Range data types

YSQL supports six built-in range data types. Each defines a range of values of a specific underlying data type. (The PostgreSQL documentation uses the term "subtype" for "underlying data type".) It's also possible to create user-defined range data types. A later version of the YSQL documentation will explain this.

Synopsis

Range data type Underlying data type
int4range integer (a.k.a. int)
int8range bigint
numrange numeric
tsrange timestamp without time zone (a.k.a. timestamp)
tstzrange timestamp with time zone (a.k.a. timestamptz)
daterange date

Description

The underlying data type of a range data type must be orderable. Each of the six built-in range data types meets this requirement. A range value is defined by its (smaller) start value and the (larger) end value. A range value therefore corresponds to the "interval" notion in mathematics. (Don't confuse this with the YSQL data type interval which denotes an amount of elapsed time and which is defined only by its absolute size.)

The "interval" notion in mathematics

See the Wikipedia article Interval (mathematics).

The following notions, and their notation, from mathematics carry over to the YSQL range data types.

  • An open interval excludes its endpoints, and is denoted with parentheses. For example, (0,1) means greater than 0 and less than 1.
  • A closed interval includes all its endpoints, and is denoted with square brackets. For example, [0,1] means greater than or equal to 0 and less than or equal to 1.
  • A half-open interval includes only one of its endpoints, and is denoted by mixing the notations for open and closed intervals. For example, (0,1] means greater than 0 and less than or equal to 1; and [0,1) means greater than or equal to 0 and less than 1.

Range values in YSQL

A range value can be specified either as a literal or using a constructor function.

First, create a table:

create table t(k int primary key, r1 tsrange, r2 tsrange);

Specify range values using literals

The same approach is used to specify range values of all range data types using literals. A text value is defined as follows:

  • It starts with an opening parenthesis or an opening square bracket.

  • It is followed by the ::text typecast representation of the lower bound, a comma and then the ::text typecast representation of the upper bound.

  • It finishes with a closing parenthesis or a closing square bracket.

This example uses tsrange.

insert into t(k, r1, r2)
values (
  1,
  '[2010-01-01 14:30, 2010-01-01 15:30)',
  '(2010-01-01 15:00, 2010-01-01 16:00]');

Inspect the bounds:

select lower(r1), upper(r1) from t where k = 1;

The reported values are insensitive to whether the bound is inclusive or exclusive. This is the result:

        lower        |        upper
---------------------+---------------------
 2010-01-01 14:30:00 | 2010-01-01 15:30:00

Are the bounds inclusive (defined using a square bracket) or not inclusive (defined using a parenthesis):

select
  (lower_inc(r1))::text as "is r1's lower inclusive?",
  (upper_inc(r1))::text as "is r1's upper inclusive?",
  (lower_inc(r2))::text as "is r2's lower inclusive?",
  (upper_inc(r2))::text as "is r2's upper inclusive?"
from t where k = 1;

This is the result:

 is r1's lower inclusive? | is r1's upper inclusive? | is r2's lower inclusive? | is r2's upper inclusive?
--------------------------+--------------------------+--------------------------+--------------------------
 true                     | false                    | false                    | true

Specify "tsrange" values using the constructor function

The constructor function for a particular range data type has the same name as the range data type. The first and second formal parameters are the lower and upper bounds. And the third, optional, formal parameter specifies the inclusive status of each bound as one of these text values:

'()'   '(]'   '[)'   '[]'

When the constructor is invoked without supplying a value for the third formal parameter, the value '[)' is used.

This INSERT statement establishes the same values for "r1" and "r2" in the row with "k = 2" that the row with "k = 1" has:

insert into t(k, r1, r2)
values (
  2,
  tsrange(
    make_timestamp(2010, 1, 1, 14, 30, 0.0),
    make_timestamp(2010, 1, 1, 15, 30, 0.0)
    ),
  tsrange(
    make_timestamp(2010, 1, 1, 15, 0, 0.0),
    make_timestamp(2010, 1, 1, 16, 0, 0.0),
    '(]'));

This SELECT statement shows that the values in the row with "k = 2" are indeed the same as the row with "k = 1" has:

select (
    (select r1 from t where k = 1) = (select r1 from t where k = 2)

    and

    (select r2 from t where k = 1) = (select r2 from t where k = 2)

  )::text "result is as expected?";

This is the result:

 result is as expected?
------------------------
 true

Unbounded ranges

Either, or both of, the lower bound and the upper bound of a range value can be set to express the semantics "unbounded". The PostgreSQL documentation uses "unbounded" and "infinite" interchangeably to denote such a range. Yugabyte recommends always using the term "unbounded" and avoiding the term "infinite". The reason for this is explained later in this section.

