Date and time operators [YSQL]

Date and time operators

Each of the comparison operators, <, <=, =, >=, >, and <>, each of the arithmetic operators, +, -, *, and /, and, of course, the :: typecast operator has one or several overloads whose two operands are among the date, time, plain timestamp, timestamptz, and interval data types. The parent section explains why timetz is not covered in this overall date-time section. This is why there are five interesting date-time data types.

The section Typecasting between values of different date-time data types shows which of the twenty potential typecasts between pairs of different interesting date-time data types are legal. (It's meaningless to think about typecasting between a pair of values of the same data type.) And for each of those ten that is legal, the section describes its semantics. Further, the typecast operator, ::, has an overload from/to each of the five listed interesting date-time data types to/from text. In other words, there are ten typecasts from/to the interesting date-time data types to/from text. The section Typecasting between date-time values and text values describes these.

The following tables show, for the comparison operators jointly, for each of the addition and subtraction operators separately, and for the multiplication and division operators jointly, which of the twenty-five nominally definable operand pairs are legal. Links are given to the sections that describe the semantics.

Always write the typecast explicitly.

Some of the expressions that use the binary operators that this section covers are legal where you might expect them not to be. The reason is that, as the section Typecasting between values of different date-time data types shows, typecasts, and corresponding implicit conversions, are defined between the data types of the operand pairs that you might not expect. Yugabyte recommends that you consider very carefully what your intention is when you take advantage of such conversions between values of different data type.

There are two reasons for writing the typecast explicitly:

  • You advertise to the reader that typecasting is being done and that they need to understand the typecast semantics.

  • You specify explicitly whether the typecast is to be done on the left operand's value to the right operand's data type, or vice versa, rather than having you, and other readers of your code, rely on remembering what the default behavior is.

Overloads of the comparison operators

All of the comparison operators, <, <=, =, >=, >, and <> are legal for pairs of values of the same date-time data type. The semantics of comparing two moment values is straightforward because a moment value is a scalar. The section How does YSQL represent an interval value? explains that an interval value is actually a three-component [mm, dd, ss] tuple. The semantics of interval-interval comparison, therefore, needs careful definition—and the section Comparing two interval values does this. Yugabyte recommends that you avoid the complexity that the non-scalar nature of interval values brings by adopting the approach that the section Custom domain types for specializing the native interval functionality describes. (It shows you how two define three interval flavors, pure months, pure days, and pure hours, so that their values are effectively scalars.)

When the data types differ, the comparison is sometimes simply illegal. The attempt then causes this error:

42883: operator does not exist...

Otherwise, even though the data types differ, the comparison is legal. The section Test the date-time comparison overloads presents code that tests all of the comparison overloads whose syntax you can write. This table summarizes the outcomes. An empty cell means that the overload is illegal.

left operand\right operand DATE TIME PLAIN TIMESTAMP TIMESTAMPTZ INTERVAL
DATE ok ok ok
TIME ok ok
PLAIN TIMESTAMP ok Note ok ok
TIMESTAMPTZ ok Note ok ok
INTERVAL ok ok

If a comparison is legal between values of two different data types, then (with the caveat that the immediately following note states) it's legal both when values of the two data types are used, respectively, as the left and right operands and when they're used as the right and left operands. In all of these cases, the mutual typecast between values of the pair of data types is legal in each direction.

Note: In just two cases, the comparison is illegal between values of a pair of data types where the typecast is legal. The table calls out these cases. In both these cases, the typecast operator is legal from the left operand to the right operand, but not vice versa.

If you think that it makes sense, then you can execute the comparison by writing the explicit typecast. Try this:

set timezone = 'UTC';
with c as (
  select
    '13:00:00'                ::time        as t,
    '02-01-2020 12:30:00'     ::timestamp   as ts,
    '02-01-2020 13:30:00 UTC' ::timestamptz as tstz
  )
select
  (ts  ::time > t)::text as "ts > t",
  (tstz::time > t)::text as "tstz > t"
from c;

It runs without error and produces these two values:

 ts > t | tstz > t
--------+----------
 false  | true

The subsections plain timestamp to time and timestamptz to time, in the section Typecasting between values of different date-time data types, explain the semantics of these two typecasts. The outcomes here, false and true, are consistent with those explanations.

