Supported Versions: Current (16) / 15 / 14 / 13 / 12
Development Versions: devel
Unsupported versions: 11 / 10 / 9.6 / 9.5 / 9.4 / 9.3 / 9.2 / 9.1 / 9.0 / 8.4 / 8.3 / 8.2 / 8.1 / 8.0 / 7.4 / 7.3 / 7.2 / 7.1
This documentation is for an unsupported version of PostgreSQL.
You may want to view the same page for the current version, or one of the other supported versions listed above instead.

17.2. Views and the Rule System

17.2.1. Implementation of Views in Postgres

Views in Postgres are implemented using the rule system. In fact there is absolutely no difference between a

    CREATE VIEW myview AS SELECT * FROM mytab;
compared against the two commands
    CREATE TABLE myview (same attribute list as for mytab);
    CREATE RULE "_RETmyview" AS ON SELECT TO myview DO INSTEAD
        SELECT * FROM mytab;
because this is exactly what the CREATE VIEW command does internally. This has some side effects. One of them is that the information about a view in the Postgres system catalogs is exactly the same as it is for a table. So for the query parser, there is absolutely no difference between a table and a view. They are the same thing - relations. That is the important one for now.

17.2.2. How SELECT Rules Work

Rules ON SELECT are applied to all queries as the last step, even if the command given is an INSERT, UPDATE or DELETE. And they have different semantics from the others in that they modify the parsetree in place instead of creating a new one. So SELECT rules are described first.

Currently, there can be only one action in an ON SELECT rule, and it must be an unconditional SELECT action that is INSTEAD. This restriction was required to make rules safe enough to open them for ordinary users and it restricts rules ON SELECT to real view rules.

The examples for this document are two join views that do some calculations and some more views using them in turn. One of the two first views is customized later by adding rules for INSERT, UPDATE and DELETE operations so that the final result will be a view that behaves like a real table with some magic functionality. It is not such a simple example to start from and this makes things harder to get into. But it's better to have one example that covers all the points discussed step by step rather than having many different ones that might mix up in mind.

The database needed to play with the examples is named al_bundy. You'll see soon why this is the database name. And it needs the procedural language PL/pgSQL installed, because we need a little min() function returning the lower of 2 integer values. We create that as

    CREATE FUNCTION min(integer, integer) RETURNS integer AS
        'BEGIN
            IF $1 < $2 THEN
                RETURN $1;
            END IF;
            RETURN $2;
        END;'
    LANGUAGE 'plpgsql';

The real tables we need in the first two rule system descriptions are these:

    CREATE TABLE shoe_data (
        shoename   char(10),      -- primary key
        sh_avail   integer,       -- available # of pairs
        slcolor    char(10),      -- preferred shoelace color
        slminlen   float,         -- miminum shoelace length
        slmaxlen   float,         -- maximum shoelace length
        slunit     char(8)        -- length unit
    );

    CREATE TABLE shoelace_data (
        sl_name    char(10),      -- primary key
        sl_avail   integer,       -- available # of pairs
        sl_color   char(10),      -- shoelace color
        sl_len     float,         -- shoelace length
        sl_unit    char(8)        -- length unit
    );

    CREATE TABLE unit (
        un_name    char(8),       -- the primary key
        un_fact    float          -- factor to transform to cm
    );
I think most of us wear shoes and can realize that this is really useful data. Well there are shoes out in the world that don't require shoelaces, but this doesn't make Al's life easier and so we ignore it.

