only the first run pays the full cost, after that, only what changed is shipped.
For example on a 20MiB repo you would upload ~11MB on the first cold run, but only 3.3KB on repeats.
I tried to make this as efficient as possible and only sending whats needed
A VM you manage yourself is cheaper per hour. An 8 vCPU / 16 GiB machine costs $0.66/hr with us, versus about $0.42/hr for an on-demand EC2 c7i.2xlarge. Roughly 1.5x. GCP's free tier has 1 GB of RAM and shared CPU, afaik, so it's pretty limited for builds.
The premium is for not doing the infra dance yourself.
Forget the EC2 instance is up and it's $10.18/day.
We bill by the second while a command runs, so against a VM you leave running all day, your break even is around 15 hours of compute.
I used to run my VMs and I would forget to tear them down and got tired of rsyncing and explaining to my agents what the setup was or where to run stuff.
That's what I wanted to fix. Your uncommitted working tree runs remotely, results come back to disk, exit codes and Ctrl-C work, and there's nothing to remember to shut down. Adding vanish cargo test just works as cargo test
If you keep on top of your VMs and don't mind the syncing, running your own will be cheaper.
Wouldn't this require you to send over your entire workspace with each call?
only the first run pays the full cost, after that, only what changed is shipped.
For example on a 20MiB repo you would upload ~11MB on the first cold run, but only 3.3KB on repeats. I tried to make this as efficient as possible and only sending whats needed
trying to understand pricing. How does this compare to getting a free google cloud vm?
A VM you manage yourself is cheaper per hour. An 8 vCPU / 16 GiB machine costs $0.66/hr with us, versus about $0.42/hr for an on-demand EC2 c7i.2xlarge. Roughly 1.5x. GCP's free tier has 1 GB of RAM and shared CPU, afaik, so it's pretty limited for builds.
The premium is for not doing the infra dance yourself. Forget the EC2 instance is up and it's $10.18/day. We bill by the second while a command runs, so against a VM you leave running all day, your break even is around 15 hours of compute.
I used to run my VMs and I would forget to tear them down and got tired of rsyncing and explaining to my agents what the setup was or where to run stuff.
That's what I wanted to fix. Your uncommitted working tree runs remotely, results come back to disk, exit codes and Ctrl-C work, and there's nothing to remember to shut down. Adding vanish cargo test just works as cargo test
If you keep on top of your VMs and don't mind the syncing, running your own will be cheaper.