Обсуждение: Strategies for converting UTC data to local windows for arbitrary resolutions and timezones

On Tue, Oct 3, 2023 at 7:33 PM Lincoln Swaine-Moore <lswainemoore@gmail.com> wrote:

Hi all,

I'm attempting to generate some reports using user-driven timezones for UTC data, and I'm having trouble writing a query that meets the following criteria:

1) Data should be averaged at one of daily, hourly, or 15 minute granularities (user-driven).
2) Data over a given period should reflect the localized value of the UTC data for a user-provided time zone. E.g.:
    a) The 1 hour period starting at '2023-10-03 12:00:00.000000-0400' (America/New_York) should reflect data with timestamps between '2023-10-03 16:00:00.000000Z' and '2023-10-03 17:00:00.000000Z'.
    b) The 1 day period starting at '2023-10-03 00:00:00.000000-0400' should reflect data with timestamps between '2023-10-03 04:00:00.000000Z' and '2023-10-04 04:00:00.000000Z'.
3) When a period interacts with a DST change in the given timezone, the data should not be clumped together. E.g.:
    a) Data points occurring at 2023-11-05 05:30:00.000000Z and 2023-11-05 06:30:00.000000Z should be treated as falling into separate buckets when the time zone is America/New_York: (2023-11-05 01:30:00.000000-0400 and 2023-11-05 01:30:00.000000-0500, respectively). This should be true for either the 15 minute or 1 hour intervals.
    b) Some clumping for day resolution seems ok! E.g. the 1 day period starting at '2023-11-05 00:00:00.000000-0400' can and probably should contain 25 hours' worth of data. Certainly it should not reflect the data falling between '2023-11-05 00:04:00.000000Z' and '2023-11-05 00:04:00.000000Z' + '24 hours'::interval (= '2023-11-06 00:04:00.000000Z'), because that would be the local times of '2023-11-05 00:00:00.000000-0400' and '2023-11-04 23:00:00.000000-0500'.
4) It would be relatively simple to do 15 minute and 1 hour periods were all timezone offsets multiples of 1 hour (in that case, all operations could be done in UTC and then converted after the fact), but unfortunately some time zones have 30 min-based offsets, which interferes with this approach.
5) Ideally, the solution would not involve messing with the server/connection's value of timezone. (Though I would be interested if there was a solution that relaxed this constraint and was relatively safe/compatible with transactions and psycopg2.)
6) Ideally, my query would return periods that are missing data (though could plausibly fill these in in the layer above). This points toward generate_series but...
7) Sigh: I can't upgrade to 16. I gather that 16 has a timezone parameter for generate_series, which I believe might help. But tragically, Digitalocean doesn't yet support 16, and it's not practical to migrate elsewhere. Based on historical release -> support timing, I'd imagine they will not support it until Q2 2024, which is too late for this feature. If anyone had the inside scoop about when they'd likely support it, I'd welcome it!

This looks pretty hairy written out as above, but I actually think it reflects most people's intuitions about what data a local period "should" correspond to (though I'd welcome feedback on this point).

Here are some thoughts about approaches that I've tried, and what their drawbacks seem to be. For all these, I'll use the following CTE to demonstrate some data that crosses a DST boundary:

```
with original_data as (
    select
        ('2023-11-05 00:00:00.000000Z'::timestamptz) + (15 * x || ' minutes')::interval as "t"
    from
        generate_series(0, 1000) as x
)
```

