wezm.net/v2/content/posts/2024/tiny-cdn.md

+++ title = "Tiny CDN" #title = "Overengineering my low traffic website" #title = "Running a tiny CDN for $3/m" #title = "Deploying my Rust web application around the world" date = 2024-11-22T08:55:40+10:00

[extra]

updated = 2024-07-26T09:58:10+10:00

+++

• Linked List is a Rust application
• I track how long it takes to generate a response. I.e. from request received to response generated
• On my Vultr VM this took 1ms or less, however some pages like the home page would sometimes take multiple milliseconds, I felt I could do better.
• The application loads the content into memory and renders on the fly for each request
• I added caching so that the body is only generated once, then served for future requests
• Response generation time now averages about a third of a millisecond 323µs
• I was satisfied with that until I saw the charts in updown.io
• Unsurprisingly with my server in Australia, the actual times people would experience when vesting could be a lot higher. [image]
• I could have slapped Clouflare or Fastly in front of the site and called it a day but a. that's no fun b. I'd prefer to avoid giving them (especially Cloudflare) even more of the Internet's traffic, no matter how miniscule.
• Based on basic stats from GoatCounter the bulk of my visitors were coming from the US or Europe, so I wanted to have the site running in those locations in addition to AU
• Being a native exeutable the application is extremely lightweight so I did some research on bargain basement VPS's on LowEndBox and found a provider that offered KVM based VMs, custom ISO support, and had US and European data centres: RackNerd.
• I created a server in New York, and another in France, they each cost about US$12 per year — probably still more than Cloudflare or Fastly but pretty cheap for a MV with 1Gb RAM.
• I did some initial testing to verify that responses from these servers were in fact going to improve things (because they were so cheap I wanted to make sure the network and underlying hardware was not over provisioned)
• I then installed Chimera Linux on each of the two new servers using my pyinfa install scripts
• todo: chimera resource usage RAM and disk
• I wrote more pyinfra code to install and configure other stuff that they needed like nginx
• I defined a cports template for the linkedlist binary so that it could be installed as an apk package on the system, with dinit config to manage the service
• SSL certs proved to be a challenge as the all needed a copy of the certs. I explored various options here, it's common problem with a bunch of hosted and self-hosted solutions. But all seems too complicated just for syncing two files between servers
• I use lego to manage certs from Let's Encrypt
• I wasn't super keen on the idea of one of the app servers needing to know about and push the certs to the others
• In the end I revived a fanless Qotom Mini PC https://qotom.net/product/29.html, again using my Chimera install scripts to set it up followed by more pyinfra code to configure it
• This machine is in my home is responsible for managing the certs with lego and pushing out updated files when they're renewed via a new hook script
• I already had a third server in Australia hosting my Chimera Linux mirror that I roped into being an AU POP for Linked List
• I coordinate thigs with a simple Makefile
• make apk will build the linkedlist apk using cbuild. This works on Arch Linux as well as my Chimera WSL2 install on my Windows ARM laptop: cross-compiling is build into cbuild/cports
• make deploy-apk pushes out the updated apk to the servers, updates them, and restarts it
• make deploy, rsyncs the content to the servers. I use xargs to do this in parallel to them all
• Now all the servers were running the application I needed to work out how to route traffic to them
• I toyed with using Deno Deploy's edge network to act as a proxy (link to sample code) but I didn't like that this added the additional latency of an extra request.
• I ended up using gcode managed DNS, which has a GeoDNS feature where you can resolve requests based on geo ip info
  • I'm happy to pay for this but I'm currently in their free tier.
• This is the config I used there:
  • x goes to y, z goes to a, b is default as it's the most powerful
• I used ping bear to check things as well as updown
• This is the end result