useFetch() struggles to handle heavy concurrent requests

[long copper]

I have an App written in Nuxt3 and recently I noticed that my app can't handle heavy concurrent requests. I tested my app with K6 Load Tester and I realized that useFetch is the problem.

I removed data fetching code from Script and the performance was way way higher than before.

Afterwards, I run my API on my local machine and tested again; the results were the same.
(My API responses under 100ms)

My question is: How apps should optimize to handle large number of users?

useFetch() struggles to handle heavy concurrent requests

[rigid hinge]

Concurrent requests as in „initial“ requests?

[long copper]

Yes, I mean when users visit the app for the first time and SSR triggers

[rigid hinge]

that's understandable. You want to use SWR to cache results (even if it is just for a second) and/or multiple instances to share the load

[long copper]

Thank you for your response, I will take your advice to account.

Quick Update: I have change the JS runtime to Deno.js and the results were significantly better.

Can you please send me a reference to follow SWR? How can I implement it with Nuxt?

[rigid hinge]

Yeah where you deploy (eg serverless functions) makes a difference too

[rigid hinge]

Check out Nuxts routeRules for SWR

[night girder]

@long copper how are you deploying?

[long copper]

Thank you so much 🙂

[long copper]

It's on a Linux machine, up on a docker container.

[night girder]

a single instance? or clustered?

https://pm2.keymetrics.io/docs/usage/cluster-mode/

[long copper]

Actually I noticed Deno handles clustering by itself, I had many instances with different process IDs on my machine.

[night girder]

have you tried bun? 🙂 it's very fast

[long copper]

Yes I did 🙂
Don't know why, but the performance was not good, It was close to Node.js

[night girder]

interesting, how were you running it?

[long copper]

I installed packages with bun package manager and built it with 'NITRO_PRESET=bun' present. Finally, It runs with
'bun run ./.output/server/index.mjs'

[long copper]

Test results with bun

Test results with Deno

[night girder]

well yeah if deno is clustering itself

bun is much faster if you look at the numbers outside of the duration/waiting, which is where clustering would have an effect

[long copper]

Do you have any Idea why useFetch holds the app from respinding? The API responses under 100ms but Nuxt holds the requests for 2 or 3 seconds?

[night girder]

how much load are you throwing at how many instances?

[long copper]

4000 requests at 6 instances in 90 seconds

[night girder]

so around 8 requests per second per instance (~45 requests per second)

what are the host machine specs?

[long copper]

I tested on my local machine with this Specs:

CPU: Core i7 8700
RAM: 16G
Storage: SSD M2

The process used 20% of CPU and 2GB of RAM