[hakunin]

You're saying that NodeJS monolith comes with no downsides compared to Rails at early stages. We disagree here.

From my brief surveys, nodejs ecosystem comes with less security out of the box, less standardized project structures, fewer thoroughly-designed and supported packages (vs cutting edge experiments), more complex upgrade paths, more projects getting abandoned, all of which can slow you down every day. Friction from these can be deadly.

On the other hand, being in the corner at scale doesn't force you to rewrite everything. Just the piece that put you there. You can leave Rails app to handle most things, and extract parts of specialized infra as needed.

Would be curious to look at specifics — what kind of pains your company is experiencing with Rails today.

[jupp0r]

"being in the corner at scale doesn't force you to rewrite everything. Just the piece that put you there. You can leave Rails app to handle most things, and extract parts of specialized infra as needed."

You can't easily, because this now puts a network boundary between your Rails app and the part you extracted. A network boundary that requires you to do IO, which will (by default) cause you Rails app to blockingly wait for a response (vs not doing so in other runtimes, ie Node).

"Would be curious to look at specifics — what kind of pains your company is experiencing with Rails today."

Some things I've seen in the last 3 month:

* Rails timing out after 30s while allocating 500MB of memory (mostly) in ActiveRecord to compute 5MB of JSON to return to an API caller

* 90% of request latency of ~10s spent waiting for downstream services to respond to requests. Most of these could be fired off concurrently (ie `Promise.all` in node). 9s/10s this Rails worker is sitting around doing nothing and eating up ~300MB of memory.

* trying to extract out Authorization to a centralized service (so that other extracted services don't have to call into the monolith in order to make authorization decisions) is a major pain as the monolith now has to make calls out to the centralized auth system to in order to make authz decisions.

[hakunin]

A internal network boundary is probably worth it for heavy jobs, since you usually don't want it to interfere with serving web requests (no matter the tech).

You probably already know what I would say to each of those examples.

> Rails timing out after 30s while allocating 500MB of memory (mostly) in ActiveRecord to compute 5MB of JSON to return to an API caller.

I can make a JS or Go program perform the same way. In fact the exact same thing happened in my shop with Go/Gorm. The key question is: how do you compute the 5mb of JSON? The devil is in those details. We changed the way we computed ours, and the issue was gone.

> 90% of request latency of ~10s spent waiting for downstream services to respond to requests. Most of these could be fired off concurrently (ie `Promise.all` in node). 9s/10s this Rails worker is sitting around doing nothing and eating up ~300MB of memory.

This sounds broken. Why is the worker doing nothing for 9 out of 10s? But like I said earlier, there are a bunch of ways to use HTTP1.1 pipelining to run them concurrently. (https://github.com/excon/excon and https://github.com/HoneyryderChuck/httpx support it, but you can also do that with Net::HTTP I believe) And you can still start threads, which are still concurrent while blocking on IO.

> trying to extract out Authorization to a centralized service (so that other extracted services don't have to call into the monolith in order to make authorization decisions) is a major pain as the monolith now has to make calls out to the centralized auth system to in order to make authz decisions.

This seems unrelated to Rails. Not sure why monolith can't continue handling authorization.

[gneray]

> This seems unrelated to Rails. Not sure why monolith can't continue handling authorization.

Agreed. You can totally keep some data in the monolith and some data in new services, and stitch them together if/as needed: https://www.osohq.com/post/distributed-authorization

[jupp0r]

"I can make a JS or Go program perform the same way. In fact the exact same thing happened in my shop with Go/Gorm. The key question is: how do you compute the 5mb of JSON? The devil is in those details. We changed the way we computed ours, and the issue was gone."

The problem is ActiveRecord in my case. The data layout is not great (lots of joins through relationships, I think 12 tables or so). ActiveRecord objects are HUGE compared to the few bytes of actual data they hold.

What do you use (except raw sql) in Ruby if you cannot use ActiveRecord? There is no other ORM that's optimized for fast reads and I don't feel like writing one.

I actually reimplemented the same API endpoint in Go using https://github.com/go-jet/jet and measured 10MB of allocations and essentially zero overhead over queries itself, a 50x speedup.

Don't get me wrong, this is not what the typical Rails shop will deal with, but it definitely shows where limitations of Rails lie and I'm dealing with stuff like that on a weekly basis in my job, and it's not even a large Rails app (3M LOC).

[hakunin]

I agree that if you make a very complex query in ActiveRecord, it could eat a lot of resources.

That said, I would highly recommend going with raw queries for this sort of complexity, no matter the language. There are usually 2 kinds of queries: normal ORM-powered CRUD operations (which can get moderately complex), and hairy specialized report-style calculations. The latter I always recommend to keep in raw, well-written, well-commented SQL form. You can still wrap it into some nice object.

You could write them in something efficient like Jet or Elixir's Ecto, but for such a complex case I'd argue that you shouldn't obfuscate SQL at all. For all other cases ActiveRecord works well.

If you are serving these results in real time, something like materialized views (in postgres) would move the burden of calculation to when data changes, rather than when data is viewed.

And to tie it back to the original convo: a very efficient concurrent language doesn't solve these root causes, rather gives you more time not to address them, and allows you to get away with more neglect. There's some value in that, but you have to weigh it against the downsides mentioned in previous comments. If the language+framework is super efficient and has no downsides to its ecosystem and ergonomics, then there's no debate, I'd just use that.