# cloud: i recommend using fly.io and cloudflare.com for most cloud needs.

as part of @/rebrand i migrated this blog to the cloud from my rusty old first gen rasperry pi. the old setup worked flawlessly but nevertheless i decided to migrate it because that's the hip thing to do in the 21st century.

# the choice

i spent ages contemplating on the cloud provider choice. oracle? google? vultr? ovhcloud? hetzner? fly.io?

i've chosen fly.io for its transparent free tier. they give you a fixed amount of free credits per month and you can spend it the way you want on their services. that's the neatest approach. even if i decide to set up other services, i think fly.io will be the most flexible. i'm aware that fly.io has bit of a bad reputation for its outages (https://community.fly.io/t/reliability-its-not-great/11253). things should be improving though. for my blog usecase i don't mind it being a bit buggy in exchange for the transparency.

but for the comments i also needed storage. using storage on fly.io makes your service less reliable because iiuc you get a block device on a specific host. your virtual machine can then be scheduled on that host only. if that host is out, the service is down which is a bummer.

so i started looking at free s3-like storage providers. i found tebi.io and cloudflare.com for this. i've chosen cloudflare after agonizing about this choice for a whole day. cloudflare comes with many other services (such as email workers which i'll also need) so i decided might as well play with that. it turned out to be a good pick.

r2 is cloudflare's s3-like storage offering. but it also has a key-value (kv) api for simpler needs. i ended using kv instead of r2.

# fly.io: deployment

the deployment is docker based. i wrote a very simple dockerfile that installs go on alpine, git pulls my blog and builds it. then it copies the binary onto another clean alpine image, git pulls the blog content and runs the server.

note that docker does lot of caching so a trick is needed to ensure that the image gets rebuilt after a git push. i use `COPY .git/refs/remotes/origin/master version` for this. see https://github.com/ypsu/blog/blob/master/Dockerfile for the glory details. i also needed a fly.toml but that was fairly straightforward.

i run "fly deploy" every time i change the code. it builds and pushes quite quickly. fly.io's cli tool is pretty handy.

i've written my server to git pull the content automatically on startup and over its lifetime so it always presents the freshest content. i don't need to rebuild the image whenever i make a new post.

# fly.io: autostopping

i have enabled auto-stopping for my server. if my server had no active requests in the past ~5 minutes, fly.io shuts the machine down. it turns it back on when the next request arrives.

this is pretty neat. my service doesn't actually need to run if nobody is talking to it. i like the energy efficiency of this.

the downside of this is that the first request is going to be slow. the vm has to start, git pull for the content must complete, i need to fetch the comments from cloudflare, and then i need to start the server. it can take up to 2 seconds and sometimes even more. but once up it's fast enough.

so far this doesn't bother me. i can very easily disable this if this starts to annoy me. see min_machines_running at https://fly.io/docs/reference/configuration/#the-http_service-section.

edit: well, this didn't last a whole day. i got annoyed by the occasional slowness. it's an always running server now.

# fly.io: idle timeouts

for both @/webchat and @/msgauth demos i need long lived idle connections. the internet claimed that this won't work: https://community.fly.io/t/is-it-possible-to-increase-the-timeout-to-120-sec/3007/5.

i had two ideas to work around this:

• buy the basic support plan for a month and then ask them to increase the timeout for my app. i wouldn't need the plan afterwards.
• emit a "." character every 50 seconds on those idle connections and have the client ignore the initial stream of dots once the request finishes.

but it turns out this isn't an issue. i had 20+ minute long idle connections that completed just fine after the event arrived on the server side.

# fly.io: dynamic request routing

right now i run a single server. but if i want to implement a leader-follower like architecture, something i was alluring to in @/scaling, this could be pretty trivial in fly.io. i'd simply use https://fly.io/docs/reference/dynamic-request-routing/ to route POST requests to the leader. GET requests i could fulfill from any replica as soon as its state catches up with state the client has last seen (stored in a cookie).

https://fly.io/blog/how-to-fly-replay/ explains this neat feature in more detail.

# cloudflare: serverless

but on fly.io i don't have a simple solution to storage. this led me to cloudflare and its whole "serverless" ideology.

the idea is that i give cloudflare a javascript snippet and they will execute it whenever my endpoint is hit. this doesn't need to start up any x86 compatible virtual machines, just a javascript environment, similar to what a browser does. the isolation the js executors give is more than adequate for most usecases. perhaps later the wasm support will further improve and then non-js languages can be used with ease too.

i realized i could implement most of my blog as a serverless service. but @/webchat or @/msgauth wouldn't work as a serverless cloudflare worker. for that i would need to use cloudflare's "durable objects" api: https://blog.cloudflare.com/introducing-workers-durable-objects/. i really like the concept and i can totally imagine myself of using it for some stuff.

so static hosting like github pages + cloudflare durable objects would be enough for this blog. there are 2 reasons i'm not regretting my old school setup with fly.io though:

• if i start using durable objects i might be too dependent on cloudflare. migrating to a different storage api is simple but durable objects changes how i architect my demos.
• and the stronger reason is that it's not free. i'd need to pay $5 per month to access the api.

but serverless computing is something i'd seriously consider for a serious application.

