The literal for a "row" type value

The word "row" has two different uses; but these uses are really different sides of the same coin. A row in a schema-level table is actually an occurrence of a "row" type—in other words, a "row" type value. In this case, the schema-level "row" type is created automatically as a side effect of executing the CREATE TABLE statement. It has the same name as the table. (This is allowed because tables and types are in different namespaces.) Further, a column in a schema-level table can have a user-defined "row" type as its data type, and in this case the "row" type need not be partnered with a table.

You might see the term "record" when you use the \df metacommand to show the signature of a function. Briefly, it's an anonymous "row" type. You produce a record instance when you use a literal that has the correct form of a "row" type but when you omit the typecast operator. If you adhere to recommended practice, and always explicitly typecast such literals, then you needn't try to understand what a record is.

You can read more about these notions in the PostgreSQL documentation here:

You need first to understand how to write a literal for a "row" type value before you can understand, as The literal for an array of "row" type values explains, how to write the literal for an array of such values.

This section uses the same approach as The literal for an array of primitive values: first it states the rules; and then it illustrates these with examples.

Statement of the rules

Just as in Statement of the rules in the "The literal for an array of primitive values" section, the statement of these rules depends on understanding the notion of the canonical form of a literal.

If you follow the rules that are stated here, and illustrated in the demonstrations below, then you will always produce a syntactically valid literal which expresses the semantics that you intend. It turns out that many other variants, especially for text[] values, are legal and can produce the result that you intend. However, the rules that govern these exotic uses will not be documented because it is always sufficient to create your literals in canonical form.

Here is the sufficient set of rules.

  • The commas that delimit successive values, and opening and closing parentheses, must not be surrounded by whitespace.
  • Do not surround the individual representations of numeric and boolean primitive values with double quotes.
  • Do surround the individual representations of timestamp values with double quotes, even though this is not strictly necessary.
  • Do surround every individual representation of a text value with double quotes, even though this is not always necessary. It is necessary for any value that itself contains, as ordinary text, any whitespace or any of the characters that have syntactic significance within the outermost curly brace pair. This is the list:
   <space>   (   )   ,   "   \
  • It's sufficient then to write all special characters ordinarily within the enclosing double quotes except for each of the double quote character itself and the backslash character. These must be escaped. The double quote character is escaped by doubling it up. And the backslash character is escaped with an immediately preceding single backslash.

  • To specify that the value for a field is NULL , you must leave no whitespace between the pair of delimiters (Left parenthesis, comma, or right parenthesis) that surround its position. (This is the only choice.)

Always write array literals in canonical form

Exactly the same considerations apply here as were explained in Always write array literals in canonical form in the section that explained the rules for literals for an array of primitive values.

Examples to illustrate the rules

It will be sufficient to consider "row" types with fields of just these data types:

  • numeric data types (like int and numeric)
  • stringy data types (like text, varchar, and char)
  • date-time data types (like timestamp)
  • the boolean data type.

Use the "row" type constructor to create representative values of each kind and inspect its ::text typecast.

"Row" type with int fields

This example demonstrates the principle:

create type row_t as (f1 int, f2 int, f3 int);
create table t(k serial primary key, v1 row_t, v2 row_t);
insert into t(v1) values (row(1, 2, 3)::row_t);
select v1::text as text_typecast from t where k = 1
\gset result_
\echo :result_text_typecast

The keyword ROW names the "row" type constructor function. It is optional, but is used here for emphasis.

The \gset metacommand was used first in this "Array data types and functionality" major section in array_agg() and unnest().

Notice that, in this example, the SELECT statement is terminated by the \gset metacommand on the next line rather than by the usual semicolon. The \gset metacommand is silent. The \echo metacommand shows this:

(1,2,3)

In this case, the value of the ::text typecast has the identical form to that of the "row" type constructor. But, as is seen below, this is not generally the case.

You can see the general form already:

  • The (text of) a "row" type literal starts with the left parenthesis and ends with the right parenthesis.

  • The items within the parentheses are delimited by commas, and there is no space between one item, the comma, and the next item. Nor is there any space between the left parenthesis and the first item or between the last item and the right parenthesis.

The next section, "Row" type with text fields, shows that more needs to be said. But the two rules that you have already noticed always hold.

