JSON timestamps are not valid ISO 8601

[normanr]

b726e2f added a ISO 8601-like output with a Z (indicating UTC) and %z (offset from UTC). I don't think the Z should be present (also the man page doesn't include the Z in the example output).

Currently json_print does a special workaround after printing microseconds to replace the Z (where snprintf wrote it's null byte). It's probably simplest to just drop the microseconds from the time format, and call strftime a second time to format the timezone after the microseconds have been printed, eg:

char *bufp = buf;
int buf_size = sizeof(buf);
int n;

n = strftime(bufp, buf_size, "%Y-%m-%dT%H:%M:%S.", ti);
bufp += n; buf_size -= n;

n = snprintf(bufp, buf_size, "%06d", ns/1000);
bufp += n; buf_size -= n;

strftime(bufp, buf_size, "%z", ti);

I think it would also make sense to move the strftime/snprintf calls before the #ifdef WITH_CJSON because it can be shared with both parts. Also buf could be made smaller, only the first 32 bytes will ever be used.

[ralight]

Thanks @normanr, I agree with your points. I've implemented it slightly differently because it's a tiny fraction more efficient this way.

[normanr]

note: the buffer is still over-sized (100 bytes), it could be smaller (maybe as small as 32 bytes, worth double-checking or giving a small amount of slack).

[normanr]

some other notes:

• I'm not sure why the size passed to snprintf is 9. That ends half-way through the timezone identifer.
• saving and restoring the overwritten char makes an assumption that the formatted string will always be the same length, I guess that's okay. I'm not sure it's worth the run-time efficiency, over having the code be safer. Is it a very hot code path?

[ralight]

Yes, the buffer is oversized, I'll drop it but I don't think it's a big deal.

The size passed to snprintf is 9 because the compiler thinks the int could 9 characters long even though we know it will be no more than 6 characters long.

If strftime doesn't give the same length output every time it's a huge bug in strftime.

[normanr]

These are both just nits, thanks for the fix!

The size passed to snprintf is 9 because the compiler thinks the int could 9 characters long even though we know it will be no more than 6 characters long.

int32 can be 10 characters, so not sure where the 9 comes from.

If strftime doesn't give the same length output every time it's a huge bug in strftime.

y10k problem 😀 but fair point. iso 8601 requires "prior agreement between the sender and the receiver", when the year is outside [0000, 9999], so unlikely to be an issue for a while.