In summary, comparison makes obvious sense only when the data types of the left and right operands are the same. These comparisons correspond to the on-diagonal cells. If you have worked out that it makes sense to use any of the comparison operations that the table above shows with ok in an off-diagonal cell, then, for clarity, you should write the explicit typecast that you intend.

Overloads of the addition and subtraction operators

The model for the intended, and therefore useful, functionality is clear and simple:

  • Subtraction between a pair of date values produces an integer value. Correspondingly, adding or subtracting an integer value to/from a date value produces a date value. (Early in PostgreSQL's history, date was the only date-time data type. Because interval was not yet available, it was natural that the difference between two date values would be an integer value.)
  • Subtraction between a pair of values of the newer data types time, plain timestamp, or timestamptz, produces an interval value. Correspondingly, adding or subtracting an interval value to/from a time, plain timestamp, or timestamptz value produces, respectively, a time, plain timestamp, or timestamptz value.
  • Adding two interval values or subtracting one interval value from another produces an interval value.

The data types that these useful operations produce are shown in bold in the cells in the tables in the sections Overloads of the addition operator and Overloads of the subtraction operator. The resulting data types in any other non-empty cells in these tables are shown in regular font—and Yugabyte recommends that you avoid using the operations that they denote in application code. (If you are sure that an operation denoted by a regular font cell makes sense in your present use case, then you should write the implied typecast explicitly.)

Overloads of the addition operator

The section Test the date-time addition overloads presents code that tests all of the addition operator overloads whose syntax you can write. This table summarizes the outcomes. An empty cell means that the overload is illegal.

left operand\right operand DATE TIME PLAIN TIMESTAMP TIMESTAMPTZ INTERVAL
DATE plain timestamp plain timestamp
TIME plain timestamp plain timestamp timestamptz time
PLAIN TIMESTAMP plain timestamp plain timestamp
TIMESTAMPTZ timestamptz timestamptz
INTERVAL plain timestamp time plain timestamp timestamptz interval

Notice that the table is symmetrical about the top-left to bottom-right diagonal. This is to be expected because addition is commutative.

When the resulting data type is rendered in bold font, this indicates that the operation makes intrinsic sense. In three of these cases, the operation has a moment value as one of the arguments and an interval value as the other. (The Yugabyte documentation refers to values of the date, time, plain timestamp or timestamptz data types as moments.) In the fourth case, both arguments are interval values.

The date-time data types are unusual with respect to addition, at least with respect to the conceptual proposition, in that you can't add two values of the same moment data type. The on-diagonal cell for each of these four data types is empty. (As mentioned, you can add a pair of interval values.)

The outcomes for the "date_value + interval_value" cell, and the converse "interval_value + date_value" cell, might surprise you. This is explained by the fact that, uniquely for subtraction between moment values, subtracting one date value from another produces an integer value. Try this:

select
  pg_typeof('2020-01-06'::date - '2020-01-01'::date)  as "data type",
            '2020-01-06'::date - '2020-01-01'::date   as "result";

This is the result:

 data type | result
-----------+--------
 integer   |      5

The inverse of this, then, is that adding an integer value to a date value produces a date value. Try this:

select
  pg_typeof('2020-01-01'::date + 5::integer)  as "data type",
            '2020-01-01'::date + 5::integer   as "result";

This is the result:

 data type |   result
-----------+------------
 date      | 2020-01-06

As mentioned above, this departure, by date, from the pattern that the other moment data types follow reflects PostgreSQL's history. Adding a sixth column and a sixth row for integer to the table would clutter it unnecessarily because only the date-integer and integer-date cells in the new column and the new row would be non-empty.

The other outcomes are intuitively clear. What could it mean to add three o'clock to five o'clock? However, the clear conceptual proposition is compromised because, as the sections time to interval and interval to time (on the Typecasting between values of different date-time data types page) show, time values can be implicitly converted to interval values, and vice-versa.

