through them all in a few seconds. ¯\_(ツ)_/¯ And it's single-threaded (except for I/O offloaded to threads -- I'm talking about the calculations).
Many companies are not throttled by server language performance. Performance in many cases will depend on your database, geolocation, and caching techniques over nodejs vs go. Nothing wrong with prioritizing developer experience.
Agree. JS/TS is not a good environment for large software projects maintained over a long period of time. I'm not 100% convinced Rust is either, yet. cargo seems to be a dumpster fire. Build tools generate TB of mystery disk bloat. Obsession with async... hopefully these issues can be resolved.
This sounds weird to me. What kind of scale / traffic did you have? Must be incredible read write heavy with millions of users?
I'm saying this because i myself have, and lots i know, have launched production sites with 100s of thousands of users on ready-made stacks like, Laravel, Rails, Flask (Php, Ruby, Python). But it's my impression that these fall short on millions of users and enormous concurrent traffic, but then you're already at huge evaluation, years into your project, or have 200+ positions, ie. you've already refactored your project multiple times.
That's why you shouldn't use slow backend technologies in the first place because you get to the point where you need preformance and it's impossible because your limited by extremely slow runtime you chose initially because it was flashy (not even easier to developer for) when you could have just chosen something better from the start.
Simple crud apps can get by fine with those technologies, but in the future I'd still never use it because you're leaving huge amounts of performance gains on the table for virtually no benefit. I don't buy the argument that javascript is just easier to develop for because it's simply not. The js ecosystem is a disaster.
What kind of application are you talking about? I agree to a degree but we haven't run into performance issues with say 100k users, what traffic did you handle, what was the business case?
Again sounds like you are running some complex math heavy operation like say a weather service, national taxi service with lots of pathfinding, a global gaming platform etc?
Curious to know what exactly you are talking about here? Because normally you just outsource the "heavy stuff" to some remote API, service, etc. you can write in a hyper efficient language anyway.
I just built an app with Rust and Svelte with ~25k LoC, the app was solid, and working on it was a joy. New team lead got hired, wasn't familiar with Rust, said we'd have trouble hiring Rust devs, and threw it all out. We're now building a Python / ReactJS app from scratch.
Speak to your academic friends for advice on how to get him fired. Do this ASAP for the sake of the company and do not drop the ball once he’s gone. You may have to step into his role
For what it’s worth, I’d be much more interested in working for your company using Rust and Svelte compared to python and react.
If you need someone with experience working with Rust and Svelte, I’m looking for work and value making quality decisions like you did. Now you can point to at least one dev who has the skills your team lead was concerned about finding.
I’m not partial to Node.js backends for many reasons, primary being that I prefer batteries-included web frameworks and the choices are very limited in the JS backend world… But I have such a hard time believing that Node.js is inherently non-performant when there are so many counter-examples of large scale applications that do just fine on it. I understand that the single-threaded nature of JS can be a bottleneck for certain workloads, but without further explanation I’m going to concur with others and say that reading your comment, it’s hard to discard you guys made a mess on your own... Which at least in my experience often comes down to how difficult it is to build in Node.js precisely because of how “minimalistic” and “unopinionated” everything tends to be… You’d better know what you’re doing when there’s no prescription or standardized way for anything from queuing systems, to ORMs, file organization and architecture, concurrency outside of web requests and CRUD, etc.
> But I have such a hard time believing that Node.js is inherently non-performant when there are so many counter-examples of large scale applications that do just fine on it.
Like LinkedIn! Except that to make it performant they had to horizontally scale and then restart the server every N hours when they ran out of memory…
Sounds like the issue wasn't the performance of TS per se, but rather the use of some ORM led to inefficient database queries being executed. This sort of screwup is possible in all languages.
i love typescript, but that's still not the same as static typing. imagine a full stack typescript app with a blog type
type Blog = { name: string }
everything works fine until you decide to refactor "name" to title. you update the backend and typescript code and deploy the change
type Blog = { title: string }
user john never closes his browser and leaves your website open. he clicks on a new blog post. his client typescript expects a name, the server gives a title, and crash, "Cannot read properties of undefined".
i still choose nodejs backends, but you obviously have to keep in mind it's not statically typed. same with "any", my coworker could take my blog and modify it. it's preventable, but with large teams and codebases still prone to error.