  • When a range value is defined using a literal, an unbounded lower or upper bound is specified as unbounded simply by omitting the value, like this:
select ('(, 5]'::int4range)::text as "canonicalized literal";

This is the result:

 canonicalized literal
-----------------------
 (,6)

Notice that the opening punctuation mark has been canonicalized to ( rather than to [ — as would be the form for a finite bound. The bound value is still omitted because "unbounded" means the same when the bound is exclusive as when it's inclusive. The closing punctuation mark has been canonicalized to ) in the normal way, bringing the consequence that the bound is rendered as "6" instead of as "5". (See the section Discrete range data types and the canonical forms of their literals below.)

Notice that no discretion is allowed for the use of whitespace: to denote that the lower bound is unbounded, there must be no whitespace between the opening punctuation mark and the comma; and to denote that the upper bound is unbounded, there must be no whitespace between the comma and the closing punctuation mark.

  • When a range value is defined using a constructor, an unbounded lower or upper bound is specified as unbounded by using NULL, like this:
select (
    '(, 5]'::int4range = int4range(null, 5, '(]')
  )::text as "literal and constructor produce the same result?";

This is the result:

 literal and constructor produce the same result?
--------------------------------------------------
 true

You can test the "unbounded" status of a bound with the boolean functions lower_inf() and upper_inf()like this:

with a as (select '(,)'::int4range as v)
select lower_inf(v)::text, upper_inf(v)::text from a;

This is the result:

 lower_inf | upper_inf
-----------+-----------
 true      | true

The example illustrates the point that, as a special case, a range might be doubly unbounded. This could be useful in a scenario where user-input determines if a query is to be restricted using a predicate on some column or if there is to be no such restriction. Dynamic SQL can be avoided by writing the query as fixed static text using the "contains" operator (see the section Is a value of a range's underlying data type contained within a range? below) and by binding in the value of the range at run time.

Note: Some data types, like for example timestamp, support a special infinity value which can also be used to define a range. Try this:

select
  ('infinity'::timestamp)             ::text as "infinity timestamp",
  ('[2020-01-01, infinity]'::tsrange) ::text as "infinity upper bound tsrange";

This is the result:

 infinity timestamp |   infinity upper bound tsrange
--------------------+----------------------------------
 infinity           | ["2020-01-01 00:00:00",infinity]

However, a range with an upper bound equal to positive infinity turns out to be different from a range that is unbounded at that end. The same holds for a range that has a lower bound equal to negative infinity. These two tests confirm this. First try this:

select (upper_inf('[2020-01-01, infinity]'::tsrange))::text as "upper bound set to infinity tests as infinity?";

This is the result:

 upper bound set to infinity tests as infinity?
------------------------------------------------
 false

And now try this:

select (
    tsrange('2020-01-01'::timestamp, 'infinity'::timestamp)
    =
    tsrange('2020-01-01'::timestamp, null      ::timestamp)
  )::text as "infinity bound same as unbounded bound?";

This is the result:

 infinity bound same as unbounded bound?
-----------------------------------------
 false

Yugabyte recommends always specifying an unbounded bound by omitting the value (in a literal) or by using NULL (in a constructor).

Using the special value infinity brings the following disadvantages:

  • Not all of the underlying data types for which there are built-in range data types support an infinity notion. In fact, only date, plain timestamp and timestamptz (in the class for which there are built-in range data types) do support an infinity notion.
  • The semantics are unclear. This is shown most dramatically by the fact that the lower_inf() and upper_inf() functions return FALSE when the corresponding bound is set to (negative or positive) infinity. Correspondingly, a pair of range values where one bound is set ordinarily to a normal value and where the other bound is set either to infinity or to "unbounded" compare as unequal.
  • There are yet other reasons. But the two that have been explained in this "recommendation" section are sufficient reason to avoid the special negative or positive infinity value.

