Xavier Bruhiere

Lambda functions for rapid prototyping

January 24, 2018 (7y ago)13 views

Abstract: Their “serverless” nature truely fullfill the cloud promises of having as less as possible to manage to put something into world. This is a new cloud paradigm that lower the barrier of prototyping, experimenting and quickly gather feedback. The idea is to show how 1. How efficient a workflow can be 2. What does it mean for early cloud projects 3. How does it fit along containers

It has been a few months already that tech conferences and blogs are agitated by the word "serverless". Somewhat a promise that specific services or frameworks can free you from thinking about servers and just let you push code out into the wild. Of course servers are still there so what is new in the current IaaS (Infrastructure as a Service) landscape ? Well this marketing bundles a much more interesting idea : lambda architectures. A paradigm that splits projects into straight functions, made available accross networks. Yes pretty much the same concept that let tech teams convince their CTO to move to microservices. But a step further. With a consistent approach on the architecture itself : one feature that just wakes up to serve its purpose, then die instead of waiting idle.

I think this is exciting. It makes projects easier to think about and much more cost effective. But like micro services you actually fight new challenges : service discovery, team education, resilience, multi-languages, multi-cloud. You need everything and everyone to be coordinated. Yet, I pretend that lambda functions, despite this, are a perfect fit for early projects and experimentations - among other things (but that is something I let for other engineers with different experiences on those matters). Read on and I hope it will give you some tools to choose the stack of your next projects.

An effective workflow

Architectures based on lambdas basically deploy one compute unit for one service that performs one thing. Needless to say when you go all-in on this and need to manage hundreds of services with execution steps triggered by each other, you can end up with spaghetti-logic-flows (of clean code, but well…).

I won't go further into the details of this situation but you probably need to hire senior architects and read interesting articles on how to manage technical debt or why micro services can bite back.

But today we will take a break and talk about fresh projects, tinkering, experimentations. This time we want it simple, clear, effective. We want to just get done the basics quickly : landing page, user management, payement. Then we can add more differentiation features. But we are not sure yet which will make the users to pay or love the product.

At the end of the day we try to iteratively and safely build the product by defining blocks we can think of in isolation and only when we need it. Dedicating lambda functions to those bricks fit this agile process and make for a clear Minimum Viable Product roadmap.

What about a sudden idea ? Whatever your existing stack, follow your intuition and quickly run a new lambda. Or kill it. Or replace it. Safe experimentations, quick iterations.

So this is the kind of micro-services like popularized by Docker but mentally cheaper ? Yes in some way. Call it nano if you want to distinguish but I believe containers or lambdas are a detail of implementaiton to enjoy the same arhitectural pattern benefits described above. Lambdas enforce 1 service by being 1 function and have the good idea of dying when they are no longer needed. It forces developers to reason on small and stateless processing tasks. Code solves one problem with minimal side effects. A great mix of good old UNIX philosophy with latest functional trend. I'm biased but it serves developer hapiness with strong community consensus isn't it ? And what does it take to be as cool as that ?

module.exports = function(context, callback) {
    callback(200, "Hello, world!\n");
}

Most frameworks and SaaS require you to expose a single function with an object that can hold secrets, HTTP query attributes. Of course nothing prevent us from writing 2000 LoC between module.exports and callback() but it quickly smells like an anti-pattern.

And especially ease of deployment will quickly make it harder to maintain than spliting our problems into manageable endpoint solutions.

$ # completely copy/pasted from http://fission.io/
 
$ fission function create --name hello --env nodejs --code hello.js
$ fission route add --function hello --url /hello
$ curl http://router.fission/hello
Hello, world!

and Bam you have some code that will greet you over HTTP thanks to Fission. It requires you to have a Kubernets setup at hand but well, in 2017, who doesn't have a self-hosted "open-source system for automating deployment, scaling, and management of containerized applications". But really if you don't have time or don't enjoy administrating distributed clusters like you should, you are still in good company. Big hosting players now offer competitive platforms to push lambda on. Webtask 101 can indeed get you setup in 30 seconds, without paying a penny.

$ # again, shamefully copy/pasted from https://webtask.io/docs/101
$ echo "module.exports = function(cb) { cb(null, 'hello world'); }" > foo.js
$ wt create foo.js

When you experienced how to kickstart a project like say React and Webpack, this is amazingly painless. Yet full featured :

$ wt cron schedule \
  -n mongocron \
  -s MONGO_URL=mongodb://webtask:supersecret@ds047592.mongolab.com:47592/webtask-examples \
  10m \
  foo.js

And the lambda function now runs periodically, a really frequent use case as you probably know.

On all aspects the barrier of entry is very low. We write code like we are used to : functions. No new Domain Specific Language, no complex configuration. Or they were quickly wrapped behind poplar frameworks like Serverless does for various cloud providers. And it matters because new paradigms that don't force you to re-learn everything and yet are immediately actionnable for our projects are usually getting a lot of tractions. And a lot of tractions means a good eco-system of services, libraries, help, articles. Which itself feeds the movement and so on.

