datetime.datetime should have a timestamp() method

[djc]

BPO	14908
Nosy	@malemburg, @tim-one, @warsaw, @jribbens, @amauryfa, @mdickinson, @abalkin, @pitrou, @catlee, @vstinner, @djc, @bitdancer
Superseder	bpo-2736: datetime needs an "epoch" method
Files	issue14908.diff

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = 'https://github.com/abalkin'
closed_at = <Date 2012-06-08.16:38:21.223>
created_at = <Date 2012-05-25.08:53:19.672>
labels = ['type-feature', 'library']
title = 'datetime.datetime should have a timestamp() method'
updated_at = <Date 2012-06-08.16:38:21.222>
user = 'https://github.com/djc'

bugs.python.org fields:

activity = <Date 2012-06-08.16:38:21.222>
actor = 'belopolsky'
assignee = 'belopolsky'
closed = True
closed_date = <Date 2012-06-08.16:38:21.223>
closer = 'belopolsky'
components = ['Library (Lib)']
creation = <Date 2012-05-25.08:53:19.672>
creator = 'djc'
dependencies = []
files = ['25865']
hgrepos = []
issue_num = 14908
keywords = ['patch']
message_count = 16.0
messages = ['161558', '161568', '162263', '162267', '162268', '162269', '162270', '162278', '162280', '162282', '162288', '162316', '162452', '162483', '162531', '162534']
nosy_count = 26.0
nosy_names = ['lemburg', 'tim.peters', 'ping', 'barry', 'jribbens', 'guettli', 'amaury.forgeotdarc', 'mark.dickinson', 'davidfraser', 'belopolsky', 'pitrou', 'andersjm', 'catlee', 'vstinner', 'tomster', 'werneck', 'djc', 'hodgestar', 'Neil Muller', 'erik.stephens', 'steve.roberts', 'r.david.murray', 'vivanov', 'python-dev', 'Jay.Taylor', 'Alexandre.Zani']
pr_nums = []
priority = 'normal'
resolution = 'duplicate'
stage = 'resolved'
status = 'closed'
superseder = '2736'
type = 'enhancement'
url = 'https://bugs.python.org/issue14908'
versions = ['Python 3.3']

[djc]

There's datetime.fromtimestamp() and datetime.timetuple(), but no datetime.timestamp(). It should be possible to round-trip a UNIX timestamp through a datetime.datetime.

[AlexanderBelopolsky]

See bpo-2736.

[warsaw]

I completely agree. As easy (but obscure) as it is, it seems quite silly to have to go through time.mktime(dt.timetuple())

[abalkin]

Barry,

Did you review the long discussion under bpo-2736? Specifically, please note the part about mxDateTime, ticks() and gmticks(). I am -0 on adding ticks() and gmticks() and as far as I can tell no one has implemented these methods for datetime without introducing some subtle bugs.

I also suspect that those who ask for a timestamp() method expect that it will do more than

EPOCH = datetime(1970, 1, 1)
def timestamp(t):
      return (t - EPOCH).total_seconds()

For example, account for leap seconds or magically guess the DST offset. If we introduce a method that implements the code above, we have to prepare to explain why it does not do those things.

If instead we introduce a shortcut for time.mktime(t.timetuple()), we will have to deal with the fact that mktime() behaves differently on different systems.

[warsaw]

On Jun 04, 2012, at 03:45 PM, Alexander Belopolsky wrote:

Did you review the long discussion under bpo-2736? Specifically, please
note the part about mxDateTime, ticks() and gmticks(). I am -0 on adding
ticks() and gmticks() and as far as I can tell no one has implemented these
methods for datetime without introducing some subtle bugs.

I skimmed it, but I don't have time to read the whole bug. I'll note that
I've recently had to convert a bunch of code from using mxDateTime to
datetime+time, for the only reason that we had to support Python 3 and mxDT
isn't yet ported.

I also suspect that those who ask for a timestamp() method expect that it
will do more than

EPOCH = datetime(1970, 1, 1)
def timestamp(t):
return (t - EPOCH).total_seconds()

Not really.

For example, account for leap seconds or magically guess the DST offset. If
we introduce a method that implements the code above, we have to prepare to
explain why it does not do those things.

That's fine, explain away!

If instead we introduce a shortcut for time.mktime(t.timetuple()), we will
have to deal with the fact that mktime() behaves differently on different
systems.

Not really. If developers even care, they will know that the epoch is
different for different systems. We don't have to deal with it, we just have
to mention it. Let them use more verbose "correct" code if they care.

Seriously, we should not over-engineer this. It's purely a more discoverable
convenience and I think implementing the moral equivalent to
time.mktime(t.timetuple()) is exactly the right thing to do.

