Microservice Hosting

Overview

Broadly speaking, there are two main approaches to writing microservices for Azure

1. Containers manged by a orchestration service such as AKS or ACS
2. Azure functions

There are trade offs to each approach...

Containers

Cloud agnostic
In general, containers will be easier to migrate (or host as part of a multi-cloud setup) than serverless. This is, of course, never straightforward because unless you deliberately set out to minimise platform dependencies from the start, you will probably be using multiple Azure PaaS services such as Key Vault, AD, Pipelines, etc. However, I think it is fair to say that it will always be easier to move an application in a container to a different host than it is to migrate, or example, Azure functions to AWS Lambda, which would be a complete rewrite. So comparatively, with containers, you are certainly reducing vendor lock-in.

Legacy applications
It is easier to migrate existing applications to containers. You can wrap an existing API in a container and update the build/deploy process. It will be a total rewrite (granted with a lot of CTRL-C, CTRL-V) to go serverless.

Development
It is less of a learning curve with containers than serverless which is an entirely different model. It is also easier to spin up local dev environments that contain multiple dependencies with containers.

Also, In the .NET world, decent tooling exists for contract first development with ASP.NET Core that allows you to auto-generate APIs from an Open API spec as part of the build process. At time of writing, no such tooling exists for for Azure functions.

Serverless

Cost
With container orchestration, you have to pay for each hosting node - which nearly always has to be over provisioned to cope with peaks which in turn results in unused capacity and unnecessary costs. With the consumption plan available in Azure functions, you are only paying for the number of invocations/requests. Overall, costs are better aligned to usage which includes scaling all the way down to 0. Its obviously use case dependent but in general serverless is going to be cheaper than running a production AKS cluster with 3+ worker nodes.

Integration
Azure Functions are easy to integrate to many types of Azure services using declarative input/output bindings. This is especially useful for event based microservices as you have triggers for the plethora of Azure eventing systems (Service Bus, Event Hubs, Event Grid).

Scaling
Auto scaling based on throughput is trivial with Azure functions. Although this is possible with AKS, in my experience at least, it is much harder to setup and maintain.

Conclusion

If you are starting a greenfield development, and don't have a requirement for massive scale or need to go multi-cloud, I believe Azure functions are the way to go. The ease of integrating with other services and simplified deployment and hosting allow you to quickly develop and deploy solutions.

If you are migrating to cloud from on premise, anticipate huge scale, or want to remain somewhat cloud agnostic, containers are a better solution.

Of course the argument is a bit of a false dichotomy as a hybrid approach is often required.