Hmm, I'm not sure if I agree. This would once again break go test ./... on a clean clone.

It's true that some test successes are different from others, but that will generally be true for many other reasons. Some tests will require specific programs to be installed, or they might require Linux or Windows to run. Should all of those also fail unless they're actually running properly?

I think you should go the other way around. In all.bash, also run by CI, you could ensure that none of the portable tests are being skipped. You could do this with something like:

NO_DB_SKIP=true go test ./...

and in the Go code to handle the case where the DB isn't available:

if noDB {
    if os.Getenv("NO_DB_SKIP") != "true" {
        t.Fatal("some error")
    }
    t.Skip("skipping")
}

That way, you have the best of both worlds. It's possible for "run all the tests properly" to be done, and you're not messing with people who just want to run the tests on a best-effort basis.

A slight variant is what @bcmills suggested in the original issue, which is to make each TestMain start up a temporary postgres on-demand, and only do the skip-or-error dance when postgres isn't installed, which is less likely to be true than "postgres isn't running".

Ah nice, thanks! I like the NO_DB_SKIP suggestion 👍

Change https://golang.org/cl/239257 mentions this issue: all.bash,internal: force DB testing in CI