@rsc
I don't know if this is a bug or WAI, but it is nicely described.

If TZ is not set, the location is always set to Local. That seems reasonable and consistent. We don't want to try to figure out exactly which timezone corresponds to /etc/localtime. That is true even in the particular case that /etc/localtime happens to be UTC.

If TZ is set to a non-empty string, then that is used to determine the location. That also seems reasonable.

I think the only question here is how to handle TZ being set to the empty string. Right now that is the same as TZ being set to UTC. That is consistent with the C library behavior.

@ianlancetaylor thanks for you thoughts. Agreed on your 2nd and 3rd points.

We don't want to try to figure out exactly which timezone corresponds to /etc/localtime. That is true even in the particular case that /etc/localtime happens to be UTC

Respectfully, why not? Go already reads /etc/localtime in this case, so there's not any syscall being avoided.

For comparison, in C:

#include <stdio.h>
#include <time.h>

int main () {
   time_t rawtime;
   struct tm *info;
   char buffer[80];

   time(&rawtime);
   info = localtime(&rawtime);

   strftime(buffer,80,"%Z", info);
   printf("Timezone: %s\n", buffer);
}

$ doas ln -sf /usr/share/zoneinfo/UTC /etc/localtime
$ env -u TZ ./a.out
Timezone: UTC
$ env TZ="" ./a.out
Timezone: UTC
$ env TZ="UTC" ./a.out
Timezone: UTC

# strace of reads to /etc/localtime
$ env -u TZ strace ./a.out 2>&1 | grep localtime
openat(AT_FDCWD, "/etc/localtime", O_RDONLY|O_CLOEXEC) = 3
$ env TZ="" strace ./a.out 2>&1 | grep localtime
$ env TZ="UTC" strace ./a.out 2>&1 | grep localtime

Respectfully, why not? Go already reads /etc/localtime in this case, so there's not any syscall being avoided.

It's a fair question. My only answer is that is how the time package has worked since the current version was introduced in https://golang.org/cl/5392041 in 2011. It means that time.LoadLocation("Local") gives you a timezone whose name is Local that represents the local time zone. It lets us do the right thing on systems like Windows and Plan 9 that don't require the IANA database.

Changing this now would have a decent chance of breaking working Go programs. I think that any such change would have to through the proposal process (https://golang.org/s/proposal-process).

Changing this now would have a decent chance of breaking working Go programs. I think that any such change would have to through the proposal process (https://golang.org/s/proposal-process).

Sensible. I brought it up because it leads to some interesting scenarios, such as reflect.DeepEqual not considering the time from time.Now() and a parsed equivalent to be equal even if they've had their clocks rounded - the timezone can still differ. Of course, the answer is to use Equal, so it's a minor point.

As for a proposal, I don't think there's a way to change this in a backward compatible manner. I'm not even sure I'd suggest any changes to the existing documentation.

Perhaps, for this one, this closed issue thread will suffice. A future soul might save a few minutes when they're investigating the oddity.

Glad to know it's at least working as designed, Thanks!