It seems Fly.io originally focused on Elixir, but has identified the product-market need for app server + Postgres.
We didn't know about Fly.io when we chose GCP for this setup.
The initial setup on GCP was exceedingly painful. We used CloudRun for our app server, with the value prop being that "it just works". It didn't. Our container failed to start with zero logs from our servers. Stackdriver was of no help. Eventually we found a Stackoverflow thread revealing that CloudRun didn't like Docker images built from Macs. As always, GCP's official docs and resources are incoherent. GCP docs address a hundred things you don't care about, and the signal-to-noise percentage is in the low teens, if we're being generous. We had to chase down half a dozen bureaucratic things to get our CloudRun app to see and talk to CloudSQL. Apparently with Fly.io, you just run a command to provision Postgres, and pass in an environment variable to your app.
We consoled ourselves that GCP was difficult to setup but now it's set-and-forget. This is also a lie. This week we saw elevated and unexplained 5xx. First was CloudRun randomly disconnecting from CloudSQL. As AWS measures reliability in terms of 9s, the way GCP DevRel responded to this bug is that this is a distributed system and therefore acceptable that things just fail a reliably human-reproducable 1%+ error rate.
Yesterday we saw botnet traffic scanning for vulnerabilities on our app. This happens if you're on the web, not inherently GCP's fault. We have GCP's Cloud Load balancer setup but it's not very smart. We were able to manually block specific IP addresses but it's no where as meaningful as Cloudflare. Not a fan of Cloudflare the company but their products address a need. The botnet somehow knocked over our "Serverless VPC connection" to CloudSQL. Basically what that is is a proxy server that you are forced to setup because CloudRun can't actually talk to cloudSQL. All the auto-scaling claims of GCP's serverless are diminished if we are forced to introduce a single point of failures like this in the loop. That serverless VPC connection requires a minimum of 2+ VMs, so the scale-to-zero of CloudRun is no more.
Our experience with GCP is constantly having to come up with workarounds and addressing their accidental complexity. This should not be the customer's problem. For example, CloudSQL doesn't have an interface to query your databases. If you use a private IP for security, you can't even use GCP's command line tools to access this. We found out that GCE VMs are automatically networked to talk to CloudSQL. We ended up creating a "bastion" GCE VM instance and setup Postgres CLI tools in order to do ad-hoc queries of our DB state. For this, we just needed to the cheapest VM but GCP makes even this difficult. As for Stackdriver, it's still been an annoyingly painful UI.
If I remember correctly Fly weren't focused on Elixir at first, and instead were just trying to be a better Heroku, but then one of their developers realised Elixir would be a good fit for one of the components they were writing, and off the back of that realised "Heroku, but with a private network to allow comms between nodes" was a really good fit for Elixir applications.
Yes, the platform came before the Elixir love; it just happened that Elixir is the development ecosystem that best takes advantage of a distributed platform with private networking, and that Phoenix LiveView really screams when you can run it within 10s of milliseconds of all your users.
We didn't know about Fly.io when we chose GCP for this setup.
The initial setup on GCP was exceedingly painful. We used CloudRun for our app server, with the value prop being that "it just works". It didn't. Our container failed to start with zero logs from our servers. Stackdriver was of no help. Eventually we found a Stackoverflow thread revealing that CloudRun didn't like Docker images built from Macs. As always, GCP's official docs and resources are incoherent. GCP docs address a hundred things you don't care about, and the signal-to-noise percentage is in the low teens, if we're being generous. We had to chase down half a dozen bureaucratic things to get our CloudRun app to see and talk to CloudSQL. Apparently with Fly.io, you just run a command to provision Postgres, and pass in an environment variable to your app.
We consoled ourselves that GCP was difficult to setup but now it's set-and-forget. This is also a lie. This week we saw elevated and unexplained 5xx. First was CloudRun randomly disconnecting from CloudSQL. As AWS measures reliability in terms of 9s, the way GCP DevRel responded to this bug is that this is a distributed system and therefore acceptable that things just fail a reliably human-reproducable 1%+ error rate. Yesterday we saw botnet traffic scanning for vulnerabilities on our app. This happens if you're on the web, not inherently GCP's fault. We have GCP's Cloud Load balancer setup but it's not very smart. We were able to manually block specific IP addresses but it's no where as meaningful as Cloudflare. Not a fan of Cloudflare the company but their products address a need. The botnet somehow knocked over our "Serverless VPC connection" to CloudSQL. Basically what that is is a proxy server that you are forced to setup because CloudRun can't actually talk to cloudSQL. All the auto-scaling claims of GCP's serverless are diminished if we are forced to introduce a single point of failures like this in the loop. That serverless VPC connection requires a minimum of 2+ VMs, so the scale-to-zero of CloudRun is no more.
Our experience with GCP is constantly having to come up with workarounds and addressing their accidental complexity. This should not be the customer's problem. For example, CloudSQL doesn't have an interface to query your databases. If you use a private IP for security, you can't even use GCP's command line tools to access this. We found out that GCE VMs are automatically networked to talk to CloudSQL. We ended up creating a "bastion" GCE VM instance and setup Postgres CLI tools in order to do ad-hoc queries of our DB state. For this, we just needed to the cheapest VM but GCP makes even this difficult. As for Stackdriver, it's still been an annoyingly painful UI.