TypeScript is definitely statically typed. It is not a 100% “sound” type system, but very few languages actually are (Haskell, Reason, etc)
And in your example of loading data from a server, if you have appropriate lint rules set up your editor and linter will warn you when you pass “any” typed data (such as that returned by fetch’s response.json()) to something expecting another type. You can use something like zod or yup to validate the data is the type you expect before passing it the rest of your typed code, thereby containing possible type errors to a known location where you can gracefully handle it.
This is a problem with any language that accepts typed data from an external system, certainly not unique to TypeScript.
typescript compiles to javascript and you run javascript with deno, nodejs or whatever javascript server you choose. Typescript helps while you develop, but at the end is just javascript.
i don't agree at all. nodejs backends are as performant as you build them. the whole premise of node was async io, so if you're doing a bunch of blocking stuff then yeah, you're going to have issues. otherwise there's nothing innately non-performant about the platform.
it's impossible to debate more without going into details on your performance issues. typical backend architecture for any platform these days is the scalable container model, if you wanted scale-to-zero i wouldn't necessarily use node because of cold starts.
prisma can be great, can be slower than writing your own query, but that's the whole point of it. most of the time, it works and you can forget about queries and typings. when it doesn't, you can just eject to raw sql anytime you want.
i'm not a total nodejs fanboy, and have used more jvm (java/kotlin/scala) in my life, but if i was building a web-app today i'd absolutely consider node for all the reasons the author listed
Nodejs backends are not "as performant as you build them" nodejs is slow! This is an undeniable fact. If you ever have to do anything computationally intensive, which every backend at some point will nodejs will become the bottleneck.
‘nodejs is too slow’ is high school level backward rationalization.
‘nodejs is too slow for my use case’ is getting there.
Most stuff we engineer maxes out the database way way waaaay before it saturates its CPU capacity regardless of the language used. You’d better know very well that you’ll be vulnerable to this as a business, otherwise your inefficient competitors will take your money faster than you can build value.
What do you think about how well uWebSockets.js performed in the TechEmpower benchmark? Just js is admittedly experimental, but it has the fourth highest score, and the highest score of any non-Rust framework as well which I found interesting. Elysia (a Bun framework) did pretty well too.
Deno would have probably scored well too (it uses Rust's Hyper crate under the hood), but they're only running a single instance of the server despite Deno supporting the Linux SO_REUSEPORT socket option, which is important because the test is run on three servers with Intel Xeon Gold 5120 Processors that have 14 cores and 28 hyperthreads [0].
Just go to its repo https://github.com/uNetworking/uWebSockets.js and see the uWebSockets submodule dependency that it's written in C++ (90.9%) and C (6.8%)...so sure, a "fair" comparison indeed with vanilla implementation!
Ok...Node is itself written in C++. The V8 and JavaScriptCore JS engines have no concept of a server (they were designed for web browsers), so under the hood JS runtimes implement servers in a lower level language capable of interfacing with Linux syscalls. Bun uses Zig, Deno uses Rust, Node uses C++, etc. I don't think that necessarily makes this an unfair comparison. The database drivers used in these benchmarks are also usually written in C++. Look at the repo for the native Postgres binding for Node: https://github.com/brianc/node-libpq
I wouldn't use find commands beyond the simplest Join. The beauty of Prisma is raw queries, because they're compiled and type-checked against your Entities. That's the whole appeal, IMHO.
"In many ways, I think I picked a great stack to move quickly and ship things. However, considering all the time spent debugging some of issues that stemmed from using more cutting edge infra providers, next time I'll more carefully consider setting up parts of my infra with a more traditional solution like AWS."
It really depends on which parts of AWS you're using. Spinning up a bunch of EC2 instances and treating them like colo'ed servers has pretty much always been boring (other than the excitement of not having drive to the DC when you need to reboot something).
Tend to agree, though I don't always practice it. Boring is also subjective, even if it's a metaphor for mature.
The same sort of arguments could be said for "choose what you know", "choose what you understand", "choose what you can hire for", "choose what you can afford".. the list goes on.
Like most things in life there are no cheat codes.
This article doesn't really help me choose and based on the title I expected it would. I don't think OP made a "hype" driven decision, but this is certainly one of the most popular modern stacks and what I would like to see instead is a comparison between the popular and less popular options and why OP chose exactly this stack.
