8 ms·

Error handling is, to me, the most important thing when writing any code. Promoting "if err != nil { return }" even more by giving it a keyword seems like a da

Error handling is, to me, the most important thing when writing any code. Promoting "if err != nil { return }" even more by giving it a keyword seems like a dangerous road to walk down. Are we proposing solutions to make developers not have to type "if err != nil" or are we looking at how we can help developers handle errors better and be more productive? I honestly don't have anything to offer as far as ideas for changes to the language, but I really hope we think through the long-term ramifications of these changes.

jchw 7y ago
I think the way Go error handling works is fantastic, and one of the few problems it has is repetition. There’s C++ macros for various environments that let you do the assign-or-propagate-error stuff with a bit less repetition, but I dislike them because they feel opaque; people don’t often feel the potential consequences and just use it as a way to not think about error handling. With Go, the error handling behavior is so obvious and in your face that it’s painful. (This doesn’t stop bad error handling hygiene, but it has definitely helped me.) Sadly, there’s no obvious dumb way to reduce that repetition. I would not mind a language mechanism similar to defer, like the handle/check proposal for example. The most obvious issue I take with try is that it looks like a function call but is much more magical :( and also, I dislike named returns.
- marcus_holmes 7y ago
  I've spent the last year working on a Go system, writing Go every day. After a couple of months, I just stopped seeing/worrying about "if err != nil {". It has become punctuation, the Go equivalent of semicolons; I don't even use snippets; I manually type that every time. Which is good, because it does make me think about whether this function call can error, and what I should do about it if it does. 90% of the time it's ("just pass it up"), 5% of the time it's ("nothing, I don't actually care if this routine fails") but 5% of the time it's ("right, yes, this needs to be dealt with here"). When reading code, I just skip over it if it doesn't do anything interesting. I understand that in English reading, we don't notice "they said"; our brains just skip over it, and you can use "said" to open every quote and it won't feel repetitive. That's how I've got with "if err != nil {" so... I'm against this "try" stuff, because it feels like it's been proposed by people who haven't worked with the language too much. I'm not against making it more friendly for new people to learn, but this doesn't feel like that. This just feels like "ugh, really? I have to handle all the errors?!"...which is not a reason to change it.
  - pjmlp 7y ago
    I used to write code like that during the 80 and 90's, until settling down in languages with first class support for exceptions. So yeah, I did it for around 20 years, and don't miss it.
    - jchw 7y ago
      I don’t like exceptions at all. They do not make it obvious what is going on, I never am completely sure if I’m handling them right. When I want to throw an exception I’m often unsure which would be right, and sometimes your API can have multiple reasons to throw the same exception. Go error handling is not like that. But I can sure as hell say fairly that Go error handling is likely also not similar to what you did for 20 years in the 80s and 90s either. Rob Pike and friends surely knew a lot about what programming was like at that time. I, being relatively a youngster, don’t first hand, but I can tell you my experiences with C++, PHP, Python have not been nearly as good as Go with error handling. For one thing, C++ has no rigid standard for how to handle errors. Some people use exceptions, some used error methods on classes (including the standard library,) some used special integer or enumeration values (...including the standard library,) and some had libraries and frameworks have their own magic error handling mechanisms. This cognitive overhead was horrible. C wasn’t much better; atoi is a case study in why error handling in C sucks. Libraries that tried to standardize it, like SDL, were bearable if it was all you used, but it probably wasn’t, and some APIs, like Win32, made it even worse. (And I suppose it is worth at least mentioning setjmp/longjmp error handling. I don’t think it’s necessary to comment on why it’s not good.) Python exception handling is admittedly better, but its not really wonderful. Exception handling code in Python is prone to breakage that is first detected at runtime. If a function implementation changes, and the set of exceptions it might throw changes, that’s an invisible API change that may cause an unhandled exception in production. Not so great. Also, on a vaguely related note, you can’t really do error values using multiple returns like in Go, because Python doesn’t support multiple returns, only tuples, and refactoring between returning values and tuples is likely to run into accidental runtime errors (though you can paper over this issue a bit with type checking.) PHP error handling sucks, I will withhold from elaborating. All of the exception handling mechanisms suffer from one problem I really don’t like: it’s another nearly invisible part of the API. It makes the wrong thing (not handling errors) easy, and the right things (handling the appropriate errors correctly) hard. Your dependencies have to care about your call tree, and if it changes in refactoring it’s anyones bet what kinds of exceptions your function might throw. You could catch all exceptions, but because language errors like syntax errors (JavaScript) and index out of range and property name errors (Python) can also be exceptions, you rarely want to catch all exceptions. Not to mention, your call itself would be caught, so any exceptions caused by anything else in the try block would also be conflated. Go does some things that are mostly not new, but haven’t all been packaged together this way before exactly: - Custom errors via implicit interfaces, allowing easy, arbitrary data to be passed through errors while maintaining full control of error messages - Deep separation of programming errors and operational errors tend to be passed as error values. Programming errors, like indices being out of bound or misusing an API, typically results in a panic, whereas operational errors. Very seldom do you actually care what the error is, but when you do you can inspect the error as any other value, because it is just any other value. - Ecosystem-wide standards for how to pass errors. It’s almost 100% universal that errors are passed at the end of the return list. This makes it easy to parse for humans and easy to lint for machines. Linters can warn you about unused error values, and if a function suddenly has an error return it’s an API break, forcing you to fix existing code to properly handle errors. - Good library support for error types, including fmt.Errorf for one-off errors that don’t need special handling, and (third party) a myriad of error wrapping/helper libraries. They’re not needed at all, but can be quite handy. It works a lot better imo. You have more ceremony but less guessing. You can read a function and see almost every edge case, and when all of the functions perform good error hygiene you no longer need to guess about what refactoring your code will do. I’ll take my repetition.

