3 ms·
When I saw this I got super excited because I thought it was going to be "per-minute low-latency distcc hosts" which was probably set up by having a ton of comm
by saurik 10y ago
When I saw this I got super excited because I thought it was going to be "per-minute low-latency distcc hosts" which was probably set up by having a ton of common toolchains ready to go at all times, and my big questions were "did they bother to support MinGW (easy)" and "do they have a reasonable story for iOS (though I can't imagine they have armv6 working right)", but while they got my hopes up when they said one could even upload custom toolchains I realized that didn't make much sense and a few paragraphs later I got the disappointing news that this is just some extremely thin wrapper over ECS that is limited to doing builds on single computers which max out at 8 vCPU (why?!). I mean, I guess this makes CI slightly more accessible to some people, but it isn't anything terribly exciting and is mostly going to help people with extremely small projects: this isn't going to scale up to the kinds of builds where you'd expect a service billing itself on scalability to be most valuable.
...Except for this part: CloudBuild has per-minute billing!! This is one of the major complaints people have about EC2 (and all the services Amazon builds over it), and is one of the major downsides of using it over Google's Compute Engine. If you have any kind of task that can possibly be thought of as a "build"--one which can be expressed as a container of software configured to access some external asset as input and which generates a concrete output "artifact" (and maybe even not, right? to support some silly things people do in their builds like "check out code from npm", you likely get network access, and your build output could always be an empty file)--this now seems like a depressingly hilarious way to trick Amazon's infrastructure into giving you per-minute billing for random tasks which take less than 20 minutes to run (important limit, as they are charging a 3x overhead vs the on demand price for an equivalent instance: for 8 vCPU / 15 GB instance, a c4.2xlarge costs $0.419 per hour and a build.general1.large costs $.02 per minute, which would be $1.20 per hour).
In other words: I will argue that this service really can and maybe should just be looked at as a different pricing model for ECS, to support any "small" task (not just building code): if it takes less than 20 minutes and doesn't require a massive computer, CodeBuild is not only cheaper but probably easier to use (as it already models the problem in terms of a task queue, so you don't have to do that part either).