The earlier recommendation to avoid the term "infinite" when describing a bound and to use only "unbounded" should now be clear: all inflexions of "infinity" serve only to bring confusion.

Operations on range values and values of the underlying data type

You can find out if a single range value is empty or if two range values intersect; you can derive a new range value as the intersection of two range values; and you can find out if a value of the underlying data type is contained within a range value.

Is a range value empty?

Try this:

select isempty(r1)::text from t where k = 1;

The return data type of isempty() is boolean. This is the result:

 isempty
---------
 false

Now try this:

select isempty(numrange(1.5, 1.5, '[)'))::text;

The third actual argument specifies the default for the corresponding formal parameter as a self-documentation device. This is the result:

 isempty
---------
 true

Here is an alternative formulation of the same test:

select (numrange(1.5, 1.5, '[)') = 'empty'::numrange)::text as "is empty?";

This (of course) is the result:

 is empty?
----------
 true

In other words, the literal 'empty'::<some range data type> represents the special "empty range" value for that range data type.

If the third actual is changed to '[]' to denote that both ends of the range are inclusive, then the result becomes false. Notice that a range's upper bound must be greater than, or equal to, the lower bound. Try this:

select isempty(numrange(1.5, 1.0))::text;

The attempt fails with the 22000 error. (This maps to the data_exception exception.)

Do two range values intersect?

Try this:

select (numrange(1.0, 2.0) && numrange(1.5, 2.5))::text as "range values intersect?";

The return data type of the && intersection operator is boolean. This is the result:

 range values intersect?
-------------------------
 true

It you change the lower and upper bounds of the second range value to 3.0 and 4.0, then the result becomes false, of course.

Produce a new range value as the intersection of two range values

Try this:

select (numrange(1.0, 2.0) * numrange(1.5, 2.5))::text as "the intersection";

This is the result:

[1.5,2.0)

If the two operands of the * range intersection operator don't intersect, then the result is the empty range. Try this:

do $body$
declare
  e      constant numrange not null := 'empty';
  r1     constant numrange not null := numrange(1.0, 2.0);
  r2     constant numrange not null := numrange(3.0, 4.0);
  result constant numrange not null := r1 * r2;
begin
  assert result = e, 'assert failed';
end;
$body$;

The block completes silently showing that the assertion holds.

Is a value of a range's underlying data type contained within a range?

Try this:

select (17::int <@ int4range(11, 42))::text as "is value within range?";

This is the result:

 is value within range?
------------------------
 true

The query can be formulated the other way round, thus:

select (int4range(11, 42) @> 17::int)::text as "does range contain value?";

This is the result:

 does range contain value?
---------------------------
 true

Discrete range data types and the canonical forms of their literals

The underlying data types of the int4range, int8range, and daterange data types are discrete: int4 and int8 accommodate only exact integer steps; and date accommodates only steps of exact whole days. This means that the definitions of an inclusive bound and an exclusive bound lead to the possibility of defining identical ranges in different ways. Consider the int range that includes 3 as its lower bound and that includes 7 as its upper bound. It can be written in four different ways:

create table t(
  k int primary key,
  r1 int4range not null,
  r2 int4range not null,
  r3 int4range not null,
  r4 int4range not null);

insert into t(k, r1, r2, r3, r4) values(1, '[3, 8)', '[3, 7]', '(2, 8)', '(2, 7]');

select (
  (select r1 = r2 from t where k = 1) and
  (select r1 = r3 from t where k = 1) and
  (select r1 = r4 from t where k = 1)
  )::text as "all the same?";

This is the result:

 all the same?
---------------
 true

Now do this:

select r1::text, r2::text, r3::text, r4::text from t where k = 1;

This is the result:

  r1   |  r2   |  r3   |  r4
-------+-------+-------+-------
 [3,8) | [3,8) | [3,8) | [3,8)

Of course, because as has been seen, the values of "r1", "r2", "r3", and "r4" are identical, it's of no consequence how these were established. The canonical form of the literal is the form that is used to display the value. The test results show that this is the form where the lower bound is inclusive (denoted by the square bracket) and where the upper bound is exclusive (denoted by the parenthesis).

Current restriction

See GitHub Issue #7353. It tracks the fact that you cannot create an index on a column list that includes a column with a range data type. Correspondingly, you cannot define a primary key constraint on such a column list.