This is odd to me as well. The value of PaaS scales with developers. But Heroku et al target small shops.
If you have 100 dev teams, infra experience is a lot of “overhead.” If you have 5, you just make sure some of the devs know how to strace and tcpdump and can collab in an incident. Frankly, building infrastructure is easy. Maintaining it is the hard part. Preventative maintenance, security, recovering from incidents are all much harder than figuring out enough for it to run.
It was a brilliant idea. Let devs be devs. But the price has always been so high that, at scale, it wasn’t far off the HR cost.
We’ve gone 180 on that with k8s. Now you have to understand a datacenter with different vocabulary to effectively leverage it. You see low adoption at “I have an app” because it’s more complicated than it’s worth. Higher adoption above that. But almost 0 adoption at I need 100+ clusters. And 90% of what I see assumes to exist in a single cluster as if etcd never hangs itself.
We’ve built all of these abstractions, but ultimately you have to understand the thing below them (ie Linux and VMs) when performance sucks or something goes boom.
I don’t know how someone walks in and learns it all at once vs my life of learning with the growing system from Unix on physicals to now.
It’s like my life with Java. Most devs don’t understand where to start troubleshooting performance or the JVM, despite rich tooling (magical APM, JFRs, etc).
If you have 100 dev teams, infra experience is a lot of “overhead.” If you have 5, you just make sure some of the devs know how to strace and tcpdump and can collab in an incident. Frankly, building infrastructure is easy. Maintaining it is the hard part. Preventative maintenance, security, recovering from incidents are all much harder than figuring out enough for it to run.
It was a brilliant idea. Let devs be devs. But the price has always been so high that, at scale, it wasn’t far off the HR cost.
We’ve gone 180 on that with k8s. Now you have to understand a datacenter with different vocabulary to effectively leverage it. You see low adoption at “I have an app” because it’s more complicated than it’s worth. Higher adoption above that. But almost 0 adoption at I need 100+ clusters. And 90% of what I see assumes to exist in a single cluster as if etcd never hangs itself.
We’ve built all of these abstractions, but ultimately you have to understand the thing below them (ie Linux and VMs) when performance sucks or something goes boom.
I don’t know how someone walks in and learns it all at once vs my life of learning with the growing system from Unix on physicals to now.
It’s like my life with Java. Most devs don’t understand where to start troubleshooting performance or the JVM, despite rich tooling (magical APM, JFRs, etc).