jchw 7y ago

I think the way Go error handling works is fantastic, and one of the few problems it has is repetition. There’s C++ macros for various environments that let you do the assign-or-propagate-error stuff with a bit less repetition, but I dislike them because they feel opaque; people don’t often feel the potential consequences and just use it as a way to not think about error handling. With Go, the error handling behavior is so obvious and in your face that it’s painful. (This doesn’t stop bad error handling hygiene, but it has definitely helped me.) Sadly, there’s no obvious dumb way to reduce that repetition. I would not mind a language mechanism similar to defer, like the handle/check proposal for example. The most obvious issue I take with try is that it looks like a function call but is much more magical :( and also, I dislike named returns.

marcus_holmes 7y ago
I've spent the last year working on a Go system, writing Go every day. After a couple of months, I just stopped seeing/worrying about "if err != nil {". It has become punctuation, the Go equivalent of semicolons; I don't even use snippets; I manually type that every time. Which is good, because it does make me think about whether this function call can error, and what I should do about it if it does. 90% of the time it's ("just pass it up"), 5% of the time it's ("nothing, I don't actually care if this routine fails") but 5% of the time it's ("right, yes, this needs to be dealt with here"). When reading code, I just skip over it if it doesn't do anything interesting. I understand that in English reading, we don't notice "they said"; our brains just skip over it, and you can use "said" to open every quote and it won't feel repetitive. That's how I've got with "if err != nil {" so... I'm against this "try" stuff, because it feels like it's been proposed by people who haven't worked with the language too much. I'm not against making it more friendly for new people to learn, but this doesn't feel like that. This just feels like "ugh, really? I have to handle all the errors?!"...which is not a reason to change it.
- pjmlp 7y ago
  I used to write code like that during the 80 and 90's, until settling down in languages with first class support for exceptions. So yeah, I did it for around 20 years, and don't miss it.
  - jchw 7y ago
    I don’t like exceptions at all. They do not make it obvious what is going on, I never am completely sure if I’m handling them right. When I want to throw an exception I’m often unsure which would be right, and sometimes your API can have multiple reasons to throw the same exception. Go error handling is not like that. But I can sure as hell say fairly that Go error handling is likely also not similar to what you did for 20 years in the 80s and 90s either. Rob Pike and friends surely knew a lot about what programming was like at that time. I, being relatively a youngster, don’t first hand, but I can tell you my experiences with C++, PHP, Python have not been nearly as good as Go with error handling. For one thing, C++ has no rigid standard for how to handle errors. Some people use exceptions, some used error methods on classes (including the standard library,) some used special integer or enumeration values (...including the standard library,) and some had libraries and frameworks have their own magic error handling mechanisms. This cognitive overhead was horrible. C wasn’t much better; atoi is a case study in why error handling in C sucks. Libraries that tried to standardize it, like SDL, were bearable if it was all you used, but it probably wasn’t, and some APIs, like Win32, made it even worse. (And I suppose it is worth at least mentioning setjmp/longjmp error handling. I don’t think it’s necessary to comment on why it’s not good.) Python exception handling is admittedly better, but its not really wonderful. Exception handling code in Python is prone to breakage that is first detected at runtime. If a function implementation changes, and the set of exceptions it might throw changes, that’s an invisible API change that may cause an unhandled exception in production. Not so great. Also, on a vaguely related note, you can’t really do error values using multiple returns like in Go, because Python doesn’t support multiple returns, only tuples, and refactoring between returning values and tuples is likely to run into accidental runtime errors (though you can paper over this issue a bit with type checking.) PHP error handling sucks, I will withhold from elaborating. All of the exception handling mechanisms suffer from one problem I really don’t like: it’s another nearly invisible part of the API. It makes the wrong thing (not handling errors) easy, and the right things (handling the appropriate errors correctly) hard. Your dependencies have to care about your call tree, and if it changes in refactoring it’s anyones bet what kinds of exceptions your function might throw. You could catch all exceptions, but because language errors like syntax errors (JavaScript) and index out of range and property name errors (Python) can also be exceptions, you rarely want to catch all exceptions. Not to mention, your call itself would be caught, so any exceptions caused by anything else in the try block would also be conflated. Go does some things that are mostly not new, but haven’t all been packaged together this way before exactly: - Custom errors via implicit interfaces, allowing easy, arbitrary data to be passed through errors while maintaining full control of error messages - Deep separation of programming errors and operational errors tend to be passed as error values. Programming errors, like indices being out of bound or misusing an API, typically results in a panic, whereas operational errors. Very seldom do you actually care what the error is, but when you do you can inspect the error as any other value, because it is just any other value. - Ecosystem-wide standards for how to pass errors. It’s almost 100% universal that errors are passed at the end of the return list. This makes it easy to parse for humans and easy to lint for machines. Linters can warn you about unused error values, and if a function suddenly has an error return it’s an API break, forcing you to fix existing code to properly handle errors. - Good library support for error types, including fmt.Errorf for one-off errors that don’t need special handling, and (third party) a myriad of error wrapping/helper libraries. They’re not needed at all, but can be quite handy. It works a lot better imo. You have more ceremony but less guessing. You can read a function and see almost every edge case, and when all of the functions perform good error hygiene you no longer need to guess about what refactoring your code will do. I’ll take my repetition.