Cursors

This section explains:

• what a cursor is;
• how you can manipulate a cursor explicitly, when your use case calls for this, using either a SQL API or a PL/pgSQL API.

The SQL API is exposed by the declare, move, fetch, and close statements. Each of these specifies its cursor using an identifier for the cursor's name.

The functionally equivalent PL/pgSQL API is exposed by the executable statements open, move, fetch, and close. Each of these specifies its cursor using an identifier for the name of a value of the dedicated data type refcursor. You can declare a refcursor, just as you declare other variables, in the PL/pgSQL source code's plpgsql_declaration_section. Notice that one flavor of the declaration syntax lets you specify the defining subquery (see below) for the underlying cursor that it denotes. Alternatively, you can specify the data type of a PL/pgSQL subprogram's formal argument as refcursor. The value of a variable or argument whose data type is refcursor is text and is simply the name of the underlying cursor that it denotes.

What is a cursor?

A cursor is an artifact that you create with the declare SQL statement—or with the equivalent PL/pgSQL open statement. A cursor's duration is limited to the lifetime of the session that creates it, it is private to that session, and it is identified by a bare name that must conform to the usual rules for SQL names like those of tables, schemas, and so on. In the sense that it's session-private, that its name isn't schema-qualified, that it has at most session duration, and that it has no explicit owner, it resembles a prepared statement. A session's currently extant cursors are listed in the pg_cursors catalog view.

A cursor is defined by a subquery (typically a select statement, but you can use a values statement) and it lets you fetch the rows from the result set that the subquery defines one-at-a-time without (in favorable cases) needing ever to materialize the entire result set in your application's client backend server process—i.e. the process that this query lists:

select application_name, backend_type
from pg_stat_activity
where pid = pg_backend_pid();

create table pg_temp.count(n int not null);

prepare stmt as
with
  c(n) as (
    insert into pg_temp.count(n)
    select count(*)
    from pg_class
    where relkind = 'v'
    returning n)
select relname from pg_class where relkind = 'v';

execute stmt;

declare cur cursor for
with
  c(n) as (
    insert into pg_temp.count(n)
    select count(*)
    from pg_class
    where relkind = 'v'
    returning n)
select relname from pg_class where relkind = 'v';

Tautologically, then, a cursor is an artifact that is characterized thus:

• It is created with the declare SQL statement (or the equivalent open PL/pgSQL statement).
• Its maximum duration is either the session in which it's created or the transaction in which it's created, according to a choice that you make when you create it.
• Its lifetime can be terminated deliberately with the close SQL statement or the same-spelled PL/pgSQL statement.
• It is defined by its name, its subquery and some other boolean attributes.
• Its name and its other attributes are listed in the pg_cursors catalog view.
• It lets you fetch consecutive rows, either one at a time, or in batches whose size you choose, from the result set that its subquery defines;
• It supports the move SQL statement, and the same-spelled PL/pgSQL statement, to let you specify any row, by its position, within the result set as the current row.
• It supports the fetch SQL statement, and the same-spelled PL/pgSQL statement, that, in one variant, lets you fetch the row at the current position or a row at any other position relative to the current position (either ahead of it or behind it).
• It supports another variant of the fetch statement (but here only in SQL) that lets you fetch a specified number of rows either forward from and including the row immediately after the current position or backward from and including the row immediately before the current position.

The "current position" notion is defined by imagining that the cursor's defining subquery always includes this select list item:

row_number() over() as current_position

Here, over() spans the entire result set. If the overall query has no order by clause, then over() has no such clause either. But if the overall query does have an order by clause, then over() has the same order by clause within its parentheses.

When you execute the move statement, the current position is left at the result to which you moved. And execute the fetch statement (fetching one or several rows in either the forward or the backward direction) the current position is left at the last-fetched result.

Simple demonstration

Set the psql variables db and u to, respectively, a convenient sandbox database and a convenient test role that has connect and create on that database. Then create a trivial helper that will delay the delivery of each row from a subquery's result set:

\c :db :u
drop schema if exists s cascade;
create schema s;

create function s.sleep_one_second()
  returns boolean
  set search_path = pg_catalog, pg_text
  language plpgsql
as $body$
begin
  perform pg_sleep(1.0);
  return true;
end;
$body$;

Use it to define a view whose result set is ten rows:

create view s.ten_rows(v) as
select a.v from generate_series(1, 10) as a(v) where s.sleep_one_second();

Now execute a query in the obvious way at the ysqlsh prompt:

select v from s.ten_rows;

It takes about ten seconds before you see any results—and then you see all ten rows effectively instantaneously. (Use the \timing on meta-command to time it.) In other words, all ten rows were first materialized in the backend server process's memory before being passed, in a single round trip, to the application.