The views are created as

    CREATE VIEW shoe AS
        SELECT sh.shoename,
               sh.sh_avail,
               sh.slcolor,
               sh.slminlen,
               sh.slminlen * un.un_fact AS slminlen_cm,
               sh.slmaxlen,
               sh.slmaxlen * un.un_fact AS slmaxlen_cm,
               sh.slunit
          FROM shoe_data sh, unit un
         WHERE sh.slunit = un.un_name;

    CREATE VIEW shoelace AS
        SELECT s.sl_name,
               s.sl_avail,
               s.sl_color,
               s.sl_len,
               s.sl_unit,
               s.sl_len * u.un_fact AS sl_len_cm
          FROM shoelace_data s, unit u
         WHERE s.sl_unit = u.un_name;

    CREATE VIEW shoe_ready AS
        SELECT rsh.shoename,
               rsh.sh_avail,
               rsl.sl_name,
               rsl.sl_avail,
               min(rsh.sh_avail, rsl.sl_avail) AS total_avail
          FROM shoe rsh, shoelace rsl
         WHERE rsl.sl_color = rsh.slcolor
           AND rsl.sl_len_cm >= rsh.slminlen_cm
           AND rsl.sl_len_cm <= rsh.slmaxlen_cm;
The CREATE VIEW command for the shoelace view (which is the simplest one we have) will create a relation shoelace and an entry in pg_rewrite that tells that there is a rewrite rule that must be applied whenever the relation shoelace is referenced in a query's rangetable. The rule has no rule qualification (discussed later, with the non SELECT rules, since SELECT rules currently cannot have them) and it is INSTEAD. Note that rule qualifications are not the same as query qualifications! The rule's action has a query qualification.

The rule's action is one querytree that is a copy of the SELECT statement in the view creation command.

Note: The two extra range table entries for NEW and OLD (named *NEW* and *CURRENT* for historical reasons in the printed querytree) you can see in the pg_rewrite entry aren't of interest for SELECT rules.

Now we populate unit, shoe_data and shoelace_data and Al types the first SELECT in his life:
    al_bundy=> INSERT INTO unit VALUES ('cm', 1.0);
    al_bundy=> INSERT INTO unit VALUES ('m', 100.0);
    al_bundy=> INSERT INTO unit VALUES ('inch', 2.54);
    al_bundy=> 
    al_bundy=> INSERT INTO shoe_data VALUES 
    al_bundy->     ('sh1', 2, 'black', 70.0, 90.0, 'cm');
    al_bundy=> INSERT INTO shoe_data VALUES 
    al_bundy->     ('sh2', 0, 'black', 30.0, 40.0, 'inch');
    al_bundy=> INSERT INTO shoe_data VALUES 
    al_bundy->     ('sh3', 4, 'brown', 50.0, 65.0, 'cm');
    al_bundy=> INSERT INTO shoe_data VALUES 
    al_bundy->     ('sh4', 3, 'brown', 40.0, 50.0, 'inch');
    al_bundy=> 
    al_bundy=> INSERT INTO shoelace_data VALUES 
    al_bundy->     ('sl1', 5, 'black', 80.0, 'cm');
    al_bundy=> INSERT INTO shoelace_data VALUES 
    al_bundy->     ('sl2', 6, 'black', 100.0, 'cm');
    al_bundy=> INSERT INTO shoelace_data VALUES 
    al_bundy->     ('sl3', 0, 'black', 35.0 , 'inch');
    al_bundy=> INSERT INTO shoelace_data VALUES 
    al_bundy->     ('sl4', 8, 'black', 40.0 , 'inch');
    al_bundy=> INSERT INTO shoelace_data VALUES 
    al_bundy->     ('sl5', 4, 'brown', 1.0 , 'm');
    al_bundy=> INSERT INTO shoelace_data VALUES 
    al_bundy->     ('sl6', 0, 'brown', 0.9 , 'm');
    al_bundy=> INSERT INTO shoelace_data VALUES 
    al_bundy->     ('sl7', 7, 'brown', 60 , 'cm');
    al_bundy=> INSERT INTO shoelace_data VALUES 
    al_bundy->     ('sl8', 1, 'brown', 40 , 'inch');
    al_bundy=> 
    al_bundy=> SELECT * FROM shoelace;
    sl_name   |sl_avail|sl_color  |sl_len|sl_unit |sl_len_cm
    ----------+--------+----------+------+--------+---------
    sl1       |       5|black     |    80|cm      |       80
    sl2       |       6|black     |   100|cm      |      100
    sl7       |       7|brown     |    60|cm      |       60
    sl3       |       0|black     |    35|inch    |     88.9
    sl4       |       8|black     |    40|inch    |    101.6
    sl8       |       1|brown     |    40|inch    |    101.6
    sl5       |       4|brown     |     1|m       |      100
    sl6       |       0|brown     |   0.9|m       |       90
    (8 rows)
It's the simplest SELECT Al can do on our views, so we take this to explain the basics of view rules. The 'SELECT * FROM shoelace' was interpreted by the parser and produced the parsetree
    SELECT shoelace.sl_name, shoelace.sl_avail,
           shoelace.sl_color, shoelace.sl_len,
           shoelace.sl_unit, shoelace.sl_len_cm
      FROM shoelace shoelace;
and this is given to the rule system. The rule system walks through the rangetable and checks if there are rules in pg_rewrite for any relation. When processing the rangetable entry for shoelace (the only one up to now) it finds the rule '_RETshoelace' with the parsetree
    SELECT s.sl_name, s.sl_avail,
           s.sl_color, s.sl_len, s.sl_unit,
           float8mul(s.sl_len, u.un_fact) AS sl_len_cm
      FROM shoelace *OLD*, shoelace *NEW*,
           shoelace_data s, unit u
     WHERE bpchareq(s.sl_unit, u.un_name);
Note that the parser changed the calculation and qualification into calls to the appropriate functions. But in fact this changes nothing.