My bias was always to err on the side of popularity. I thought that I would be less likely to encounter frustrating bugs and edge cases using choices that have been battle tested for a least a few years.
What I overlooked is that while the platform might be popular and mostly trustworthy (such as Vercel), there are specific solutions like the Vercel Cron Job are still relatively new (released Feb 2023) and still not ready for serious use.
Why does 'modern' nowadays mean 'host everything on managed services' like Vercel or Netlify? Why is running your own database so frowned upon? Is *SQL/SQLite so insecure that a default password-protected setup can be pwned within minutes?
Managed hosting can give you a head start, but also increased costs. There are quite a few horror stories when it comes to billing.
> Why is running your own database so frowned upon?
It's a meme that running your own DB means that you'll get owned within seconds of going live, or lose all your data in the first day because fires in your data center are a daily occurence :-/.
It's the modern equivalent of "MongoDB is Web Scale".
> Managed hosting can give you a head start, but also increased costs.
Yeah ... buuuuuuut ... it's a very tiny head start.
Sure, you may save a few hours by using a hosted postgres service, but you save that only once over the entire lifetime of the database server!
IOW, if you set up a PostgreSQL server on a couple of cheap VPS instances, you can continue using that server for each new product until you hit performance/security problems.
I've got exactly one PostgreSQL server set up, and all my little experiments happen on that one server, with me creating new databases as and when needed.
A true background job would be impossible with this stack. Because I was on the Pro plan, in theory I could set the serverless function timeout to 900s.
Definitely would reach for remix before next, and I would go with a big cloud provider container app solution such as Azure container apps (if azure was better) or Google Cloud Run (good, but do you trust google?) before Vercel.
Personally I think ORMs are mostly bad and Prisma is more of the same. Zapatos at least gives you typed SQL results without a lot of overhead.
Next is buggy, Remix mostly just works, and I have zero interest in Vercel which is the Next selling point IMO.
I haven't played with either Drizzle or Kysely, looking at the docs Drizzle in particular looks slimmer and more sql-like which is a plus, though in general I think having an environment like dbeaver or data grip that gives you completions for your queries and developing them there then copy/pasting the code and getting type safety on raw queries with zapatos is still superior for power users who know sql.
Fascinating discussion. What I'm looking for in a tech stack for my web app:
CRUD - simple requirements at first but could be developed over time to include more social/community functions.
Good dev experience, I'm a relative newbie and will be learning on the go.
Inexpensive to host, no unexpected bills.
My tech stack should have the potential to scale nicely, in case it ever needs it.
So I'm not looking for much, just
Easy - Powerful - Future Proof
Author's code is vulnerable to timing attacks and probably others as well.
This is why people should use established backend frameworks such as Laravel/Django/Rails if they really want to focus on their product and not on reinventing the square wheel.
To me this is a valuable write-up because it (a) clearly describes paper cuts accumulated when building with unfamiliar tech and (b) showcases how much nonsense developers are willing to tolerate in the name of "new" or maybe "speed".
Some of the paper cuts described include:
- Inexplicable interaction between the data layer and a hosting service's environment variable
- Migrations that don't work reliably
- Cron jobs that don't work reliably
- Inexplicable caching behavior for serverless functions, requiring an obtuse workaround
- An auth library that makes it hard to... get the ID of the logged-in user, apparently
I dunno, I suppose I'm happy to stick with boring old tech.
Thanks Dave! That's a great summary of my experience.
Most of my experience has been with a more "boring" stack, so I was interested in trying some of the more "hyped" technology choices, to see if I was missing out on anything.
I’m also in a phase where I get to start a lot of new low-stakes web projects, so I’ve made a point of adopting at least one new-to-me tool for each. I love typescript, and am generally happy with modern front-end tooling, but the node back-end ecosystem drives me bonkers — too many paper-cuts from too many tools along the lines you’ve documented here — and has me regularly retreating back to Python or Ruby land.
Pocketbase is just superb. I am using pocketbase together with scalajs and svelte. It saved me so much time setting up the basics and you won't regret it once the project grows. Anything works with pocketbase but go or any js based stack has an easy onboarding.
Do you eventually have to move off pocketbase? Looks like it only supports SQLite, and therefor obviously can only run a single server instance. How do you plan to migrate off it in the case that you have higher availability or traffic needs?
