How do you test a product with distributed logic and non-deterministic components? Well, it depends on the product and when the question is asked. This post is one answer, for now.
A service I work on exposes an API for internal and external clients and calls multiple other services. It orchestrates a reasonably complex journey for a user, with their inputs and the responses of the services determining each step.
Our customers can configure the service to provide a variant of the standard journey for their users. This gives them choices on the available steps, the order of the steps, and, importantly, some of the logic for taking particular steps. This configuration lives in a different repository to the service, not owned by us, and it can be changed and deployed independently from the service.
As you can imagine, we test our service in multiple ways at different granularities using the usual tooling, including unit tests for the core pieces of functionality of the service, integration tests that hit the service's endpoints in different scenarios, and end-to-end tests that run against a live service and dependencies.
The usual trade-offs of all of those approaches apply but the concern that has exercised us for a while is the gap between those tests and the additional logic described in the configurations. That, and the fact that some of the other services are not owned by us, not hosted by us, or are non-deterministic because LLMs.
Naturally we have in-repo checks to cover variant configurations, but it's the combination of service and configuration that determine the available paths and the scenarios in which specific options are allowed. If we change either, then we'd like to know that the journeys our customers want are still available, under the conditions they have specified, long before that combination hits production. (And let's take it as read that we can't move the external logic into our world at the moment.)
What we felt we needed was something along these lines: use real configurations from a nominated branch, use a nominated service version, hit the API not the FE, run locally or in CI, and cover all customer configurations ... but target the important special cases only. Also the usual stuff: be readable, understandable, maintainable, and extensible; provide useful debugging information; run reliably and quickly.
Where we're at currently is a test suite in the service's repo using a single Dockerised Wiremock instance for all of the dependency services. In CI we run our service in the Docker container too. There's a thin DSL for journey navigation and a clear structure to the tests and test data.
The core of this setup is that mocking makes the runs fast and deterministic; the control we have with the mocks means we can target very specific paths; collecting mocked responses by scenario makes them much easier to find and read than having them buried code; using a single Wiremock instance makes housekeeping straightforward.
Allowing any service version to run against any set of configurations gives us an easy way to compare any combo that's interesting at any point and restricting to just the key scenarios (we hope) helps us to keep a lid on maintenance.
--00--
To sketch in some detail, Wiremock's standalone server deals with two different file types: stub responses (in a mappings directory) and JSON payloads (in __files). Stub files map an incoming URL to an outgoing response, optionally using a payload from __files.
Here's how our Wiremock configuration is structured:
Each scenario has its own response files for the services that are involved in that scenario. For example, scenario 2 involves two other services, but scenario 3 only one.
We write stub files per service, setting them up so that their response payload comes from a file of the same name, so mappings/service1.json returns the contents of current/service1.json.
The __files/current directory is mounted by the Docker container. As part of the setup for each scenario, we copy the relevant files into it and ask Wiremock to re-read them. Although Wiremock provides an admin API for updating responses, we found that copying files on disk was more reliable for us even it it's frankly pretty clunky.
The test code has functions like the ones below in classes per scenario, using operators like submit from our DSL to move through steps in the journey and standard test framework assertions on expected state.
@ConfigurationTest(only = [Config.A, Config.B])
fun Scenario2_AB(config: Configuration) {
setMockResponses("scenario2")
journey(config) {
expect { ... }
submit(...) expect { Result_AB }
...
}
}
@ConfigurationTest(only = [Config.C, Config.D])
fun Scenario2_CD(config: Configuration) {
setMockResponses("scenario2")
journey(config) {
expect { ... }
submit(...) expect { Result_CD }
...
}
}The setMockResponses() call does test setup and the @ConfigurationTest custom annotation arranges a parameterised test and the values to run it on. In the example above, the only key shows clearly which configs can share a particular test function and in the example below, the except keyword specifies the case that is different to the rest.
@ConfigurationTest(except = [Config.E])
fun Scenario3_Default(config: Configuration) {
setMockResponses("scenario3")
journey(config) {
...
}
}
@ConfigurationTest([Config.E])
fun Scenario3_E(config: Configuration) {
setMockResponses("scenario3")
journey(config) {
...
}
}The mock responses are JSON payloads and are as real as they need to be to exercise the particular logic combination that the scenario cares about and otherwise as self-describing as they can be. See Real vs Clear for more about my position on that.
--00--
The suite we've built runs quickly through a range of scenarios and is notably not flaky, unlike the E2E tests against live services that we run alongside it using shared machinery for navigating the user journeys.
The speed comes from two places: first the mocks return instantly unlike the high latency on some of the external calls and, second, we can use the mocks to shorten the journeys. If the logic we want to test exists on say, step 13 of a journey, the mocked test can receive the state from step 12 as its first response, effectively skipping all other steps. This acceleration is powerful because it also makes the test logic, data, and crucially intent, much easier to comprehend.
Interestingly, or perhaps predictably and depressingly, one of the headaches around our team discussions about the gap this suite fills was naming. We didn't have a shared language for talking about whatever kinds of tests these are, and it took time for the vocabulary to (a) be suggested, (b) be agreed upon, and (c) be used consistently.
What we've ended up calling them is isolated tests and they sit as siblings of the E2E tests I mentioned in a hierarchy of API tests. There's still some quibbles about the appropriateness of "E2E" given that those tests don't touch the user interface or necessarily complete a user journey once the desired logic is covered. I think we've made our peace with that though.
For me, it remains to be seen whether the approach scales. The thing is readable today but there are more scenarios to add, including migrating some things covered by the E2E tests, and there will be new feature and configuration combinations to test in future.
Exhaustive enumeration of files per scenario aids readability, but there's the later tension with maintainability to consider. I am expecting that the APIs involved are sufficiently stable and future changes will be sufficiently predictable, that widespread edits to the test data will be rare and can be automated easily. A one-time cost for that against the every-time cost that hard-to-read code and data impose.
--00--
The problem I posed at the top was how to test a product with distributed logic and non-deterministic components?
If I had to summarise the approach we've iterated to it's this:
- identify a place to exercise the combination from
- find a way to deliver the scenarios we want to test
- drive the combination to the important states
- check the important features of that state
Yes, at its heart, this just the standard Arrange, Act, Assert. Despite that, I can imagine some people will object to the approach because it's deliberately not hitting live services.
I take that point, but I think I've shown that we've thought hard about what we're doing and, to be honest, the other alternatives we've tried gave us nothing of value.
Better isolate than never, you might say.
Code highlighting: pinetools