Create another helper function to fetch one row from a cursor and to return its value together with the time taken to fetch it;

create function s.next_row(cur in refcursor)
  returns text
  set search_path = pg_catalog, pg_text
  language plpgsql
as $body$
declare
  t0  float8 not null := 0;
  t1  float8 not null := 0;
  t   int    not null := 0;
  v   int;
begin
  t0 := extract(epoch from clock_timestamp());
  fetch next  from cur into v;
  t1 := extract(epoch from clock_timestamp());

  t := (round(t1 - t0)*1000.0)::int;
  return rpad(coalesce(v::text, '<no row>'), 9)||'-- '||to_char(t, '9999')||' ms';
end;
$body$;

Now use a cursor to fetch the ten rows one by one:

\t on
start transaction;
  declare "My Cursor" no scroll cursor without hold for
  select v from s.ten_rows;

  select s.next_row('My Cursor');
  select s.next_row('My Cursor');
  select s.next_row('My Cursor');
  select s.next_row('My Cursor');
  select s.next_row('My Cursor');
  select s.next_row('My Cursor');
  select s.next_row('My Cursor');
  select s.next_row('My Cursor');
  select s.next_row('My Cursor');
  select s.next_row('My Cursor');
  select s.next_row('My Cursor');

rollback;
\t off

Now you see the rows delivered one by one, every second. This is the result. (Blank lines were removed manually.)

 1        --  1000 ms
 2        --  1000 ms
 3        --  1000 ms
 4        --  1000 ms
 5        --  1000 ms
 6        --  1000 ms
 7        --  1000 ms
 8        --  1000 ms
 9        --  1000 ms
 10       --  1000 ms
 <no row> --     0 ms

When you execute "select v from ten_rows" ordinarily using ysqlsh, you have to wait until the entire result set has been materialized in the memory of its backend server process before it's delivered to the client application as a unit. This incurs a memory usage cost as well as a time-delay irritation. But when you declare a cursor for that select statement, you materialize the results one row at a time and deliver each to the client as soon as its available. When you use this approach, no more than a single row needs ever to be concurrently materialized in the backend server process's memory.

In real applications, you'll use the piecewise result set delivery that a cursor supports only when the result set is vast; and you'll fetch it in batches of a suitable size: small enough that the backend server process's memory isn't over-consumed; but large enough that the round-trip time doesn't dominate the overall cost of fetching a batch.

Transactional behavior — holdable and non-holdable cursors

A cursor can be declared either as so-called holdable—or not. See the account of the with hold or without hold choice in the section for the declare statement. Try this:

\c :db :u
select count(*) from pg_cursors;

start transaction;
  declare "Not Holdable" cursor without hold for select 17;
  declare "Is Holdable"  cursor with    hold for select 42;
  select name, is_holdable::text from pg_cursors order by name;
commit;
select name, is_holdable::text from pg_cursors order by name;

\c :db :u
select count(*) from pg_cursors;

An invocation of "select count(*) from pg_cursors", immediately after starting a session, will inevitably report that no cursors exist. The first pg_cursors query (within the ongoing transaction) produces this result:

     name     | is_holdable 
--------------+-------------
 Is Holdable  | true
 Not Holdable | false

And the pg_cursors query immediately after committing the transaction produces this result:

     name     | is_holdable 
--------------+-------------
 Is Holdable  | true

In other words, a non-holdable cursor will vanish when the transaction within which it was declared ends—even if the transaction is committed. Because a non-holdable cursor cannot exist outside of an ongoing transaction, this attempt:

declare "Not Holdable" cursor without hold for select 17;

causes the 25P01 error: DECLARE CURSOR can only be used in transaction blocks. The wording is slightly confusing because this causes no such error:

declare "Is Holdable"  cursor with    hold for select 42;

See the section The transaction model for top-level SQL statements. The assumption is that you're running ysqlsh with the default setting of 'on' for the psql variable AUTOCOMMIT.

Notice that the transactional behavior of a cursor differs critically from that of a prepared statement:

\c :db :u
select count(*) from pg_prepared_statements;

start transaction;
  prepare stmt as select 42;
rollback;
select name from pg_prepared_statements;

Like is the case for cursors, an invocation of "select count(*) from pg_prepared_statements", immediately after starting a session, will inevitably report that no prepared statements exist. But even when a statement is prepared within a transaction that is rolled back, it continues to exist after that until either the session ends or it is deallocated. (If you create a holdable cursor, within an on going transaction and then roll back the transaction, then it vanishes.)