To use the text of the literal that was produced to create a value, you must enquote it and typecast it. Do this with the \set metacommand:

\set canonical_literal '\'':result_text_typecast'\''::row_t
\echo :canonical_literal

. The \echo metacommand now shows this:

'(1,2,3)'::row_t

Next, use the canonical literal that you produced to update "t.v2" to confirm that the value that the row constructor created was recreated:

update t set v2 = :canonical_literal where k = 1;
select (v1 = v2)::text as "v1 = v2" from t where k = 1;

It shows this:

 v1 = v2
---------
 true

As promised, the canonical form of the "row" type literal does indeed recreate the identical value that the "row" type constructor created.

"Row" type with text fields

Use "Row" type with int fields as a template for this and the subsequent sections. The example sets array values each of which, apart from the single character a, needs some discussion. These are the characters (or, in one case, character sequence), listed here "bare" and with ten spaces between each:

     a       '       a b       ()       ,       "       \       null
create type row_t as (f1 text, f2 text, f3 text, f4 text, f5 text, f6 text, f7 text, f8 text);
create table t(k serial primary key, v1 row_t, v2 row_t);
insert into t(v1) values (
  ('a', $$'$$, 'a b', '()', ',', '"', '\', null)::row_t);

select v1::text as text_typecast from t where k = 1
\gset result_
\echo :result_text_typecast

Here, the ROW keyword in the "row" type constructor function is omitted to emphasize its optional status.

The \echo metacommand shows this:

(a,',"a b","()",",","""","\\",)

This is rather hard (for the human) to parse. To make the rules easier to see, the syntactically significant commas are surrounded with three spaces on each side:

(   a   ,   '   ,   "a b"   ,   "()"   ,   ","   ,   """"   ,   "\\"   ,   )

Note: The introduction of spaces here, to help readability, is done only for that purpose. Unlike is the case for an array literal, doing this actually affects the value that the literal produces. You will demonstrate this at the end of this section.

In addition to the first two rules, you notice the following.

  • Double quotes are used to surround a value that includes any spaces. (Though the example doesn't show it, this applies to leading and trailing spaces too.)
  • The comma has been surrounded by double quotes. This is because it does have syntactic significance, as the value delimiter, within the parentheses of a "row" type literal.
  • The parentheses have been surrounded by double quotes. This is because these do have syntactic significance.
  • The single quote is not surrounded with double quotes. Though it has syntactic significance in other parsing contexts, it is insignificant within the parentheses of a "row" type literal. This holds, also, for all sorts of other punctuation characters like ; and : and [ and ] and so on.
  • The double quote has been escaped by doubling it up and this has been then surrounded with double quotes. This is because it does have syntactic significance, as the (one and only) quoting mechanism, within the parentheses of a "row" type literal.
  • The backslash has also been escaped with another single backslash and this has been then surrounded with double quotes. This is because it does have syntactic significance, as the escape character, within the parentheses of a "row" type literal.
  • NULL is represented in a "row" type literal by the absence of any characters between two successive delimiters: between the left parenthesis and the first comma, between two successive commas, or between the last comma and the right parenthesis.

There's another rule that the present example does not show. Though not every comma-separated value was surrounded by double quotes, it's never harmful to do this. You can confirm this with your own test, Yugabyte recommends that, for consistency, you always surround every text value within the parentheses of a "row" type literal with double quotes.

To use the text of the literal that was produced to create a value, you must enquote it and typecast it. Do this, as you did for the int example above, with the \set metacommand. But you must use dollar quotes because the literal itself has an interior single quote.

\set canonical_literal '$$':result_text_typecast'$$'::row_t
\echo :canonical_literal

The \echo metacommand now shows this:

$$(a,',"a b","()",",","""","\\",)$$::row_t

Next, use the canonical literal that you produced to update "t.v2" to confirm that the value that the row constructor created was recreated:

update t set v2 = :canonical_literal where k = 1;
select (v1 = v2)::text as "v1 = v2" from t where k = 1;

It shows this:

 v1 = v2
---------
 true

So, again as promised, the canonical form of the array literal does indeed recreate the identical value that the "row" type constructor created.