1) date_trunc: it seems like as of v12, date_trunc accepts a third argument of timezone, which essentially plays the role of the server timezone setting for the scope of the function. This is very handy, and *I believe* solves my issues for the hour/day periods:
```
[etc]
select
    date_trunc('day', t, 'America/New_York'),
    min(t),
    max(t),
    count(*)
from original_data
group by 1
order by 1;

       date_trunc       |          min           |          max           | count
------------------------+------------------------+------------------------+-------
 2023-11-04 04:00:00+00 | 2023-11-05 00:00:00+00 | 2023-11-05 03:45:00+00 |    16
 2023-11-05 04:00:00+00 | 2023-11-05 04:00:00+00 | 2023-11-06 04:45:00+00 |   100
 2023-11-06 05:00:00+00 | 2023-11-06 05:00:00+00 | 2023-11-07 04:45:00+00 |    96
 2023-11-07 05:00:00+00 | 2023-11-07 05:00:00+00 | 2023-11-08 04:45:00+00 |    96
 2023-11-08 05:00:00+00 | 2023-11-08 05:00:00+00 | 2023-11-09 04:45:00+00 |    96
 2023-11-09 05:00:00+00 | 2023-11-09 05:00:00+00 | 2023-11-10 04:45:00+00 |    96

[etc]
```

This checks out, but unfortunately doesn't seem to work for 15 minutes. I think, by the way, that this behavior is identical to what I would've gotten if my server timezone was "America/New_York" and I ran it without the 3rd argument, though I'd be curious to hear if there are discrepancies.

On this point, I've read up on some of the history around this feature, and was a little puzzled by this assertion in this thread: https://www.postgresql.org/message-id/87in1k73nr.fsf@news-spur.riddles.org.uk:

> If you have a timestamp-in-UTC column and want to do a date_trunc in some other specified zone (that's not the session timezone), you need FOUR uses of AT TIME ZONE to do it correctly:
> date_trunc('day', col AT TIME ZONE 'UTC' AT TIME ZONE 'Asia/Kathmandu') AT TIME ZONE 'Asia/Kathmandu' AT TIME ZONE 'UTC'

I *think* that this doesn't comport with behavior I've seen, because at a DST boundary the initial `col AT TIME ZONE 'UTC' AT TIME ZONE 'Asia/Kathmandu'` would drop the offset, which from then on would be unrecoverable. For example:

```
select date_trunc('hour', '2023-11-05 05:30:00.000000'::timestamp AT TIME ZONE 'UTC' AT TIME ZONE 'America/New_York') AT TIME ZONE 'America/New_York' AT TIME ZONE 'UTC';
      timezone
---------------------
 2023-11-05 06:00:00

select date_trunc('hour', '2023-11-05 06:30:00.000000'::timestamp AT TIME ZONE 'UTC' AT TIME ZONE 'America/New_York') AT TIME ZONE 'America/New_York' AT TIME ZONE 'UTC';
      timezone
---------------------
 2023-11-05 06:00:00
```

This is what I meant above by "clumping" in point (3). I believe this issue is ameliorated by setting the database timezone:

```
set timezone to 'America/New_York';
select date_trunc('hour', '2023-11-05 06:30:00.000000'::timestamp AT TIME ZONE 'UTC');
       date_trunc
------------------------
 2023-11-05 01:00:00-05

select date_trunc('hour', '2023-11-05 05:30:00.000000'::timestamp AT TIME ZONE 'UTC');
       date_trunc
------------------------
 2023-11-05 01:00:00-04
```

The same (correct, in my opinion) behavior seems to occur when using the third argument to date_trunc, so all is well on that front, but I'm being pedantic about this because if I was misunderstanding and usage of four `at time zone`s was a legitimate strategy, it could plausibly be applied to the below. Hoping to be incorrect here!


2) date_bin: this will take a "stride", which should accommodate all the periods I'm interested in, and "origin" to help deal with initial day/hour offsets. But I'm having trouble getting it to respect time zone/DST stuff. For example:

```
[etc]
select
    date_bin('24 hours', t, '2023-11-03 00:00:00.000000-0400'::timestamptz),
    min(t),
    max(t),
    count(*)
from
    original_data
group by 1
order by 1;

        date_bin        |          min           |          max           | count
------------------------+------------------------+------------------------+-------
 2023-11-04 00:00:00-04 | 2023-11-04 20:00:00-04 | 2023-11-04 23:45:00-04 |    16
 2023-11-05 00:00:00-04 | 2023-11-05 00:00:00-04 | 2023-11-05 22:45:00-05 |    96
 2023-11-05 23:00:00-05 | 2023-11-05 23:00:00-05 | 2023-11-06 22:45:00-05 |    96
 2023-11-06 23:00:00-05 | 2023-11-06 23:00:00-05 | 2023-11-07 22:45:00-05 |    96
 2023-11-07 23:00:00-05 | 2023-11-07 23:00:00-05 | 2023-11-08 22:45:00-05 |    96
 2023-11-08 23:00:00-05 | 2023-11-08 23:00:00-05 | 2023-11-09 22:45:00-05 |    96
 2023-11-09 23:00:00-05 | 2023-11-09 23:00:00-05 | 2023-11-10 22:45:00-05 |    96
[etc]
```

