Error inspection proposal has two functions - Is and As. Both iterate over a chain of errors, both attempt to find an error in that chain. The only difference in those functions is that one works with values and another with types or so it seems.
Problem
When you use sentinel errors and try to wrap an error with a sentinel error its impossible to implement it in a way that allows Is to find that sentinel error in the error chain. You can see the problem here. The problem stems from the fact that Is as currently proposed interacts with errors only in two ways - equality comparison and Unwrap method from Wrapper interface. To find an error in the chain you need both target error and error in the chain to be equal. To wrap an error you need to be able to actually store wrapped error somewhere. And you can't do both - you are either equal to an error that doesn't know anything about what it wrapping or you store an error that you wrap and stop being equal to the sentinel error you attempt to find.
With As and using typed errors you can easily test for errors wrapping another error like here.
Is behaviour is not a problem by itself - it might be exactly the right thing to do. The problem is that Is and As behave differently while being so similar on paper - their names are very alike, arguments are quite similar, both do about the same thing and have very similar documentation.
When you read the docs nothing suggests that their behaviour is different, you expect that they do exactly the same thing on exactly the same data (chain of errors), one works with types and another with values.
Solutions
There are several ways to solve the issue:
- Remove
Isaltogether. - Rename
Isto something else (Cause, whatever) and state specifically that it cannot find sentinel errors that wrap another error and that such behaviour is unwanted. - Allow errors to implement optional
Ismethod.
Opinion
I don't think option 1 is viable as you have legitimate reasons for Is to exist and all it will accomplish is force people to implement their own Is.
Option 2 solves the issue nicely. It is especially good if you consider that in general typed errors are better than sentinel errors (you cannot overwrite typed error like a sentinel error, you can add behaviour to typed errors, etc). This option pushes people to use typed errors instead of sentinel errors. The problem is it might end up with the same result as option 1 - people will implement their own Is method.
Option 3 brings both Is and As on par in terms of functionality and behaviour but complicates things as is written in the original proposal.
Whatever the decision will be, if behaviour of Is is left as it is I strongly believe it should be better differentiated from As as to not cause confusion.