Finally in this section, consider the meaning-changing effect of surrounding the comma delimiters with whitespace. Try this:

create type row_t as (f1 text, f2 text, f3 text);
select '(   a   ,   "(a b)"   ,   c   )'::row_t;;

It shows this:

 ("   a   ","   (a b)   ","   c   ")

You understand this by realizing that the entire run of characters between a pair of delimiters is taken as the value. And double quotes act as an interior escaping mechanism. This model holds when, but only when, the value between a pair of delimiters is interpreted as a text value (because this is the data type of the declared "row" type field at this position).

This rule is different from the rule for an array literal. It's also different from the rules for JSON documents. In these cases, the value is entirely within the double quotes, and whitespace around punctuation characters outside of the double-quoted values is insignificant.

Note: There is absolutely no need to take advantage of this understanding. Yugabyte recommends that you always use the "almost-canonical" form of the literal—in other words, you surround every single text value with double quotes, even when these are not needed, and you allow no whitespace between these double-quoted values and the delimiter at the start an end of each such value.

"Row" type with timestamp fields

This example demonstrates the principle:

create type row_t as (f1 timestamp, f2 timestamp);
create table t(k serial primary key, v1 row_t, v2 row_t);
insert into t(v1) values (('2019-01-27 11:48:33', '2020-03-30 14:19:21')::row_t);
select v1::text as text_typecast from t where k = 1
\gset result_
\echo :result_text_typecast

The \echo metacommand shows this:

("2019-01-27 11:48:33","2020-03-30 14:19:21")

To use the text of the literal that was produced to create a value, you must enquote it and typecast it. Do this with the \set metacommand:

\set canonical_literal '\'':result_text_typecast'\''::row_t
\echo :canonical_literal

. The \echo metacommand now shows this:

'("2019-01-27 11:48:33","2020-03-30 14:19:21")'::row_t

Next, use the canonical literal that you produced to update "t.v2" to confirm that you have recreated the value that the row constructor created:

update t set v2 = :canonical_literal where k = 1;
select (v1 = v2)::text as "v1 = v2" from t where k = 1;

It shows this:

 v1 = v2
---------
 true

Once again, as promised, the canonical form of the array literal does indeed recreate the identical value that the "row" type constructor created.

"Row" type with boolean fields

This example demonstrates the principle:

create type row_t as (f1 boolean, f2 boolean, f3 boolean);
create table t(k serial primary key, v1 row_t, v2 row_t);
insert into t(v1) values ((true, false, null)::row_t);
select v1::text as text_typecast from t where k = 1
\gset result_
\echo :result_text_typecast

The \echo metacommand shows this:

 (t,f,)

To use the text of the literal that was produced to create a value, you must enquote it and typecast it. Do this with the \set metacommand:

\set canonical_literal '\'':result_text_typecast'\''::row_t
\echo :canonical_literal

. The \echo metacommand now shows this:

'(t,f,)'::row_t

Next, use the canonical literal that you produced to update "t.v2" to confirm that the value that the row constructor created was recreated:

update t set v2 = :canonical_literal where k = 1;
select (v1 = v2)::text as "v1 = v2" from t where k = 1;

It shows this:

 v1 = v2
---------
 true

Yet again, as promised, the canonical form of the array literal does indeed recreate the identical value that the "row" type constructor created.

Further examples

There are other cases of interest like this:

  • a "row" type whose definition include one or more fields whose data types are other user-defined "row" types.

The rules for such cases can be determined by induction from the rules that this section has stated and illustrated.

"Row" type literal versus "row" type constructor

The two notions, type constructor and literal, are functionally critically different. You can demonstrate the difference using a DO block, because this lets you use a declared variable. It's more effort to do this using a SQL statement because you'd have to use a scalar subquery in place of the PL/pgSQL variable. The ROW keyword is deliberately omitted here to emphasize its optional status.

create type rt as (n numeric, s text, t timestamp, b boolean);

do $body$
declare
  n constant numeric := 42.17;
  s constant text := 'dog house';
  t constant timestamp := '2020-04-01 23:44:13';
  b constant boolean := true;
  r1 constant rt := (n, s, t, b)::rt;
  r2 constant rt := '(42.17,"dog house","2020-04-01 23:44:13",t)'::rt;
begin
  assert r1 = r2, 'unexpected';
end;
$body$;

You can use the "row" type constructor as an expression in the array[] value constructor). But, of course, you can use only the literal for a "row" type value within the literal for an array. The literal for an array of "row" type values explains this.