As you can see, the days get "bumped" by DST, and are off by one (until the spring). I actually think this makes intuitive sense for this concept of "stride"--it just means that it's not as useful for timezones with DST.


3) generate_series: If I could get this working, it would be ideal, because it would also help fill in gaps in my data with null rows. But again, I can't seem to, except in v16, have it respect timezones other than the server's:

```
set timezone to 'UTC';
select
    days as start_time,
    lead(days) over (order BY days) as end_time
from generate_series(
    '2023-11-03 00:00:00.000000-0400'::timestamptz,
    '2023-11-07 00:00:00.000000-0500'::timestamptz,
    '1 day'::interval
) days;

       start_time       |        end_time
------------------------+------------------------
 2023-11-03 04:00:00+00 | 2023-11-04 04:00:00+00
 2023-11-04 04:00:00+00 | 2023-11-05 04:00:00+00
 2023-11-05 04:00:00+00 | 2023-11-06 04:00:00+00
 2023-11-06 04:00:00+00 | 2023-11-07 04:00:00+00
[etc.]


set timezone to 'America/New_York';
[same as above]

       start_time       |        end_time
------------------------+------------------------
 2023-11-03 00:00:00-04 | 2023-11-04 00:00:00-04
 2023-11-04 00:00:00-04 | 2023-11-05 00:00:00-04
 2023-11-05 00:00:00-04 | 2023-11-06 00:00:00-05
 2023-11-06 00:00:00-05 | 2023-11-07 00:00:00-05
[etc.]
```

The latter being correct for these purposes, but not seeming super reliable/practical (see: point 5).

Can anyone think of any other approaches to this problem? I'd be thrilled if I could manage it without resorting to manually post-processing in pandas or something, because I have a number of datapoints, and each individual one is quite wide.

Apologies for the length of this message--just trying to be thorough. I sincerely appreciate any help or pointers!

Best,
Lincoln

--

Lincoln Swaine-Moore

That's a long email to digest but a couple thoughts that may help.

I've always disliked the term "timestamp with time zone" and always mentally translate that to "point in time". As long as data is stored as a timestamp with time zone (point in time), it can be manipulated in any time zone you desire including handling DST.

Time calculation always involves some assumption of meaning and PostgreSQL makes some reasonable ones. I'm in US Pacific time and if I run:

select '2023-11-05'::timestamptz ;

I get:                    
     timestamptz        
------------------------
2023-11-05 00:00:00-07

If I add a day (crossing the DST boundary):

select '2023-11-05'::timestamptz + '1 day'::interval;
       ?column?         
------------------------
2023-11-06 00:00:00-08

Observe that I end up at midnight the following day. But if I instead add 24 hours:

select '2023-11-05'::timestamptz + '24 hours'::interval;
       ?column?         
------------------------
2023-11-05 23:00:00-08

24 hours is exactly what I get.

You are generating 15-minute intervals the hard way. You can do it directly and have DST handled for you:

select generate_series('2023-11-05'::timestamptz, '2023-11-06'::timestamptz, '15 minutes'::interval);

    generate_series      
------------------------
2023-11-05 00:00:00-07
2023-11-05 00:15:00-07
2023-11-05 00:30:00-07
2023-11-05 00:45:00-07
2023-11-05 01:00:00-07
2023-11-05 01:15:00-07
2023-11-05 01:30:00-07
2023-11-05 01:45:00-07
2023-11-05 01:00:00-08
2023-11-05 01:15:00-08

...

