cmd/vet: vet should recommend "os.IsNotExist(err)" over "err == os.ErrNotExist"

[go101]

It looks the ErrNotExist variable in os package shouldn't be exported at all.

[ianlancetaylor]

Is this mistake really common enough that it deserves a vet check?

[josharian]

@dominikh

[dominikh]

In my own corpus (most public Go packages), there are 229 matches for [!=]= os\.ErrNotExist and 36618 matches for os\.IsNotExist\(.

Among the 229 results are matches in camlistore, rkt and docker.

[go101]

@ianlancetaylor, I seldom write file io programs. I think it is so natural that use "err == os.ErrNotExist" to check if a file exists or not when I used the io.Open function the first time. But I was wrong.

[bradfitz]

Among the 229 results are matches in camlistore, rkt and docker.

But in Camlistore we have interfaces documenting that os.ErrNotExist is the contract between our own interfaces and our own implementations.

A sample:

make.go:// The error is os.ErrNotExist if GOPATH is unset or the directory
make.go:                return path, os.ErrNotExist
make.go:        return path, os.ErrNotExist
pkg/blob/fetcher.go:    // os.ErrNotExist should be returned for the error (not a wrapped
pkg/blob/fetcher.go:    // offset or length is negative, or os.ErrNotExist if the blob
pkg/blobserver/b2/b2.go:                return nil, 0, os.ErrNotExist
pkg/blobserver/blobpacked/subfetch.go:// check. The returned error should be os.ErrNotExist if the blob

I think this vet check would be noisier than it would be helpful.

@robpike?

[robpike]

I don't think this merits a vet check. Remember it must satisfy the criteria specified in the readme of correctness, frequency, and precision. It fails for correctness and is marginal for frequency and precision.