To expand the view, the rewriter simply creates a subselect rangetable entry containing the rule's action parsetree, and substitutes this rangetable entry for the original one that referenced the view. The resulting rewritten parsetree is almost the same as if Al had typed

    SELECT shoelace.sl_name, shoelace.sl_avail,
           shoelace.sl_color, shoelace.sl_len,
           shoelace.sl_unit, shoelace.sl_len_cm
      FROM (SELECT s.sl_name,
                   s.sl_avail,
                   s.sl_color,
                   s.sl_len,
                   s.sl_unit,
                   s.sl_len * u.un_fact AS sl_len_cm
              FROM shoelace_data s, unit u
             WHERE s.sl_unit = u.un_name) shoelace;
There is one difference however: the sub-query's rangetable has two extra entries shoelace *OLD*, shoelace *NEW*. These entries don't participate directly in the query, since they aren't referenced by the sub-query's join tree or targetlist. The rewriter uses them to store the access permission check info that was originally present in the rangetable entry that referenced the view. In this way, the executor will still check that the user has proper permissions to access the view, even though there's no direct use of the view in the rewritten query.

That was the first rule applied. The rule system will continue checking the remaining rangetable entries in the top query (in this example there are no more), and it will recursively check the rangetable entries in the added sub-query to see if any of them reference views. (But it won't expand *OLD* or *NEW* --- otherwise we'd have infinite recursion!) In this example, there are no rewrite rules for shoelace_data or unit, so rewriting is complete and the above is the final result given to the planner.

Now we face Al with the problem that the Blues Brothers appear in his shop and want to buy some new shoes, and as the Blues Brothers are, they want to wear the same shoes. And they want to wear them immediately, so they need shoelaces too.

Al needs to know for which shoes currently in the store he has the matching shoelaces (color and size) and where the total number of exactly matching pairs is greater or equal to two. We teach him what to do and he asks his database:

    al_bundy=> SELECT * FROM shoe_ready WHERE total_avail >= 2;
    shoename  |sh_avail|sl_name   |sl_avail|total_avail
    ----------+--------+----------+--------+-----------
    sh1       |       2|sl1       |       5|          2
    sh3       |       4|sl7       |       7|          4
    (2 rows)
Al is a shoe guru and so he knows that only shoes of type sh1 would fit (shoelace sl7 is brown and shoes that need brown shoelaces aren't shoes the Blues Brothers would ever wear).

