Imperative and Declarative API: Another Pile of Marketing Deja-Moo

Looks like some vendor marketers (you know, the same group of people who brought us the switching/routing/bridging stupidity) felt the need to go beyond the usual SDN and intent-based hype and started misusing the imperative versus declarative concepts. Unfortunately some networking engineers fell for the ploy; here’s a typical feedback along these lines I got from one of my readers:

I am frustrated by most people’s shallow understanding API’s, especially the differences between declarative (“what”) and imperative (“how”) API’s, and how that impacts one’s operations. Declarative APIs are the key pillar of what many vendors call “policy” or “intent-based” networking.

Let’s try to unravel that.

You keep using that word…

According to Wikipedia:

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.

I haven’t seen a single API call that would tell the server how to do stuff. They always tell the server what we want to get done. Admittedly, some API calls are more abstracted than others, but that’s a different story.

Conclusion: All API calls are declarative.

Back to Wikipedia:

In computer science, imperative programming is a programming paradigm that uses statements that change a program’s state. In much the same way that the imperative mood in natural languages expresses commands, an imperative program consists of commands for the computer to perform.

All Create, Update, or Delete (three quarters of CRUD) API calls are executed to change the state of the system. Admittedly, some API calls tell the system what to do in vague manner, while others are more precise.

Conclusion: Many API calls are imperative.

Looks like we got to a bit of an absurd situation, and the only way to resolve it is to conclude that it’s all just another pile of marketing deja-moo. QED.

Deja Moo – when you know you’ve seen that bullshit before

Summary

Obviously that won’t stop the vendor marketers from proudly creating slides claiming their product uses declarative API. Make sure to make fun of them whenever you see one – it probably won’t do any more harm than mr. Quixote did to the windmills, but it might feel good (he wasn’t so lucky).

A note to marketers

I wrote this in 2015 but of course nobody ever listens:

I understand you have to do what you have to do, but you don’t have to make yourself looking totally ridiculous in the process.

Feedback on Twitter

Kristian Larsson wrote a lengthy Twitter thread explaining his views on the topic, and allowed me to copy that thread into this blog post:


I don’t think the meaning of declarative and imperative is the same for an API as it is for a computing paradigm or programming language, which seems to be the Wikipedia definition you’ve gone by. Primarily, your mapping of imperative to how leads to an incorrect conclusion. For an API, it’s not about telling a program how to perform a command. I think you should see the ‘how’ as what low level commands you need to string together to achieve your high level intent.

I think of it like this; an imperative API exposes verbs that you then call. add_user() is an imperative command. Calling add_user(‘ivan’, pwd) twice in a row will fail due to username uniqueness. How do I achieve “user Ivan ssh-key is X”. Since calling add_user(‘ivan’, pwd) fails if the user exists, I first have to rm_user(‘ivan’), then add_user(‘ivan’, pwd). Reverse the order of those commands, and it means something entirely different.

In a declarative API on the other hand, you merely specify a list of users - your intent - and let device figure it out. If the user exists before doesn’t matter, then nothing happens, or it’s updated with new parameters. If it didn’t exist, it’s created etc. You could say that add_user() could be made idempotent, which is true and that would make that API call less imperative and more declarative, but it is still fundamentally imperative. The declarative approach is to declare the list of users, not performing actions on that list.

Declarative APIs are generally easier to work with since we effectively delegate the sisyphean task of reaching the intended configuration to the device. Pretty much everything is imperative at a low level. Idempotency is a way to build declarative things out of imperative ones.

Higher level APIs tend to be more declarative than the underlaying low level things they abstract over, which in turn tend to be more imperative. In a stack of abstractions, declarativeness tend to rise the higher up in the stack you go and vice versa.

NETCONF / YANG is a mostly declarative API. The configuration you declare is declarative. You tell the device that there should be 4 BGP neighbors, you don’t do add_bgp_neighbor() or rm_bgp_neighbor(). But NC / YANG also supports imperative things, like restarting a device is not something you configure, it is naturally something you tell it to do, a imperative verb - “restart!” These are modeled in YANG as actions, since that is what they are! Thus it is possible to build an imperative API in NETCONF YANG through actions that modify the applied configuration - but who would want that?

I think describing APIs as declarative or imperative is a useful classification and helps in understanding the capabilities and behavior of them. As usual, it’s not black and white though. Soon someone is going to come along with “but but, in NETCONF, the RPC for editing the supposedly declarative configuration is called ’edit-config’ which sounds very verbish!?”. Yes, but edit-config takes a declarative config. There is no add-config or remove-config RPC :) NETCONF / YANG is fundamentally model driven, because you teach your client once about edit-config, then it can modify (add, delete or modify) any configuration described by a YANG model. When you add ABC (successor of BGP) then you don’t need to teach it about add_abc_neighbor()


While I agree with most everything he wrote, and abstraction layers and idempotency are important, I still don’t think it’s OK to take well-defined terms and reuse them in a vague way (when you find a non-marketing definition of imperative versus declarative API please do let me know).

Also keep in mind that many APIs that people would claim are declarative because they offer a layer of abstraction are not really idempotent (= you cannot declare your intent twice).

More Information

I plan to cover intent-based networking and declarative- and policy APIs in Network Automation Concepts webinar (one of those things I feel need to get done before I disappear for that coffee break).

Revision History

2021-01-14
Added comments by Kristian Larsson

Latest blog posts in Network Infrastructure as Code series

1 comments:

  1. Great thoughts! In addition to atomicity (in my mind this is tracking changes and rolling back the whole set if any fail) I want APIs to fully resolve the order of operations problem entirely. Simply, I think just having a start / stop point where all changes are rolled back if they are not all accepted is an important piece as well as being able to supply a list of desired outcomes per object database, but the ability to declare a configuration set which may reference multiple object databases, and be able to supply all of that in any order in a single transaction request and seeing it completed or rejected as a whole is I think my most desired feature. This may be part of what you meant in the article too and I'm just agreeing with you :).

    For example, when I want to create a firewall policy or a new BGP neighbor configuration, I currently may need to create objects in several places in just the exact correct order before creating the policy / BGP config otherwise it gets rejected.

    In my mind, if vendor equipment has some requirement of a specific order of operations for a configuration set I think THEY really should be supplying the logic to make that happen, not the API consumer. The API consumer should only have to focus on the right settings in the agreed understandable format and then submit it via a transaction.

    I think the order of operations issue is by far the biggest frustration with many APIs. Needing to learn how to reference the specific object databases, learn the specific object models, AND the specific ORDER to accomplish a larger "task" puts the burden on the consumer to become experts on the ins and outs of how the consumed resource thinks. I really think that should be a vendor responsibility.

    Just my opinion! I'd personally be a lot happier if REST APIs (my favorite offender) did this better!

Add comment
Sidebar