Scrollable cursors

When you choose, at cursor creation time, either the scroll or the no scroll options, the result of your choice is shown in the is_scrollable column in the pg_cursors view. Try this:

start transaction;
  declare "Not Scrollable" no scroll cursor without hold for
    select g.v from generate_series(1, 5) as g(v);
  declare "Is Scrollable"     scroll cursor without hold for
    select g.v from generate_series(1, 5) as g(v);

  select name, is_scrollable::text from pg_cursors order by name;
rollback;

This is the result:

      name      | is_scrollable 
----------------+---------------
 Is Scrollable  | true
 Not Scrollable | false

The term of art scrollable reflects a rather unusual meaning of scrollability. In, for example, discussions about GUIs, scrolling means moving forwards or backwards within a window or, say, a list. However:

• When pg_cursors.is_scrollable is false, this means that you can change the current position in the cursor's result set (using either move or as a consequence of fetch) only in the forward direction.
• When pg_cursors.is_scrollable is true, this means that you can change the current position in the cursor's result set both in the forward direction and in the backward direction.

In other words:

• When you create a cursor and specify no scroll, you're saying that you will allow changing the current position in the result set in only the forward direction.
• When you create a cursor and specify scroll, you're saying that you will allow changing the current position in the result set in both the forward direction and the backward direction.

Notice that your choice with the move statement, to change the current position by just a single row or by many rows is an orthogonal choice to the direction in which you move. Similarly, your choice with the fetch statement, to fetch just a single row or many rows is an orthogonal choice to the direction in which you fetch.

• There is no way to create a cursor so that changing the current position in the result set by more than one row, except by consecutive fetches, is prevented.

"no scroll" cursor demonstration

Do this:

set client_min_messages = error;
start transaction;
  declare cur no scroll cursor without hold for
    select g.v from generate_series(1, 5) as g(v);
  fetch next       from cur;
  move relative 2  in   cur;
  fetch forward 2  from cur;

So far, this runs without error and produces these results, as expected:

 1

 4
 5

Now do this:

  fetch relative 0 from cur;
rollback;

The fetch causes the 55000 error: cursor can only scan forward because fetching a cursor's current row tautologically does not move the current position forward and is therefore not allowed when you create the cursor using no scroll.

"scroll" cursor demonstration

Now do this:

set client_min_messages = error;
start transaction;
  declare cur scroll cursor without hold for
    select g.v from generate_series(1, 5) as g(v);

  -- These are the same four statements that the "no scroll" demo used.
  -- But now, "fetch relative 0" succeeds.
  fetch next         from cur;
  move  relative 2   in   cur;
  fetch forward 2    from cur;
  fetch relative 0   from cur;

  -- Add these statements.
  fetch backward     from cur;
  fetch backward all from cur;
  move  last         in  cur;
  fetch relative 0   from cur;
  move  first        in   cur;
  fetch relative 0   from cur;
rollback;

Now every fetch and move statement succeeds. These are the results:

 1

 4
 5

 5

 4

 3
 2
 1

 5

 1

Caching a cursor's result set

• The result set for a with hold cursor is always cached when the transaction that creates it commits.

• The result set for a without hold cursor might, or might not, be cached.

  • If the execution plan for the cursor's defining subquery can be executed in either the forward or the backward direction, then the result set will not be cached.

  • But if the plan can be executed only in the forward direction, then the result set must be cached if you specify scroll when you create the cursor.

PostgreSQL, and therefore YSQL, do not expose metadata to report whether or not a cursor's result set is cached. Nor does the documentation for either RDBMS attempt to specify the rules that determine whether caching will be done. However, it's possible to reason, about certain specific select statements, that their plans cannot be executed backward. For example the plan for a query that includes row_number() in the select list cannot be run backward because the semantics of row_number() is to assign an incrementing rank to each new row in the result set as it is produced when the plan is executed in the forward direction—and the planner cannot predict how many rows will be produced to allow row_number() to be calculated by decrementing from this for each successive row when the plan is run backward. (If the select statement has no order by, then the rows are produced in physical order (i.e. in an order that's determined by how the table data is stored).

You can, however, support pedagogy by including a user-defined function in the where clause that always returns true and that uses raise info to report when it's invoked.

First, do this set-up. All the caching tests will use it:

set client_min_messages = error;
drop schema if exists s cascade;
create schema s;

create table s.t(k serial primary key, v int not null);
insert into s.t(v) select generate_series(1, 5);
create view s.v(pos, v) as select row_number() over(), v from s.t;

create function s.f(i in int)
  returns boolean
  set search_path = pg_catalog, pg_temp
  volatile
  language plpgsql
as $body$
begin
  raise info 'f() invoked';
  return i = i;
end;
$body$;

"with hold", "no scroll" cursor

Do this:

start transaction;
  declare cur no scroll cursor with hold for
    select pos, v from s.v where s.f(v);

It completes silently without error. Now do this:

commit;

This is when you see the raise info output—five times in total, i.e. once for each row that's tested:

INFO:  f() invoked
INFO:  f() invoked
INFO:  f() invoked
INFO:  f() invoked
INFO:  f() invoked