# cloudflare: dns management

i pointed iio.ie's domain nameservers to cloudflare. i didn't enable cloudflare's ddos protection for my service. so the iio.ie requests go directly to fly.io.

it's an unsupported interaction anyway because i really wanted my fly.io instance to only talk https. but in order for fly.io to generate the ssl cert, it wants the domain to be pointing at fly.io's ip address. that won't be the case if the domain points at cloudflare's proxy.

https://community.fly.io/t/cloudflare-525-error-randomly-occurs/1798 explains some workarounds. basically turn off https enforcement in fly.io's proxy level and do it yourself (or not do it at all). i think the fly.io app would need a shared ipv4 address for that. then you can certainly bypass fly.io's cert management limitations. but that sounds like pain.

nevertheless i wanted to host the cloudflare workers on my domain for the completeness' sake. this way i could change my storage backend without needing to change anything in my server. so i pointed api.iio.ie to some dummy ip address (100:: and 192.0.2.1) and i enabled cloudflare for it. then i configured cloudflare to route only a specific path to my cloudflare worker. this way a silly bot stumbling onto api.iio.ie's frontpage won't eat into my worker quota.

when i initially configured my cloudflare worker to talk to my fly.io server, it didn't work. the fetch request errored out with a too many redirects error. for some reason cloudflare really insisted talking http with my server which always got a redirect response. i've fixed this by asking cloudflare to always use https. more specifically, i switched my domain's ssl encryption mode to "full (strict)" in the cloudflare dashboard.

# cloudflare: request workers

i've created a cloudflare worker and pointed a specific path under api.iio.ie to them. here's how my worker looks like implementing the some operations:

  async function handleFetch(request: Request, env: Env, ctx: ExecutionContext): Promise < Response > {
    let method = request.method
    let path = (new URL(request.url)).pathname
    let params = (new URL(request.url)).searchParams
    if (env.devenv == 0 && request.headers.get('apikey') != env.apikey) {
      return response(403, 'unathorized')
    }

    let value, list, r
    switch (true) {
      case path == '/api/kv' && method == 'GET':
        let value = await env.data.get(params.get('key'))
        if (value == null) return response(404, 'key not found')
        return response(200, value)

      case path == '/api/kv' && method == 'PUT':
        await env.data.put(params.get('key'), await request.text())
        return response(200, 'ok')

      case path == '/api/kv' && method == 'DELETE':
        await env.data.delete(params.get('key'))
        return response(200, 'ok')

    ...

my blog server uses this to upload, fetch, delete individual comments. i also have further endpoints for listing and fetching all comments in a single request.

cloudflare's cli tool is pretty easy to use too. i run `wrangler dev -r` to run the worker locally. it is then pointed to a test kv namespace so i can have test comments. and when i'm ready, i use `wrangler deploy` to push it to production.

it's pretty straightforward, i love it.

# cloudflare: email workers

for my @/msgauth demo i need to parse incoming emails. previously i used a hacky smtp server for that without any spoofing protections. but cloudflare has this: https://developers.cloudflare.com/email-routing/email-workers/.

the worker runs for each email received on the preconfigured address. in my case it just sends a POST request to my server to notify about the email's receipt. it's super straightforward and it allowed me to delete lot of hacky code.

# logs

both `fly logs` and `wrangler tail` allow me to stream the debug logs in real time. but neither of them offers historical access or analytics for them. on fly.io i could simply configure an external service to push them to. on cloudflare this needs a paid plan already.

but meh, i can live without logs. having access to logs would make me just obsess about them. if i need to debug something then i hope the real time logs will be more than enough. they were certainly enough when i was moving my blog to the cloud.

# github actions

oh, by the way, i am also using github actions. whenever i push to github, a github action rebuilds the backup page of this blog that is linked on the frontpage. the action also calls an endpoint on the blog to run `git pull`. this way this blog always serves fresh content without needing to re-deploy the service on fly.io.

# takeaway

it's only been a day since this blog is on the cloud using those two providers but so far i'm super happy with the results. i can't think of anything that could go wrong so i hope it stays this way.

next time someone asks me how to create an online service, i will probably suggest to look into cloudflare. cloudflare also has s3-like storage, sql database, cron triggers, etc so it should cover most needs. and if they need something for more computationally intensive tasks, then for those parts i'll recommend fly.io. i think that's a super sensible approach in 2023.

i'm sure the other cloud providers are nice too. i haven't really used them. well, i used google cloud for @/gdsnap but i found its dashboard (and its whole cloud offering tbh) just painfully overwhelming. my choices only offer a few apis but they do them relatively well.

published on 2023-09-08

to the frontpage