Those frameworks and services are especially important for Serverless because the whole point is to save you the hassle from managing stuff unrelated to your core project. But it uses non-trivial technologies and manage complex infrastructures of ephemeral compute instances dynamically mapped onto gateways to route traffic.

Self-hosted alternatives like IBM OpenWhisk brings in-house huge control over your serverless stack, without too fancy technologies (proxies, containers). So may be you have the devops team and the investissement ready to make that happen. No more third-party reliance, full customization allowed for very specific business cases (or shortcut solutions). But servers are back so be aware of the challenges (provisioning, configuration, maintenance, monitoring, performances, security).

In the meantime let us focus on how we can leverage existing tools for fun and profit.

Early cloud projects

We saw that code was fast to write, deployments were cheap. Idea to online is as frictionless as it can be and this is a double edged result. Much like twitter you can end up just putting poor vanity idea in the cloud in no time or you can live by the Lean Startup and use it to iterate fast. Whatever the quality of your initial thoughts, make it public or share it privately to gather feedback and experience. It can be a good proxy to assess if your initial design or tech is simply relevant. Or build a landing page and gauge interest. Either you will be able to focus on the graal, something people want, or you will be fast to realise it doesn't worth to put more time in it. Engineers usually don't like to cut an arm and give up on something they really wanted to build (except for doing it again from scratch or switch to another fancy tech). I can't help you fighting this syndrome but at least we can have the right numbers to decide.

Our leitmotiv is to go fast and we can insist a little more : Lambdas can be used to be DRY at the architecture level. I feel this argument is a bit weaker than the other because like many good practices in computer science, it is only as good as the author implemented it. Yet, Serverless workflow promotes decouple services with a single purpose, hence you can probably share some of them, like payement processing, static site rendering, user registration and so on. I definitely did and as you develop projects you will start to extract patterns and share more an more services to go faster at the prototyping level. I do believe though that it will probably break as project grow. From experience we tend to specialise bricks to a point where it solves really specific edge cases too. Keeping things as simple as possible, implementing only features we really need, following the open/closed principle are no silver bullets but it can go a long way.

But let's add some open-source and community ingredients here to our thinking. Since we are talking about decoupled service with clear boundaries, sharing code can extend further than our own stack. Stdlib raised recently 2 millions for offering a common library of functions accessible through the network. Everything seems to point in this direction but i'm not sure we are here yet. Maybe there is a lot to solve in trusting third-party codes, in-house development trade-offs or adoption rate but I didn't meet a lot of projects that leverage this opportunity. Yet, as a small idea, this is something we can actually do to, once again, to go faster. Every services that are common to early SaaS for example but don't represent a competitive edge could be just plug so that we focus on this so-called competitive seed.

Although we listed some limitations of this approach, it is worth noting this is something achievable inside an organisation too to empower development teams.

Tech landscape

So far I fired most of the arguments I had in favour of using lambda technologies for small prototypes. I hope it resonates with your experience or may be push you to try it for your next move. But may be you caught many points or keywords shared by microservices and containers articles. So I want to clear confusions and situate where lambdas fit on the current devops landscape (a small part of it to be fair).

First, none of those terms are mutually exclusives : you can design a microservices architecture using lambdas that run code inside containers. This is no accident serverless is coming few years after Docker and containers popularity. They unlocked powerful cloud orchestration and the agility lambdas need to fire on events and die right after processing it.

From my point of view it comes down to developers usage and business goals. Nothing will prevent us from putting elefants in containers or deploying microservices on bare-metal machines. Those are tools you can leverage and combine depending of your constraints. Kubernetes for example markets itself as a "Production grade container orchestration" but tweak some bolts and you get a "serverless framework". So :

  • Microservices is an architectural pattern that promote decoupled services with clear, small boundaries and interfaces.
  • Containers are a linux feature that allow lighter isolation of processes than VMs thanks to cgroups.
  • Lambdas are ephemeral functions that wake up on events, process it and die after that.

Conclusion

The startup where I work organized its first hackathon last week. We built many projects in 24 hours and many of us ended up deploying lambdas functions. Because it was fast. Because it was cheap (free actually). Because we could split the work between team of three or four. The scope was clearly defined and it was a perfect fit for event based projects (like developing slack bots). Some of us discovered the framework for the first time and yet they were operational in a few minutes, writing, deploying and monitoring lambda code.

All of this quickly breaks when projects grow, like we mentioned. Yet as we improve developer tooling I think this paradigm takes us in the right direction: lazy execution, effective costs, ease of usage, ease of sharing. We probably don't need to migrate all our professional infrastructures under the lambda hat. We probably need les frameworks and more winners that just work. And we definitely need to define how to deploy and maintain hundreds of functions.

So jump aboard. This is already an exciting technology for actionable reasons, and we still have a lot more to contribute.