[AlexanderBelopolsky]

On Mon, Jun 4, 2012 at 11:57 AM, Barry A. Warsaw <report@bugs.python.org> wrote:

I've recently had to convert a bunch of code from using mxDateTime to
datetime+time,

That's a valuable experience. How big of a deal was the lack of
.ticks() and .gmticks()? How did you work around it?

[warsaw]

On Jun 04, 2012, at 04:03 PM, Alexander Belopolsky wrote:

That's a valuable experience. How big of a deal was the lack of
.ticks() and .gmticks()? How did you work around it?

It was *much* less of an issue than all the magic date format parsing that
mxDT supports. That's actually something that I think is a more serious
deficiency, since .strptime() is pretty limited (e.g. how would you parse ISO
8601 dates both with and without the 'T'?). A Python 3 compatible time format
parser would make for a very nice separate library (PyPI) and/or addition to
the stdlib (eventually).

FWIW, here's the changeset.

http://bazaar.launchpad.net/~gwibber-committers/gwibber/trunk/revision/1354

[AlexandreZani]

I think the easiest and most intuitive approach is to simply define timestamp() as being the reverse of fromtimestamp(). Don't worry about leap seconds and all that stuff.

If non-1970 epochs are a concern, this could be renamed to posix_timestamp or some such with perhaps a generic timestamp function that takes both a time and epoch.

But let's not let such a useful function not happen just because it won't solve everyone's problem.

[AlexanderBelopolsky]

On Mon, Jun 4, 2012 at 1:33 PM, Alexandre Zani <report@bugs.python.org> wrote:

I think the easiest and most intuitive approach is to simply define timestamp()
as being the reverse of fromtimestamp().

I would like to invite everyone to review the discussion leading to
closing of bpo-2736. We cannot implement "the reverse of
fromtimestamp()" because fromtimestamp() is not reversible in presence
of DST and because float cannot represent all values that datetime
supports. Both issues can be resolved, but I did not see any solution
that I would call intuitive.

Is anyone motivated enough to port mxDT's ticks() method to datetime?
I don't think we need to reinvent the wheel.

[AlexandreZani]

I'm still reading through the issue you mentioned. (It's a painful read I have to admit) One major obstacle seems to be that during the DST switch over, an hour gets repeated and so the datetime object is ambiguous. (are you on the first or second hour?) I would argue that it isn't a problem that this function needs to solve. The ambiguity isn't tied to the conversion. It's tied to the datetime object itself.

Let's add an optional parameter to specify the DST status, doc the pitfall and not worry overmuch about it.

Side note: Let me know if I misunderstood Alexander, but if I didn't this should be documented with the datetime object. Based upon my understanding, the datetime object is a bad choice if you care about that ambiguity. That's not really clear.

[AlexanderBelopolsky]

On Mon, Jun 4, 2012 at 3:21 PM, Alexandre Zani <report@bugs.python.org> wrote:

Let me know if I misunderstood Alexander, but if I didn't this should be documented with
the datetime object. Based upon my understanding, the datetime object is a bad choice
if you care about that ambiguity. That's not really clear.

The datetime object is neither good nor bad if you want to deal with
the DST ambiguity. If you want to store times as local time, the best
practice is to include the UTC offset. The datetime module now
provides the facilities to do so. If you neglect the timezone and
deal with naive datetime objects instead, you should probably avoid
mixing it with POSIX timestamps in the same application.

Here is an example where naive datetime is quite appropriate: a
timekeeping application for a small business with a single location.
If you need to generate opening hours for a given week,
[datetime(y,m,d, 9, 0) + datetime(i) for i in range(5)] is an adequate
solution, but if you compare it to [datetime.fromtimestamp(x +
24*3600*i) for i in range(5)], you may be off by one hour if your week
spans the DST change.

In most applications, however, keeping local time without UTC offset
is a bad choice. POSIX's timestamp only advantage is that it does no
allow keeping local time at all.

[abalkin]

Barry,

I looked at your changeset at http://bazaar.launchpad.net/~gwibber-committers/gwibber/trunk/revision/1354 and at the specification at http://developers.facebook.com/docs/reference/api/event/.

Am I correct that you are parsing "string containing an ISO-8601 formatted date/time or a UNIX timestamp; if it contains a time zone (not recommended), it will be converted to Pacific time before being stored and displayed"?

If so, it looks like converting input to POSIX timestamps is the wrong thing to do because users who specify naive time expect it to be displayed without conversion.

The way I read the specs, a better implementation would store input in naive datetime objects, converting UNIX timestamps or TZ-aware strings to Pacific timezone.