2023-11-06 00:00:00-08
(101 rows)

Note that 01:00:00-07 is a different point in time than 01:00:00-08. Sticking with timestamp with time zone, aka a fully qualified point in time, removes any ambiguity. Also observe that I get the correct number of "bins". If I run the same thing but for March 12 2023 (spring forward)  I'll get 93 rows but on "normal days" there will be 97.

I suspect your best bet will be to store the data as type time stamp with time zone and to set the time zone before running your queries.

Note that setting the time zone is a client/connection setting so if you set it within a transaction, it will stay set when the transaction concludes. But time manipulation is tricky and trying to DIY reinvent the wheel is painful and often buggy. Let PostgreSQL do the work for you.

Cheers,

Steve

For the sake of simplicity, I only included the possible date-bin variants so that the results can be compared; ordering and grouping with just one date_bin etc. can be easily customized... SET TIMEZONE='Etc/UTC';
SELECT
    sub.gs
    ,date_bin('15 minutes', sub.gs, '2023-01-01') AS norm15minutes
    ,date_bin('1 hours', sub.gs, '2023-01-01') AS norm1hour
    ,date_bin('1 days', sub.gs, '2023-01-01') AS norm1day
FROM (SELECT generate_series('2023-11-03 00:00:00Z'::timestamptz, '2023-11-06 00:00:00Z'::timestamptz, '5 minutes'::interval) AS gs) AS sub;

For the WHERE clause also everything in UTC (the conversion of the parameters of "user time zone" takes place before).

Conversion of the results into the "user time zone" takes place in the client app.

Steve Crawford <scrawford@pinpointresearch.com> writes:
> On Tue, Oct 3, 2023 at 7:33 PM Lincoln Swaine-Moore <lswainemoore@gmail.com>
> wrote:
>> 5) Ideally, the solution would not involve messing with the
>> server/connection's value of timezone. (Though I would be interested if
>> there was a solution that relaxed this constraint and was relatively
>> safe/compatible with transactions and psycopg2.)

> Note that setting the time zone is a client/connection setting so if you
> set it within a transaction, it will stay set when the transaction
> concludes. But time manipulation is tricky and trying to DIY reinvent the
> wheel is painful and often buggy. Let PostgreSQL do the work for you.

Expanding on that philosophy: you should be able to set the timezone
locally within a function, so that it wouldn't be that hard to make a
wrapper for generate_series that emulates the 4-argument version added
in v16.

Rather than messing with manually saving and restoring the prevailing
zone, I'd let the function SET infrastructure do it for me.  Sadly,
that SET clause only takes literal constant arguments, so it'd go
roughly like this:

create function generate_series(timestamptz, timestamptz, interval, text)
returns setof timestamptz
strict immutable language plpgsql as
$$
begin
  perform set_config('timezone', $4, true);
  return query select generate_series($1, $2, $3);
end
$$ set timezone = 'UTC';

Setting the zone to UTC is a useless step, but that triggers
restoring the previous zone when the function exits; simpler
and probably faster than coding the save/restore explicitly.

Side note: whether this is really "immutable" is a matter for
debate, since time zone definitions tend to change over time.
But we chose to mark the new 4-argument version that way,
so you might as well do so too.

            regards, tom lane

Steve Crawford <scrawford@pinpointresearch.com> writes:

On Tue, Oct 3, 2023 at 7:33 PM Lincoln Swaine-Moore <lswainemoore@gmail.com>
wrote:

5) Ideally, the solution would not involve messing with the
server/connection's value of timezone. (Though I would be interested if
there was a solution that relaxed this constraint and was relatively
safe/compatible with transactions and psycopg2.)

Note that setting the time zone is a client/connection setting so if you
set it within a transaction, it will stay set when the transaction
concludes. But time manipulation is tricky and trying to DIY reinvent the
wheel is painful and often buggy. Let PostgreSQL do the work for you.

