Exceptions should be treated just like domain models. Each service works with their own domain models and should have their own set of exception models as well. When communicating with external systems, the service should convert external exceptions to its domain exceptions as soon as possible. Basically I'm saying go with solution #2.
Lets consider the communication from service A -> B. Service A should first of all have an interface defined to decouple the business logic from the implementation of requests to B. In your example A is an account service and B is a user service. So let's call the interface UserService. This interface would have a set of (ideally) compiler-checked exceptions.
interface UserService
def getUser(id): User throws UserNotFoundException, UserServiceException
You should implement HTTP client for service B so that any service that needs to depend on service B imports the common HTTP client. The error responses from requests to service B will be defined in this HTTP client component. That way they're only defined once.
class BHttpClient
def getUserById(id) =
response = http.get("/users/${id}").send
if (response.status == 404) throw new UnknownUserException
else if (response.status == 500) throw new InternalServerException
else return json.parse[User](response.content)
The implementation of UserService, HttpUserService will use that HTTP client to communicate with B, should catch HTTP and transport exceptions from the client and wrap them in the appropriate "domain" exception.
class HttpUserService(client: BHttpClient) implements UserService
def getUser(id) =
try {
client.getUserById(id)
} catch {
case e: UnknownUserException => throw new UserNotFoundException(e)
case e: InternalServerException => throw new UserServiceException(e)
}
Cons: Service A will need to catch and expect that Server B is capable of returning a bunch of errors, adding coupling between Service A and B.
Service A will catch errors from the http client in HttpUserService and wrap them in meaningful errors for service A. The business logic in service A is decoupled from service B through the UserService interface. HttpUserService is coupled to BHttpClient, but decoupled from service B because you can mock service B at the transport level.
Even if you choose to use a different architecture like @Laiv describes in the comments, you'll still want to decouple yourself from message and events you receive by converting the message models and exceptions into domain exceptions in each service. I don't agree with @Laiv, that it's as cut and dry as asynchronous message architecture or you might as well implement a monolith. There are still big gains that can be made by a synchronous, distributed service oriented architecture like you've described. The first and hardest step of getting the right architecture is to decouple the components. By dividing into microservices early, you can more easily adopt an asynchronous approach later if you need it.
