If you’ve read this blog before, or are unfortunate enough to have an actual personal relationship with me, you’ll know that I have strong opinions and can be, shall we
Pulumi code ends up looking like a DSL anyway with all the stuff you end up using from the top-level pulumi package to do anything vaguely complicated.
Only now, compared with Terraform, you need to worry about resource ordering and program flow, because when you have a dependency between resources, the resource object you depend on has to be instantiated (within the program flow, I mean - Pulumi handles calculating the ordering of actual cloud resource creation) before the dependent resource. This gets old really quickly if you’re iterating on a module that creates more than a few interdependent resources. So much cut, paste, reorder. FWIW CDK has the same issue, and for the same reason - because it’s using a general-purpose programming language to model a domain which it doesn’t fit all that well.
I like Pulumi and it’s got a lot going for it, especially if you have complex infrastructure requirements. You get a bunch of little quality of life enhancements that I wish Terraform would adopt, like cloud state management by default, and a built-in mechanism for managing secrets in a sane way. Python/TypeScript etc. modules are much more flexible than Terraform modules, and really help with building large chunks of reusable infrastructure. The extra programmability can be useful, though you need to be extra-careful of side-effects. You get more power, but you also get some extra work.
But for most people deploying a bit - or even quite a lot - of cloud infrastructure, Terraform is honestly just easier. It’s usually some fairly simple declarative config with some values passed from one resource to another, and a small amount of variation that might require some limited programmability. Which is exactly what Terraform targets with HCL. It’s clear to me that Pulumi sees this too, since they introduced the YAML syntax later on. But IMO HCL > YAML for declarative config.
Pulumi code ends up looking like a DSL anyway with all the stuff you end up using from the top-level pulumi package to do anything vaguely complicated.
Only now, compared with Terraform, you need to worry about resource ordering and program flow, because when you have a dependency between resources, the resource object you depend on has to be instantiated (within the program flow, I mean - Pulumi handles calculating the ordering of actual cloud resource creation) before the dependent resource. This gets old really quickly if you’re iterating on a module that creates more than a few interdependent resources. So much cut, paste, reorder. FWIW CDK has the same issue, and for the same reason - because it’s using a general-purpose programming language to model a domain which it doesn’t fit all that well.
I like Pulumi and it’s got a lot going for it, especially if you have complex infrastructure requirements. You get a bunch of little quality of life enhancements that I wish Terraform would adopt, like cloud state management by default, and a built-in mechanism for managing secrets in a sane way. Python/TypeScript etc. modules are much more flexible than Terraform modules, and really help with building large chunks of reusable infrastructure. The extra programmability can be useful, though you need to be extra-careful of side-effects. You get more power, but you also get some extra work.
But for most people deploying a bit - or even quite a lot - of cloud infrastructure, Terraform is honestly just easier. It’s usually some fairly simple declarative config with some values passed from one resource to another, and a small amount of variation that might require some limited programmability. Which is exactly what Terraform targets with HCL. It’s clear to me that Pulumi sees this too, since they introduced the YAML syntax later on. But IMO HCL > YAML for declarative config.