Recently, in Container Solutions’ engineering Slack channel, a heated argument ensued amongst our engineers after a Pulumi-related story was posted. I won’t recount the hundreds of posts in the thread, but the first response was “I still don’t know why we still use Terraform”, followed by a still-unresolved ping-pong debate about whether Pulumi is declarative or imperative, followed by another debate about whether any of this imperative vs declarative stuff really matters at all, and why can’t we just use Pulumi please?
This article is my attempt to
prove that I was right and everyone else was wrong calmly lay out some of the issues and help you understand both what’s going on and how to respond to your advantage when someone says your favoured tool is not declarative and therefore verboten.
What does declarative mean, exactly?
Answering this question is harder than it appears, as the formal use of the term can vary from the informal use within the industry. So we need to unpick first the formal definition, then look at how the term is used in practice.
The formal definition
Let’s start with the Wikipedia definition of declarative:
“In computer science, declarative programming is a programming paradigm—a style of building the structure and elements of computer programs—that expresses the logic of a computation without describing its control flow.”
This can be reduced to:
“Declarative programming expresses the logic of a computation without describing its control flow.”
This immediately begs the question: ‘what is control flow?’ Back to Wikipedia:
“In computer science, control flow (or flow of control) is the order in which individual statements, instructions or function calls of an imperative program are executed or evaluated. Within an imperative programming language, a control flow statement is a statement that results in a choice being made as to which of two or more paths to follow.”
This can be reduced to:
“Imperative programs make a choice about what code is to be run.”
According to Wikipedia, examples of control flow include if statements, loops, and indeed any other construct that allows changes which statement is to be performed next (e.g. jumps, subroutines, coroutines, continuations, halts).
Informal usage and definitions
In debates around tooling, people rarely stick closely to the formal definitions of declarative and imperative code. The most commonly heard informal definition saw heard is: “Declarative code tells you what to do, imperative code says how to do it”. It sounds definitive, but discussion about it quickly devolves into definitions of what ‘what’ means and what ‘how’ means.
Any program tells you ‘what’ to do, so that’s potentially misleading, but one interpretation of that is that it describes the state you want to achieve.
For example, by that definition, is this pseudo-code declarative or imperative?
if exists(ec2_instance_1): create(ec2_instance_2) create(ec2_instance_1)
Firstly, strictly speaking, it’s definitely not declarative according to a formal definition, as the second line may or may not run, so there’s control flow there.
It’s definitely not idempotent, as running once does not necessarily result in the same outcome as running twice. But an argument put to me was: “The outcome does not change because someone presses the button multiple times”, some sort of ‘eventually idempotent’ concept. Indeed, a later clarification was: “Declarative means for me: state eventually consistent”.
It’s not just engineers in the field who don’t cling to the formal definition. This Jenkinsfile documentation describes the use of conditional constructs whilst calling itself declarative.
So far we can say that:
- The formal definitions of imperative vs declarative are pretty clear
- In practice and general discussion, people get a bit confused about what it means and/or don’t care about the formal definition
Are there degrees of declarativeness?
In theory, no. In practice, yes. Let me explain.
What is the most declarative programming language you can think of? Whichever one it is, it’s likely that either there is a way to make it (technically) imperative, or it is often described as “not a programming language”.
SQL is to many a canonical declarative language: you describe what data you want, and the database management system (DBMS) determines how that data is retrieved. But even with SQL you can construct conditionals:
insert into table1 where exists ( select 1 from table2 where "some value" == table2.column1 )
The insert to table1 will only run conditionally, i.e. if there’s a row in table two that matches the text “some value”. You might think that this is a contrived example, and I won’t disagree. But in a way this backs up my central argument: whatever the technical definition of declarative is, the difference between most languages in this respect is how easy or natural it is to turn them into imperative languages.
Now consider this YAML, yanked from the internet:
job: script: "echo Hello, Rules!" rules: - if: '$CI_MERGE_REQUEST_TARGET_BRANCH_NAME == "master"' when: always - if: '$VAR =~ /pattern/' when: manual - when: on_success
This is clearly effectively imperative code. It runs in an order from top to bottom, and has conditionals. It can run different instructions at different times, depending on the context it is run in. However, YAML itself is still declarative. And because YAML is declarative, we have the hell of Helm, kustomize, and different devops pipeline languages that claim to be declarative (but clearly aren’t) to deal with, because we need imperative, dynamic, conditional, branching ways to express what we want to happen.
It’s this tension between the declarative nature of the core tool and our immediate needs to solve problems that creates the perverse outcomes we hate so much as engineers, where we want to ‘break out’ of the declarative tool in order to get the things we want done in the way that we want it done.
Terraform and Pulumi
Which brings us neatly to the original subject of the Slack discussion we had at Container Solutions.
Anyone who has used Terraform for any length of time in the field has probably gone through two phases. First, they marvel at how the declarative nature of it makes it in many ways easier to maintain and reason about. And second, after some time using it, and as complexity in the use case builds and builds, they increasingly wish they could have access to imperative constructs.
It wasn’t long before Hashicorp responded to these demands and introduced the ‘count’ meta-argument, which effectively gave us some kind of loop concept, and hideous bits of code like this abound to give us if statements by the back door:
count = var.something_to_do ? 1 : 0
There’s also ‘for’ and ‘for_each’ constructs, and the ‘local-exec’ provisioner, which allows you to escape any declarative shackles completely and just drop to the (decidedly non-declarative) shell once the resource is provisioned.
It’s often argued that Pulumi is not declarative, and despite protestations to the contrary, if you are using it for its main selling point (that you can use your preferred imperative language to declare your desired state), then Pulumi is effectively an imperative tool. If you talk to the declarative engine under Pulumi’s hood in YAML, then you are declarative all the way down (and more declarative than Terraform, for sure).
The point here is that not being purely declarative is no bad thing, as it may be that your use case demands a more imperative language to generate a state representation. Under the hood, that state representation describes the ‘what’ you want to do, and the Pulumi engine figures out how to achieve that for you.
Some of us at Container Solutions worked some years ago at a major institution that built a large-scale project in Terraform. For various reasons, Terraform was ditched in favour of a python-based boto3 solution, and one of those reasons was that the restrictions of a more declarative language produced more friction than the benefits gained. In other words, more control over the flow was needed. It may be that Pulumi was the tool we needed: A ‘Goldilocks’ tool that was the right blend of imperative and declarative for the job at hand. It could have saved us writing a lot of boto3 code, for sure.
How to respond to ‘but it’s not declarative!’ arguments
Hopefully reading this article has helped clarify the fog around declarative vs imperative arguments. First, we can recognise that purely declarative languages are rare, and even those that exist are often contorted into effectively imperative tooling. Second, the differences between these tools is how easy or natural they make that contortion.
There are good reasons to make it difficult for people to be imperative. Setting up simple Kubernetes clusters can be a more repeatable and portable process due to its declarative configuration. When things get more complex, you have to reach for tools like Helm and kustomize which may make you feel like your life has been made more difficult.
WIth this more nuanced understanding, next time someone uses the “but it’s not declarative” argument to shut you down, you can tell them two things: That that statement is not enough to win the debate; and that their suggested alternative is likely either not declarative, or not useful. The important question is not: “Is it declarative?” but rather: ‘How declarative do we need it to be?”