This module provides time types that supports positive and negative infinity, as well as some functions for converting to and from strings.

Also, this module also contains commentary regarding postgresql's timestamp types, civil timekeeping in general, and how it relates to postgresql-simple. You can read more about PostgreSQL's date and time types at https://www.postgresql.org/docs/9.5/static/datatype-datetime.html, and the IANA time zone database at https://en.wikipedia.org/wiki/Tz_database.

Stack Overflow also has some excellent commentary on time, if it is a wiki page or a highly upvoted question and answer. If the answer regarding time has not received about a hundred upvotes at least, then the answer is almost invariably completely and painfully wrong, even if it's the chosen answer or the most highly upvoted answer to a question.

PostgreSQL's timestamp with time zone (hereafter, timestamptz) can be converted to Haskell's UTCTime and ZonedTime types, because values of these types represent a self-contained, unambiguous point in time. PostgreSQL's timestamp without time zone (hereafter, timestamp) can be converted to Haskell's LocalTime, because values of these types are ambiguous by themselves, and require context to disambiguate.

While this behavior may be superficially counterintuitive because the names might suggest otherwise, this behavior is correct. In fact, the "timezone" language in both the postgresql and haskell types would be better read as "offset (from UTC)", thus we have postgresql's "timestamp with offset" corresponding to Haskell's "time with the offset 'zero'" and Haskell's "time with an offset (that might be nonzero)". Similarly, postgresql's "timestamp without an offset" corresponds to Haskell's "local time (without an offset)".

It's important to distinguish between an offset, a standard time, and a time zone. An offset is simply a difference of a local time from UTC, such as +00, -05, or +05:30. A standard time specifies an offset (which may vary throughout the year, due to daylight savings) that a region follows, such as Universal Coordinated Time (UTC), Eastern Standard Time/Eastern Daylight Time (EST/EDT), or India Standard Time (IST). And a time zone, much like a standard time, is a function from timestamps to offsets.

A time zone is different from a standard time because different regions inside a standard time can be governed by different civil authorities with different laws and thus have different histories of civil time. An IANA time zone is any region of the world that has had the same history of civil time since 1970-01-01 00:00+00.

For example, as of today, both America/New_York and America/Indiana/Indianapolis are on the EST/EDT time standard, but Indiana used to be on Central Standard Time until 1942, and did not observe daylight savings time (EST only) until 2006. Thus, the choice between these two time zones still matters if you are dealing with timestamps prior to 2006, and could become relevant again if (most of) Indiana moves back to Central Time. (Of course, if the Central to Eastern switch was the only difference, then these two time zones would be the same in IANA's eyes, due to their cutoff date of 1970-01-01.)

Getting back to practicalities, PostgreSQL's timestamptz type does not actually store an offset; rather, it uses the offset provided to calculate UTC, and stores the timestamp as UTC. If an offset is not provided, the given timestamp is assumed to be a local time for whatever the timezone variable is set to, and the IANA TZ database is consulted to calculate an offset from UTC for the time in question.

Note that while most (local timestamp, time zone) pairs correspond to exactly one UTC timestamp, some correspond to two UTC timestamps, while others correspond to none at all. The ambiguous case occurs when the civil time is rolled back, making a calendar day longer than 24 hours. In this case, PostgreSQL silently chooses the second, later possibility. The inconsistent case occurs when the civil time is moved forward, making a calendar day less than 24 hours. In this case, PostgreSQL silently assumes the local time was read off a clock that had not been moved forward at the prescribed time, and moves the clock forward for you. Thus, converting from local time to UTC need not be monotonic, if these inconsistent cases are allowed.

When retrieving a timestamptz, the backend looks at the time zone connection variable and then consults the IANA TZ database to calculate an offset for the timestamp in the given time zone.

Note that while some of the information contained in the IANA TZ database is a bit of a standardized fiction, the conversion from UTC time to a (local time, offset) pair in a particular time zone is always unambiguous, and the result can always be unambiguously converted back to UTC. Thus, postgresql-simple can interpret such a result as a ZonedTime, or use the offset to convert back to UTCTime.

By contrast, the timestamp type ignores any offsets provided to it, and never sends back an offset. Thus, postgresql-simple equates this with LocalTime, which has no concept of an offset. One can convert between timestamptz and timestamp using the AT TIME ZONE operator, whose semantics also demonstrates that timestamptz is UTCTime whereas timestamp is LocalTime.