Expanding on that philosophy: you should be able to set the timezone
locally within a function, so that it wouldn't be that hard to make a
wrapper for generate_series that emulates the 4-argument version added
in v16.

Rather than messing with manually saving and restoring the prevailing
zone, I'd let the function SET infrastructure do it for me.  Sadly,
that SET clause only takes literal constant arguments, so it'd go
roughly like this:

create function generate_series(timestamptz, timestamptz, interval, text)
returns setof timestamptz
strict immutable language plpgsql as
$$
begin  perform set_config('timezone', $4, true);  return query select generate_series($1, $2, $3);
end
$$ set timezone = 'UTC';

Setting the zone to UTC is a useless step, but that triggers
restoring the previous zone when the function exits; simpler
and probably faster than coding the save/restore explicitly.

Side note: whether this is really "immutable" is a matter for
debate, since time zone definitions tend to change over time.
But we chose to mark the new 4-argument version that way,
so you might as well do so too.
			regards, tom lane

Lincoln Swaine-Moore <lswainemoore@gmail.com> writes:
>>> create function generate_series(timestamptz, timestamptz, interval, text)
>>> returns setof timestamptz
>>> strict immutable language plpgsql as
>>> $$
>>> begin
>>> perform set_config('timezone', $4, true);
>>> return query select generate_series($1, $2, $3);
>>> end
>>> $$ set timezone = 'UTC';

> This is a nice suggestion, and in fact, it would be fine from my
> perspective to reset to UTC every time. My concern is only around the
> safety of the final `set timezone`. Under what circumstances/when can I
> count on that being set? E.g. if a query using that function was cancelled
> before finishing, would the connection timezone remain as $4?

No.  The function call mechanism will ensure that timezone goes back
to its previous state at function exit.  (In the case of an error
exit, that's actually handled by the transaction abort logic, but the
result is the same.)  Because of that, I think it doesn't really
matter whether the set_config call says "true" or "false", but saying
that it's a local setting seems less confusing.

> Does that mean I could run this
> outside this context of a function, and expect the setting to go back to
> UTC on a rollback?

An actual rollback would undo the effects of set_config, yes.  You
only need this function wrapper to ensure that subsequent operations
in the same transaction don't see the setting change.

            regards, tom lane

Lincoln Swaine-Moore <lswainemoore@gmail.com> writes:
> Excellent, thank you. So just to be explicit here, I could either run this
> function, or set/run my query/set back, with the same behavior/safety
> guarantees as if I was using the generate_series function with timezone
> from v16?

Yeah.  One thing to keep in mind is that that might have different
behavior in terms of the evaluation of the arguments to the function,
ie which timezone setting is your input parsed according to.

            regards, tom lane

Lincoln Swaine-Moore <lswainemoore@gmail.com> writes:
>> Yeah.  One thing to keep in mind is that that might have different
>> behavior in terms of the evaluation of the arguments to the function,
>> ie which timezone setting is your input parsed according to.

> I see. You mean, in the event that it doesn't conform to an entry in
> `pg_timezone_names`? I do have control over the possible options the user
> can provide, so it should be straightforward to make sure those all have
> entries.

If you mean that your input will always include an explicit zone
specification, then this doesn't affect you.  What I was thinking
about was that

    select generate_series('2023-10-04 13:30', ...)

is going to mean different things depending on the zone setting
that prevails when that constant is parsed.