This demonstrates that the result set has been cached. Now fetch all the rows and close the cursor;

fetch all from cur;
close cur;

The fetch all statement brings this SQL output:

   1 | 5
   2 | 1
   3 | 4
   4 | 2
   5 | 3

"without hold", "no scroll" cursor

Do this:

start transaction;
  declare cur no scroll cursor without hold for
    select pos, v from s.v where s.f(v);

    fetch next from cur;
    fetch next from cur;
    fetch next from cur;
    fetch next from cur;
    fetch next from cur;
rollback;

It's easier to distinguish the raise info output and the SQL output from the statements that bring these if you save this code to, say, t.sql, and then execute it at the ysqlsh prompt. This is what you see.

INFO:  f() invoked
   1 | 5

INFO:  f() invoked
   2 | 1

INFO:  f() invoked
   3 | 4

INFO:  f() invoked
   4 | 2

INFO:  f() invoked
   5 | 3

Notice that the same v values are paired with the same pos values as with the "with hold", "no scroll" cursor test. But here, nothing suggests that results are cached—and they don't need to be because the cursor doesn't allow moving backwards in the result set. (However, you can't design a test to demonstrate that results are not cached.)

"without hold", "scroll" cursor

Do this:

start transaction;
  declare cur scroll cursor without hold for
    select pos, v from s.v where s.f(v);

  -- Moving forward.
  fetch next       from cur;
  fetch next       from cur;
  fetch next       from cur;
  fetch next       from cur;
  fetch next       from cur;

  -- Moving backward.
  fetch prior      from cur;
  move  absolute 3 in   cur;
  fetch relative 0 from cur;
  move  absolute 1 in   cur;
  fetch relative 0 from cur;
rollback;

Apart from replacing no scroll with scroll in the declaration of the cursor, the code is identical, through all the "Moving forward fetches, to that for the "without hold", "no scroll" cursor test. And the output is identical, too.

Now continue with the tests that move backwards: You no longer see any raise info output because the entire result set is now cached. Here are the SQL results:

   4 | 2

   3 | 4

   1 | 5

Notice that you see the same v values for the same pos values as before. This is required by the semantics of moving forward and backward through the unique result set that the cursor's subquery defines.

"without hold", "scroll" cursor — variant

The output from the previous test shows that the result set is cached incrementally. This variant demonstrates it more vividly by interleaving forward and backward fetches and moves. First, increase the number of rows in s.t and inspect s.v:

delete from s.t;
insert into s.t(v) select generate_series(1, 10);
select pos, v from s.v;

This is the result:

   1 |  6
   2 |  8
   3 |  7
   4 | 10
   5 |  1
   6 |  2
   7 |  4
   8 |  5
   9 |  3
  10 |  9

You might see the v values in a different order. See the note The physical order cannot be predicted above.

Now do this:

start transaction;
  declare cur scroll cursor without hold for
    select pos, v from s.v where s.f(v);

This succeeds silently without error. Now do this:

  fetch forward  3 from cur;
  fetch backward 2 from cur;

It causes this output:

INFO:  f() invoked
INFO:  f() invoked
INFO:  f() invoked
   1 | 6
   2 | 8
   3 | 7

   2 | 8
   1 | 6

We don't see any raise info output when we fetch the rows with pos = 2 and pos = 1 again because they're already cached. Now do this:

  fetch forward  5 from cur;
  fetch backward 3 from cur;

It causes this output:

INFO:  f() invoked
INFO:  f() invoked
INFO:  f() invoked
   2 |  8
   3 |  7
   4 | 10
   5 |  1
   6 |  2

   5 |  1
   4 | 10
   3 |  7

Because we're now seeing the three rows with pos = 4, pos = 5, and pos = 6 for the first time, we see the raise info output three times. Then, when we revisit the rows with pos = 5, pos = 4, and pos = 3, we don't see any raise info output because, again, they're already cached. Now do this:

  fetch forward  5 from cur;
  fetch backward 1 from cur;

It causes this output:

INFO:  f() invoked
INFO:  f() invoked
   4 | 10
   5 |  1
   6 |  2
   7 |  4
   8 |  5

   7 | 4

Here, we're seeing the rows with pos = 7 and pos = 8 for the first time—which brings the raise info output twice. And then, because we've already (just) seen the row with pos 7, there's now no raise info output.

Finally, do this:

  move  absolute 3 in cur;
  fetch forward  7 from cur;
rollback;

It causes this output:

INFO:  f() invoked
INFO:  f() invoked
   4 | 10
   5 |  1
   6 |  2
   7 |  4
   8 |  5
   9 |  3
  10 |  9

At this stage, we've seen all rows in the result set but those with pos = 9 and pos = 10, so we see the raise info for just each of those two rows.

Beware Issue #6514