This explains the non-empty cells in the time row and the time column that are rendered in normal font.

Try this:

do $body$
declare
  t0  constant time     not null := '12:00:00';
  i0  constant interval not null := '12:00:00';

  t   constant time     not null := i0;
  i   constant interval not null := t0;
begin
  assert t = t0, 'interval > time failed.';
  assert i = i0, 'time > interval failed.';
end;
$body$;

The block finishes silently, showing that both assertions hold. Notice the practice recommendation above. If you work out that you can make use of the semantics of these conversions, you should write the typecasts explicitly rather than rely on implicit conversion.

Note: you might argue that you can give a meaning to the strange notion of adding a pair of timestamptz values like this:

drop function if exists sum(timestamptz, timestamptz) cascade;

create function sum(t1 in timestamptz, t2 in timestamptz)
  returns timestamptz
  stable
  language plpgsql
as $body$
declare
  e1      constant double precision not null := extract(epoch from t1);
  e2      constant double precision not null := extract(epoch from t2);
  result  constant timestamptz      not null := to_timestamp(e1 + e2);
begin
  return result;
end;
$body$;

Test it like this:

set timezone = 'UTC';
select sum('1970-01-01 01:00:00 UTC', '1970-01-01 02:00:00 UTC')::text;

This is the result:

 1970-01-01 03:00:00+00

It's exactly what the semantics of extract(epoch from timestamptz_value) and to_timestamp(double_precision_value) tell you to expect. You could even create a user-defined operator + based on the function sum(timestamptz_value, timestamptz_value) as defined here. But it's hard to see how the semantics brought by this might be generally useful.

Overloads of the subtraction operator

The section Test the date-time subtraction overloads presents code that tests all of the subtraction addition operator overloads whose syntax you can write. This table summarizes the outcomes. An empty cell means that the overload is illegal.

left operand\right operand DATE TIME PLAIN TIMESTAMP TIMESTAMPTZ INTERVAL
DATE plain timestamp interval interval plain timestamp
TIME interval time
PLAIN TIMESTAMP interval plain timestamp interval interval plain timestamp
TIMESTAMPTZ interval timestamptz interval interval timestamptz
INTERVAL interval interval

Notice that the table is symmetrical about the top-left to bottom-right diagonal. This is to be expected because subtraction is commutative in this sense:

(a - b) = -(b - a)

When the resulting data type is rendered in bold font, this indicates that the operation makes intrinsic sense. You should avoid the operations that the non-empty cells in regular font denote unless you are convinced that a particular one of these makes sense in your present use case. Then, as recommended above, you should write the implied typecast operator explicitly.

Notice that the number of non-empty cells (seven in all) in bold font in this table for the overloads of the subtraction operator is the same as the corresponding number in the table for the overloads of the addition operator. This reflects the complementary relationship between addition and subtraction.

However, there are more non-empty cells in regular font (eleven in all) in this table for the overloads of the subtraction operator than there are in the table for the overloads of the addition operator (eight in all). You might think that this is odd. This outcome reflects the complex rules for when implicit typecasting may be invoked.

Overloads of the multiplication and division operators

The multiplication and division operators are illegal between all possible pairs of values of the date-time data types. This is consistent with the common-sense expectation: what could it mean, for example, to multiply or divide one timestamp value by another? For completeness, the sections Test the date-time multiplication overloads and Test the date-time division overloads present code that confirms that this is the case—in other words, that there are no implicit data type conversions, here, to confuse the simple statement of rule.

Multiplication and division are legal only when you multiply or divide an interval value by a real or integral number. (Again, what could it mean, for example, to multiply or divide a timestamp value by a number?) Moreover, division needs no specific discussion because dividing the interval value i by the number n is the same as multiplying i by 1/n. The section Multiplying or dividing an interval value by a number explains the semantics. This needs careful definition because of the non-scalar nature of interval values.

Yugabyte recommends that you avoid the complexity that the non-scalar nature of interval values brings by adopting the approach that the section Custom domain types for specializing the native interval functionality describes.