Has the “serverless revolution stalled”? I dunno. I like serverless. Taught a popular course about it. But I reviewed and published an article written by Bernard Brode that made that argument, and it sparked a lot of discussion. If we can agree that serverless computing means building an architecture out of managed services that scale to zero—we’re not strictly talking about function-as-a-service—that’s a start. Has this serverless model crossed the chasm from early adopters to an early majority? I don’t think so. And the data shows that usage of FaaS—still a fundamental part of most people’s serverless architecture—has flattened a bit. Why is that? I’m no expert, but I wonder if some of the inherent friction of the 1st generation FaaS gets in the way.
We’re seeing a new generation of serverless computing that removes that friction and may restart the serverless revolution. I’m talking here about Google Cloud Run. Based on the Knative project, it’s a fully managed service that scales container-based apps to zero. To me, it takes the best attributes from three different computing paradigms:
|Platform-as-a-Service||– focus on the app, not underlying infrastructure|
– auto-wire networking components to expose your endpoint
|Container-as-a-Service||– use portable app packages|
– develop and test locally
|Function-as-a-Service||– improve efficiency by scaling to zero|
– trigger action based on events
Each of those above paradigms has standalone value. By all means, use any of them if they suit your needs. Right now, I’m interested in what it will take for large companies to adopt serverless computing more aggressively. I think it requires “fixing” some of the flaws of FaaS, and there are four reasons Cloud Run is positioned to do so.
1. It doesn’t require rearchitecting your systems
First-generation serverless doesn’t permit cheating. No, you have to actually refactor or rebuild your system to run this way. That’s different than all the previous paradigms. IaaS? You could take existing bare metal workloads and run them unchanged in a cloud VM platform. PaaS? It catered to 12-factor apps, but you could still run many existing things there. CaaS? You can containerize a lot of things without touching the source code. FaaS? Nope. Nothing in your data center “just works” in a FaaS platform.
While that’s probably a good thing from a purity perspective—stop shifting your debt from one abstraction to another without paying it down!—it’s impractical. Simultaneously, we’re asking staff at large companies to: redesign teams for agile, introduce product management, put apps on CI pipelines, upgrade their programming language/framework, introduce new databases, decouple apps into microservices, learn cloud and edge models, AND keep all the existing things up and running. It’s a lot. The companies I talk to are looking for ways to get incremental benefits for many workloads, and don’t have the time or people to rebuild many things at once.
This is where Cloud Run is better than FaaS. It hosts containers that respond to web requests or event-based triggers. You can write functions, or, containerize a complete app—Migrate for Anthos makes it easy. Your app’s entry point doesn’t have to conform to a specific method signature, and there are no annotations or code changes required to operate in Cloud Run. Take an existing custom-built app written in any language, or packaged (or no source-code-available) software and run it. You don’t have to decompose your existing API into a series of functions, or break down your web app into a dozen components. You might WANT to, but you don’t HAVE to. I think that’s powerful, and significantly lowers the barrier to entry.
2. It runs anywhere
Lock-in concerns are overrated. Everything is lock-in. You have to decide whether you’re getting unique value from the coupling. If so, go for it. A pristine serverless architecture consists of managed services with code (FaaS) in the gaps. The sticky part is all those managed services, not the snippets of code running in the FaaS. Just making a FaaS portable doesn’t give you all the benefits of serverless.
That said, I don’t need all the aspects of serverless to get some of the benefits. Replacing poorly utilized virtual machines with high-density nodes hosting scale-to-zero workloads is great. Improving delivery velocity by having an auto-wired app deployment experience versus ticket-defined networking is great. I think it’s naive to believe that most folks can skip from traditional software development directly to fully serverless architectures. There’s a learning and adoption curve. And one step on the journey is defining more distributed services, and introducing managed services. Cloud Run offers a terrific best-of-both-worlds model that makes the journey less jarring. And uniquely, it’s not only available on a single cloud.
Cloud Run is great on Google Cloud. Given the option, you should use it there. It’s fully managed and elastic, and integrates with all types of GCP-only managed services, security features, and global networking. But you won’t only use Google Cloud in your company. Or Azure. Or AWS. Or Cloudflare. Cloud Run for Anthos puts this same runtime most anywhere. Use it in your data center. Use it in your colocation or partner facility. Use it at the edge. Soon, use it on AWS or Azure. Get one developer-facing surface for apps running on a variety of hosts.
A portable Faas, based on open source software, is powerful. And I believe, necessary, to break into mainstream adoption within the enterprise. Bring the platform to the people!
3. It makes the underlying container as invisible, or visible, as you want
Cloud Run uses containers. On one hand, it’s a packaging mechanism, just like a ZIP file for AWS Lambda. On the other, it’s a way to bring apps written in any language, using any libraries, to a modern runtime. There’s no “supported languages” page on the website for Cloud Run. It’s irrelevant.
Now, I personally don’t like dealing with containers. I want to write code, and see that code running somewhere. Building containers is an intermediary step that should involve as little effort as possible. Fortunately, tools like Cloud Code make that a reality for me. I can use Visual Studio Code to sling some code, and then have it automatically containerized during deployment. Thanks Cloud Buildpacks! If I choose to, I can use Cloud Run while being blissfully unaware that there are containers involved.
That said, maybe I want to know about the container. My software may depend on specific app server settings, file system directories, or running processes. During live debugging, I may like knowing I can tunnel into the container and troubleshoot in sophisticated ways.
Cloud Run lets you choose how much you want to care about the container image and running container itself. That’s a flexibility that’s appealing.
4. It supports advanced use cases
Cloud Run is great for lots of scenarios. Do server-side streaming with gRPC. Build or migrate web apps or APIs that take advantage of our new API Gateway. Coordinate apps in Cloud Run with other serverless compute using the new Cloud Workflows. Trigger your Cloud Run apps based on events occurring anywhere within Google Cloud. Host existing apps that need a graceful shutdown before scaling to zero. Allocate more horsepower to new or existing apps by assigning up to 4 CPUs and 4GB of RAM, and defining concurrency settings. Decide if your app should always have an idle instance (no cold starts) and how many instances it should scale up to. Route traffic to a specific port that your app listens on, even if it’s not port 80.
If you use Cloud Run for Anthos (in GCP or on other infrastructure), you have access to underlying Kubernetes attributes. Create private services. Participate in the service mesh. Use secrets. Reference ConfigMaps. Turn on Workload Identity to secure access to GCP services. Even take advantage of GPUs in the cluster.
Cloud Run isn’t for every workload, of course. It’s not for background jobs. I wouldn’t run a persistent database. It’s ideal for web-based apps, new or old, that don’t store local state.
Give Cloud Run a look. It’s a fast-growing service, and it’s free to try out with our forever-free services on GCP. 2 million requests a month before we charge you anything! See if you agree that this is what the next generation of serverless compute should look like.