Never happened to me. SQLite is faster than most people think. When you hit the limits of sqlite, chances are high that you can afford to use one of the scaling solutions mentioned above. Here are some benchmarks:
Could you recommend a stack please?
* Testing - test runner, mocking
* Database - Typesafe queries, ideally as little abstraction as possible, migrations
* Web app framework
* Dependency injections (or do you prefer doing it manually over using a library)
Unless you are very lucky, odds are your app is never going to be hitting "production" workloads. If it ever does, then you can pivot to something else (and by that point you'll presumably have the revenue to do so).
From what I can see from the demo that's a super simple application. It might make sense to use a tool such as htmx/unpoly/hotwire/etc.
What I don't get is people just blindly "following the hype" (as with anything else) and assuming it's a replacement for client side frameworks (react, vue, etc) for anything you might need out there. You already said this several times in some Podcasts I listen to, so I don't blame you. But people are just "hey bro just use HTMX and Go and done". And that might be fine if you're a backend developer that doesn't like JavaScript and have a very simple use case and probably don't have high UX standards or picky designers chasing you with every animation, loading state or performance issue.
There are many projects out there with pretty complex UX logic where this will just not work, and if it works the end result is going to be much harder to maintain. Let's not take into account finding people willing to work with these tools (already happened to us where we had to move a Rails/Hotwire project to Rails/Inertia) because no one could work on the frontend, initially created by a backend dev and that became such a mess not even him could maintain it anymore).
But again, it is not the technology to blame (it's great!), it is the usage of it. And people can misuse anything, as they misuse react/vue/etc when they're building just a landing page.
You sound ~unhinged~ (edit: "unhappy" is more like what I was thinking; pardon the negativity) with this unprovoked rant about the reasons you assume people choose a technology.
It would be refreshing to me for someone to suggest we at least evaluate HTMX on a new project before going straight to the Next.js or Remix behemoths where we'll have 45min builds pulling in thousands of NPM dependencies over half a million files in node_modules.
Edit: I was also thinking it would be even more refreshing if people focused on web standards and asked whether or not they actually need a trendy framework.
> before going straight to the Next.js or Remix behemoths where we'll have 45min builds pulling in thousands of NPM dependencies over half a million files in node_modules.
You sound unhappy as well, and you also sound like you're doing something terribly wrong here.
My experience with people following the HTMX hype is backend developers annoyed by JavaScript not being perfect and node_modules size.
node_modules is never that big, and even if it is it's not a problem as big as maintaining complicated (i.e. not trivial) projects with these tools.
At the end of the day it depends mostly on your skills and the requirements for the project you're working on. As I said if you have some UX experience and you care about loading states, page transitions, animations and in general the little details then something built with these tools will be unmaintainable by anyone other than the person that implemented it in the first place.
node_modules too big? so what? that's not all of it shipped to the browser and I have plenty of space on my computer. A mess of html attributes, logic split between jquery, backend, html attributes, etc? Now that's a bigger problem.
And why I said "hype"? Well, because people suddenly see it as the greatest thing ever because some youtubers/podcasters made it popular, when there've been similar solutions for ages and nobody cared (i.e. Unpoly). But if the right youtuber talks about it, then boom! Everyone is saying everything should be done this way from now on.
As I said in another comment, the idea is great and I like it. But as the creator of the library said MANY times: It's not a replacement for component based frameworks. It's a great tool for the simpler use cases where react/vue/etc would be overkill.
My rant is about people acting as this is the future for everything, which again, the author of the library itself says it's not, and he's a smart guy.
I'm interested it HTMX, but for a project like this I couldn't afford to lose out on React's ecosystem of pre-made components where the goal was to ship a professional looking SaaS web-app quickly.
If I wanted harden this setup, my next consideration would be either having a separate microservice that's purely responsible for auth, or using 3rd party provider like Cognito.
if only there was a network stack for it, we could use TempleOS + HolyC as a Lambda in AWS. These things are useless without data or gravitas behind them.
These technologies are great for prototyping and building a v1 release to see if what you're trying to achieve is actually possible, but you will regret it later on.
The reason I know this, I work at a startup where we literally had the same backend stack and its been nothing but preformance issue after preformance issue. And it all needs to be replaced. We would have been better off building everything with go/rust in the first place. Or even java.