            regards, tom lane

> On 4 Oct 2023, at 17:58, Lincoln Swaine-Moore <lswainemoore@gmail.com> wrote:
>
> > SELECT
> >     sub.gs AS ts_in_utc
> >     ,sub.gs AT TIME ZONE 'America/New_York' AS ts_in_local_tz
> >     ,date_bin('1 days', sub.gs AT TIME ZONE 'America/New_York',
> > '2023-01-01')
> > FROM (SELECT generate_series('2023-11-03 00:00:00Z'::timestamptz,
> > '2023-11-07 00:00:00Z'::timestamptz, '5 minutes'::interval) AS gs) AS sub
> > WHERE
> >     sub.gs >= '2023-11-05 00:00:00 America/New_York'::timestamptz AND
> > sub.gs < '2023-11-06 00:00:00 America/New_York'::timestamptz
>
> I believe this query will be funky around DST borders, because `sub.gs AT TIME ZONE 'America/New_York'` will be
localizedin a way that erases the difference between hours with different offsets, which are genuinely different. For
instance,I ran this and there are two rows within it that look like:  
>
> ` 2023-11-05 05:00:00+00 | 2023-11-05 01:00:00 | 2023-11-05 00:00:00`
> and
> ` 2023-11-05 06:00:00+00 | 2023-11-05 01:00:00 | 2023-11-05 00:00:00`
>
> I think that the non-unique second column will pose an issue for the date binning at a resolution finer than 1 day.

What I do in such cases is to add an extra column with the UTC timestamp to serve as a linear scale to the local
timestamps.That also helps with ordering buckets in reports and such during DST changes (especially the ones where an
hourrepeats). 

Filtering in the queries occurs on the UTC scale, with the local timestamps calculated back to UTC, so that it doesn’t
matterwhether the local time has 23, 24, 25 or 24.5 or 23.5 or whatever number of hours on a date-range - it all maps
backbecause UTC always has 24 hours. 

Something that I also do is to create calendar tables and views for the buckets, with 2 timestamps per bucket: the
startof the bucket and the start of the next bucket. That gives you a range to put actual timestamps between (not
BETWEENbetween, because that’s inclusive). You can store and index that, as opposed to generated results using
generate_series- basically I materialise those. 

For hours and quarter hours I found it to be fairly convenient to base a view on a join between a date calendar and an
(quarterof an) hour per UTC day table, but materialising that with some indexes may perform better (at the cost of disk
space).I do materialise that currently, but our database server doesn’t have a lot of memory so I’m often not hitting
thecache and performance suffers a bit (infrastructure is about to change for the better though). 

Regards,

Alban Hertroys
--
There is always an exception to always.

> On 4 Oct 2023, at 21:30, Lincoln Swaine-Moore <lswainemoore@gmail.com> wrote:
>
> > What I do in such cases is to add an extra column with the UTC timestamp to serve as a linear scale to the local
timestamps.That also helps with ordering buckets in reports and such during DST changes (especially the ones where an
hourrepeats). 
>
> > For hours and quarter hours I found it to be fairly convenient to base a view on a join between a date calendar and
an(quarter of an) hour per UTC day table, but materialising that with some indexes may perform better (at the cost of
diskspace). I do materialise that currently, but our database server doesn’t have a lot of memory so I’m often not
hittingthe cache and performance suffers a bit (infrastructure is about to change for the better though). 
>
> That's an interesting idea, but I'm not sure I fully understand. Assuming you're aggregating data: what do you group
by?For instance, at an hourly resolution, if you group by both the UTC timestamp and the local one, you might end up,
say,dividing an hour-long bucket in two for time zones with half-hour-based offsets, no?  
>
> Thanks for the detailed writeup! Definitely helpful to learn more about what people are using in production to handle
thissort of thing. 

Frankly, I haven’t had to deal with half-hour-based offsets since I got this idea. I’m using it with
whole-hour-offsets,where it doesn’t affect bin boundaries. 

I suppose you could enrich your data in a similar fashion by adding a (virtual) column with the (client) time zone
offset,so you could group by local timestamp + offset. That’s not going to match index expressions though, I fear… 

For sorting, UTC timestamps would probably still be a useful addition, but they’re simple to add by either converting
backfrom the local timestamps or by taking the min and max of the UTC-based column on the above grouping. Both
solutionsrequire that offset, obviously. 

Now of course there are only 2 hours a year where this happens. Our data scientists chose to ignore the problem for
simplicity’ssake and be slightly off with their numbers on those dates. 

Regards,

Alban Hertroys
--
If you can't see the forest for the trees,
cut the trees and you'll find there is no forest.