In one way or another, every developer has come in touch with an API. Either integrating a major system for a big corporation, producing some fancy charts with the latest graph library, or simply by interacting with his favorite programming language. The truth is that APIs are everywhere! They actually represent a fundamental building block of the nowadays Internet, playing a fundamental role in the data exchange process that takes place between different systems and devices. From the simple weather widget on your mobile phone to a credit card payment you perform on an online shop, all of these wouldn’t be possible if those systems wouldn’t communicate with each other by calling one another’s APIs.
So with the ever growing eco-system of heterogeneous devices connected to the internet, APIs are put a new set of demanding challenges. While they must continue to perform in a reliable and secure manner, they must also be compatible with all these devices that can range from a wristwatch to the most advanced server in a data-center.
REST to the rescue
One of the most widely used technologies for building such APIs are the so called REST APIs. These APIs aim to provide a generic and standardize way of communication between heterogeneous systems. Because they heavily rely on standard communication protocols and data representation – like HTTP, XML or JSON – it’s quite easy to provide client side implementations on most programming languages, thus making them compatible with the vast majority of systems and devices.
So while these REST APIs can be compatible with most devices and technologies out there, they also must evolve. And the problem with evolution is that you sometimes have to maintain retro-compatibility with old client versions.
Let’s build up an example.
Let’s imagine an appointment system where you have an API to create and retrieve appointments. To simplify things let’s imagine our appointment object with a date and a guest name. Something like this:
public class AppointmentDTO {
public Long id;
public Date date;
public String guestName;
}A very simple REST API would look like this:
@Path("/api/appointments")
public class AppointmentsAPI {
@GET
@Path("/{id}")
public AppointmentDTO getAppointment(@PathParam("id") String id) { ... }
@POST
public void createAppointment(AppointmentDTO appointment) { ... }
}Let’s assume this plain simple API works and is being used on mobile phones, tablets and various websites that allow for booking and displaying appointments. So far so good.
At some point, you decide it would be very interesting to start gathering some statistics about your appointment system. To keep things simple you just want to know who’s the person who booked most times. For this you would need to correlate guest between themselves and decide you need to add an unique identifier to each guest. Let’s use Email. So now your object model would look like something like this:
public class AppointmentDTO {
public Long id;
public Date date;
public GuestDTO guest;
}
public class GuestDTO {
public String email;
public String name;
}So our object model changed slightly which means we will have to adapt the business logic on our api.
The Problem
While adapting the API to store and retrieve the new object types should be a no brainer, the problem is that all your current clients are using the old model and will continue to do so until they update. One can argue that you shouldn’t have to worry about this, and that customers should update to the newer version, but the truth is that you can’t really force an update from night to day. There will always be a time window where you have to keep both models running, which means your api must be retro-compatible.
This is where your problems start.
So back to our example, in this case it means that our API will have to handle both object models and be able to store and retrieve those models depending on the client. So let’s add back the guestName to our object to maintain compatibility with the old clients:
public class AppointmentDTO {
public Long id;
public Date date;
@Deprecated //For retro compatibility purposes
public String guestName;
public GuestDTO guest;
}Remember a good thumb rule on API objects is that you should never delete fields. Adding new ones usually won’t break any client implementations (assuming they follow a good thumb rule of ignoring new fields), but removing fields is usually a road to nightmares.
Now for maintaining the API compatible, there are a few different options. Let’s look at some of the alternatives:
- Duplication: pure and simple. Create a new method for the new clients and have the old ones using the same one.
- Query parameters: introduce a flag to control the behavior. Something like useGuests=true.
- API Versioning: Introduce a version in your URL path to control which method version to call.
So all these alternatives have their pros and cons. While duplication can be plain simple, it can easily turn your API classes into a bowl of duplicated code.
Query parameters can (and should) be used for behavior control (for example to add pagination to a listing) but we should avoid using them for actual API evolutions, since these are usually of a permanent kind and therefore you don’t want to make it optional for the consumer.
Versioning seems like a good idea. It allows for a clean way to evolve the API, it keeps old clients separated from new ones and provides a generic base from all kinds of changes that will occur during your API lifespan. On the other hand it also introduces a bit of complexity, specially if you will have different calls at different versions. Your clients would end up having to manage your API evolution themselves by upgrading a call, instead of the API. It’s like instead of upgrading a library to the next version, you would upgrade only a certain class of that library. This can easily turn into a version nightmare…
To overcome this we must ensure that our versions cover the whole API. This means that I should be able to call every available method on /v1 using /v2. Of course that if a newer version on a given method exists on v2 it should be run on the /v2 call. However, if a given method hasn’t changed in v2, I expect that the v1 version would seamlessly be called.
Inheritance based API Versioning
In order to achieve this we can take advantage of Java objects polymorphic capabilities. We can build up API versions in a hierarchical way so that older version methods can be overridden by newer, and calls to a newer version of an unchanged method can be seamlessly fallen back to it’s earlier version.
So back to our example we could build up a new version of the create method so that the API would look like this:
@Path("/api/v1/appointments") //We add a version to our base path
public class AppointmentsAPIv1 { //We add the version to our API classes
@GET
@Path("/{id}")
public AppointmentDTO getAppointment(@PathParam("id") String id) { ... }
@POST
public void createAppointment(AppointmentDTO appointment) {
//Your old way of creating Appointments only with names
}
}
//New API class that extends the previous version
@Path("/api/v2/appointments")
public class AppointmentsAPIv2 extends AppointmentsAPIv1 {
@POST
@Override
public void createAppointment(AppointmentDTO appointment) {
//Your new way of creating appointments with guests
}
}So now we have 2 working versions of our API. While all the old clients that didn’t yet upgrade to the new version will continue to use v1 – and will see no changes – all your new consumers can now use the latest v2. Note that all these calls are valid:
| Call | Result |
|---|---|
GET /api/v1/appointments/123 | Will run getAppointment on the v1 class |
GET /api/v2/appointments/123 | Will run getAppointment on the v1 class |
POST /api/v1/appointments | Will run createAppointment on the v1 class |
POST /api/v2/appointments | Will run createAppointment on the v2 class |
This way any consumers that want to start using the latest version will only have to update their base URLs to the corresponding version, and all of the API will seamlessly shift to the most recent implementations, while keeping the old unchanged ones.
Caveat
For the keen eye there is an immediate caveat with this approach. If your API consists of tenths of different classes, a newer version would imply duplicating them all to an upper version even for those where you don’t actually have any changes. It’s a bit of boiler plate code that can be mostly auto-generated. Still annoying though.
Although there is no quick way to overcome this, the use of interfaces could help. Instead of creating a new implementation class you could simply create a new Path annotated interface and have it implemented in your current implementing class. Although you would sill have to create one interface per API class, it is a bit cleaner. It helps a little bit, but it’s still a caveat.
Final thoughts
API versioning seems to be a current hot topic. Lot of different angles and opinions exists but there seems to be a lack of standard best practices. While this post doesn’t aim to provide such I hope that it helps to achieve a better API structure and contribute to it’s maintainability.
A final word goes to Roberto Cortez for encouraging and allowing this post on his blog. This is actually my first blog post so load the cannons and fire at will. :)
| Reference: | REST API Evolution from our JCG partner Decio Sousa at the Roberto Cortez Java Blog blog. |