PostgreSQL's timezone is a per-connection variable that by default is initialized to 'localtime', which normally corresponds to the server's time zone. However, this default can be modified on the server side for an entire cluster, or on a per-user or per-database basis. Moreover, a client can modify their instance of the variable at any time, and can apply that change to the remaining duration of the connection, the current transaction, or the execution context of a server-side function. In addition, upon connection initialization, the libpq client checks for the existence of the PGTZ environment variable, and if it exists, modifies timezone accordingly.

With a few caveats, postgresql-simple is designed so that you can both send and receive timestamps with the server and get a correct result, no matter what the timezone setting is. But it is important to understand the caveats:

1. The correctness of server-side computations can depend on the timezone setting. Examples include adding an interval to a timestamptz, or type casting between timestamp and timestamptz, or applying the DATE function to a timestamptz.
2. The (localtime, offset) pair contained in a ZonedTime result will depend on the timezone setting, although the result will always represent the same instant in time regardless of the time zone.
3. Sending a LocalTime and interpreting it as a timestamptz can be useful, as it will be converted to UTC via the tz database, but correctness will depend on the timezone setting. You may prefer to use an explicit AT TIME ZONE conversion instead, which would avoid this contextual dependence.

Furthermore, although these following points don't involve the timezone setting, they are related to the last point above:

1. Sending a UTCTime and interpreting it as a timestamp can be useful. In practice, the most common context used to disambiguate timestamp is that it represents UTC, and this coding technique will work as expected in this situation.
2. Sending a ZonedTime and interpreting it as a timestamp is almost always the wrong thing to do, as the offset will be ignored and discarded. This is likely to lead to inconsistencies in the database, and may lead to partial data loss.

When dealing with local timestamps that refer to the future, it is often useful to store it as a local time in a timestamp column and store the time zone in a second column. One reason to do this is so that you can convert to UTC on the fly as needed, and be protected against future changes to the TZ database due to changes in local time standards. In any case, ZonedTime is not suitable for this application, because despite its name, it represents an offset and not a time zone. Time zones can change; offsets do not. In reality, we can't convert a local timestamp that occurs sufficiently far in the future to UTC, because we don't know how to do it yet.

There are a few limitations and caveats that one might need to be aware of with the current implementation when dealing with older timestamps:

For sufficiently old timestamps in almost all time zones, the IANA TZ database specifies offsets from UTC that is not an integral number of minutes. This corresponds to local mean time; that is, astronomical time in the city that defines the time zone. Different time zones moved away from local mean time to a standard time at different points in history, so "sufficiently old" depends on the time zone in question.

Thus, when retrieving a timestamptz postgresql will in some cases provide seconds in the offset. For example:

$ psql
psql (9.4.5)
Type "help" for help.

lpsmith=> SET timezone TO 'America/New_York';
SET
lpsmith=> VALUES ('1883-11-18 16:59:59+00'::timestamptz),
                 ('1883-11-18 17:00:00+00'::timestamptz);
           column1
------------------------------
 1883-11-18 12:03:57-04:56:02
 1883-11-18 12:00:00-05
(2 rows)

Both of these timestamps can be parsed as a UTCTime type, however ZonedTime will fail on the former timestamp. Because ZonedTime assumes that offsets are an integer number of minutes, there isn't an particularly good solution here.

PostgreSQL, like most software, uses the proleptic Gregorian calendar for its date calculations, extending the Gregorian calendar backwards in time before its introduction and pretending that the Julian calendar does not exist. For most purposes, the adoption of the Gregorian calendar ranges from 1582-10-15 to 1923-03-01, depending on location and sometimes even political allegiances within a single location.

Timestamps BCE are not supported. For example, PostgreSQL will emit "0045-01-01 BC" for the first proleptic Gregorian day of the year the Roman Empire adopted the Julian Calendar, but postgresql-simple does not (yet?) have the ability to either parse or generate this syntax. Unfortunately this syntax isn't convenient to print or especially parse.

Also, postgresql itself cannot parse or print dates before 4714-11-24 BC, which is the Julian date on the proleptic Gregorian Calendar. Although postgresql's timestamp types are perfectly capable of representing timestamps nearly 300,000 years in the past, using this would require postgresql-simple and other client programs to support binary parameters and results.

Dealing with years BCE is also complicated slightly by the fact that Haskell's time library has a year "0000", which is a convention often used by astronomers, while postgresql adopts the more historically accurate convention that there is no year zero, but rather "1 BCE" was immediately followed by "1 CE".