As InVision continues its efforts to decompose the monolith into a distributed system, it has given me lots of opportunity to think about "breaking changes." A "breaking change" is any change that requires the consumer of an API to make lockstep changes in order for the consuming code to continue to work properly. We often think of breaking changes in regard to API method signatures, RESTful resources, and data structures; but, this is only part of the story. In reality, a "breaking change" is any change to an "aspect" of the API that the consumer has come to depend on. This aspect could be documented. It could also be implicit in the implementation. As such, I thought it would be a fun thought-experiment to sit down and noodle on all the changes that could be considered "breaking".
| || || |
| || |
| || || |
Changing a method or resource name.
Renaming a method or changing the location of a RESTful resource is a pretty obvious breaking change. Doing so will require the consumer to update their code in order to invoke the new method name or resource location.
NOTE: This includes the concept of removing or deleting such a dependency as well. Of course, any time we delete something that a consumer depends on, it's a breaking change.
Changing an invocation signature.
If you add optional parameters to an invocation target, that can be safe. But, if you add required parameters; or, if you add optional parameters that change the default behavior of the invocation; that is a breaking change. Doing so will require the consumer to update their invocation arguments in order to adapt to the change.
Changing a response structure.
When we talk about data structures, we often talk in terms of "additive" changes vs. "breaking" changes. The idea being that if you only add new properties to an existing structure, the consumer can safely ignore them at first; then, incrementally update the code to consume the new properties as needed.
But, is this always true? What if the consumer needs to serialize the data and store it in a database column that has a constrained character size? In such a case, even an "additive" change could cause an unexpected database truncation error.
Or, imagine that the response was being logged to a file. An increase in the response payload size could change the velocity with which log files take up disk-space. This, in turn, could cause issues if the log files aren't being rotated quickly enough.
Changing a response code.
Changing a response code is like changing a response data structure. If the consumer's logic examines the response code, the consumer may have to change their logic in order to cope with the change in response code.
But, what if you do something less obvious like implement ETag support. In such a case, an API that always responded with a 200 OK could suddenly start responding with a 304 Not Modified response. Such a change could cause runtime errors for a consumer that never had to deal with a functioning ETag workflow before.
Changing error types, messages, and other details.
When it comes to error handling, consumers implement all kinds of crazy-ass logic in order to understand the errors that are coming back from a system or module. This could mean looking at error Types and Messages. It could also mean performing string-parsing on the extended data provided in an error. As such, any change to the content of an error object could, theoretically, break the error handling logic in the consuming context.
Changing event types.
An event type is akin to a method signature or a RESTful resource location. As such, if you change an event type, consumers will have to change their code in order to listen for the new events.
Changing event payloads.
An event payload is akin to a response structure. As such, changing an event payload incurs the same problems as changing a response structure. In so much as an "additive" change will likely be OK in the vast majority of cases. But, that even an "additive" change can cause unexpected errors depending on how the consumer is processing the event.
Changing testable class hierarchies.
Changing the inheritance chain or prototype chain of an object is the same as changing its Type. As such, any consumer that inspects the type of an object, either declaratively as a Type annotation or imperatively with something like instanceof, will need to change their code in order to consume the new class definition.
Changing rate limiting rules.
Changing the rate limiting strategies for a service - or implementing rate limiting for the first time - may require a consumer to refactor their invocation patterns so as not to hit a rate-limit ceiling. Or, they may have to come up with totally new strategies if a rate-limit ceiling is too low. Or, they may have to upgrade to a more substantial API subscription with higher rate limits.
Changing caching strategies.
If a change in caching strategies is completely transparent, then it's not a breaking change. But, the moment there's a possibility that stale data can be presented to a client, that's a breaking change (assuming that the consumer is not expecting the possibility of and / or the magnitude of such staleness). So, for example, updating an API response to read from a database replica instead of the database master is a definitely a breaking change.
Changing the connaissance of time.
A consumer doesn't just consume end-points or methods, it consumes a "system". As such, if a system as a whole changes the way workflows are orchestrated internally, this could be considered a breaking change. Imagine a scenario in which a DELETE request triggers a "deleted" event that gets pushed to a publish-and-subscribe mechanism. If "deleted" events went from being published instantly to being published at night in a "batch job", the only thing that changed would be the "connaissance of time". But, if a consumer expected those events to be published in a more immediate fashion, the consuming code may have to change in order to cope with the new timing.
Ok, that's maybe a bad example. But, just think of the way in which we depend on the relative timing of related actions in a system. Imagine that your database replica lag suddenly went from sub-second times to several hours. It is very likely that you would have to change your application code to deal with such a change in timing.
Changing content-distribution (CDN) strategies.
There are many benefits to moving an API behind a Content-Delivery Network (CDN). Automatic compression; image optimization; DDOS protection; to name a few. But, a CDN adds "API" semantics in the sense that it injects its own HTTP headers into both the request and the response data. A consumer may come to depend on those injected HTTP headers. For example, the Cloudflare CDN injects a "country" header that can be used to help implement geofencing internally to an application. If the application context were to be moved to another CDN that no longer injected the same HTTP headers, both internal and external consumers may have to change in order to make up for the missing data.
This list of potential breaking changes is not intended to be exhaustive by any means. This is simply the list that I could conjure up while noodling on this topic. That said, I hope that this sheds light on the complexities of evolving an application while it's being actively consumed. We often limit our thinking of "breaking changes" to invocation signatures and response objects. But, the reality is, there's a much larger set of changes that can create problems for our consumers. Evolve your applications safely and conscientiously!
Another one that I forgot about it outlined in this post:
.... the implicit binding of your method calls is part of your API. So, if you object methods go from being pre-bound (to your object instance) to being "naked function references", that's a breaking change. A consumer would likely have to update their code to avoid runtime reference errors.
Another one that popped into my head:
Changing a service's SLA.
If you change an SLA that affects things like response-times, this could end up having down-stream changes, especially with regard to monitoring and alerting. If a service has increased latency (as a new baseline), then downstream clients may have to change their monitoring logic and alerting thresholds.
Changing the semantics of a parameter or returned value.
When you preserve the method's signature, but change the semantics of one or more of its parameters or return values. Example: an integer which previously returned a distance in feet now returns the same distance but expressed in metres.
To me, this is absolutely a breaking change because clients are dependent on the behaviour and must make changes to accommodate the change in meaning.
This is a really sneaky and subtle change, because preserving the method signature means that the change could go completely undetected by clients, and the effects of the mis-interpretation could ripple through the system.
There's nothing an automated system can do to protect us here; we rely on the API provider to communicate the change to its consumers. If we are relying on documentation alone to achieve this, then the documentation must cover semantic as well as syntactic behaviours. And even then, all consumers are expected to have read and understood the documentation!
That's a great example! In fact, I just ran into something like this
in Lucee CFML. On there version that I am on, if you query the
database and use a
NULL values are returned as
However, in a recent change to Lucee, those same
values are now being returned as empty strings.
The major problem with this is that the code in place is likely
already trying to handle the
NULL response. But, the
empty-string is not null. As such, the code that handles the
response will suddenly use a different control flow, which
could have a major impact.
Even minor changes like this could lead to very unexpected results.
Interesting, I'm glad you agree; I'm currently involved in a "debate" in which the other party maintains that as long as the method signature of the API is preserved, the change can be regarded as "non-breaking" for the consumer...!
That's just crazy-pants :D