The output of the parser this time is the parsetree

    SELECT shoe_ready.shoename, shoe_ready.sh_avail,
           shoe_ready.sl_name, shoe_ready.sl_avail,
           shoe_ready.total_avail
      FROM shoe_ready shoe_ready
     WHERE int4ge(shoe_ready.total_avail, 2);
The first rule applied will be the one for the shoe_ready view and it results in the parsetree
    SELECT shoe_ready.shoename, shoe_ready.sh_avail,
           shoe_ready.sl_name, shoe_ready.sl_avail,
           shoe_ready.total_avail
      FROM (SELECT rsh.shoename,
                   rsh.sh_avail,
                   rsl.sl_name,
                   rsl.sl_avail,
                   min(rsh.sh_avail, rsl.sl_avail) AS total_avail
              FROM shoe rsh, shoelace rsl
             WHERE rsl.sl_color = rsh.slcolor
               AND rsl.sl_len_cm >= rsh.slminlen_cm
               AND rsl.sl_len_cm <= rsh.slmaxlen_cm) shoe_ready
     WHERE int4ge(shoe_ready.total_avail, 2);
Similarly, the rules for shoe and shoelace are substituted into the rangetable of the sub-query, leading to a three-level final querytree:
    SELECT shoe_ready.shoename, shoe_ready.sh_avail,
           shoe_ready.sl_name, shoe_ready.sl_avail,
           shoe_ready.total_avail
      FROM (SELECT rsh.shoename,
                   rsh.sh_avail,
                   rsl.sl_name,
                   rsl.sl_avail,
                   min(rsh.sh_avail, rsl.sl_avail) AS total_avail
              FROM (SELECT sh.shoename,
                           sh.sh_avail,
                           sh.slcolor,
                           sh.slminlen,
                           sh.slminlen * un.un_fact AS slminlen_cm,
                           sh.slmaxlen,
                           sh.slmaxlen * un.un_fact AS slmaxlen_cm,
                           sh.slunit
                      FROM shoe_data sh, unit un
                     WHERE sh.slunit = un.un_name) rsh,
                   (SELECT s.sl_name,
                           s.sl_avail,
                           s.sl_color,
                           s.sl_len,
                           s.sl_unit,
                           s.sl_len * u.un_fact AS sl_len_cm
                      FROM shoelace_data s, unit u
                     WHERE s.sl_unit = u.un_name) rsl
             WHERE rsl.sl_color = rsh.slcolor
               AND rsl.sl_len_cm >= rsh.slminlen_cm
               AND rsl.sl_len_cm <= rsh.slmaxlen_cm) shoe_ready
     WHERE int4ge(shoe_ready.total_avail, 2);
It turns out that the planner will collapse this tree into a two-level querytree: the bottommost selects will be "pulled up" into the middle select since there's no need to process them separately. But the middle select will remain separate from the top, because it contains aggregate functions. If we pulled those up it would change the behavior of the topmost select, which we don't want. However, collapsing the query tree is an optimization that the rewrite system doesn't have to concern itself with.

Note: There is currently no recursion stopping mechanism for view rules in the rule system (only for the other kinds of rules). This doesn't hurt much, because the only way to push this into an endless loop (blowing up the backend until it reaches the memory limit) is to create tables and then setup the view rules by hand with CREATE RULE in such a way, that one selects from the other that selects from the one. This could never happen if CREATE VIEW is used because for the first CREATE VIEW, the second relation does not exist and thus the first view cannot select from the second.

17.2.3. View Rules in Non-SELECT Statements

Two details of the parsetree aren't touched in the description of view rules above. These are the commandtype and the resultrelation. In fact, view rules don't need this information.

There are only a few differences between a parsetree for a SELECT and one for any other command. Obviously they have another commandtype and this time the resultrelation points to the rangetable entry where the result should go. Everything else is absolutely the same. So having two tables t1 and t2 with attributes a and b, the parsetrees for the two statements

    SELECT t2.b FROM t1, t2 WHERE t1.a = t2.a;

    UPDATE t1 SET b = t2.b WHERE t1.a = t2.a;
are nearly identical.
  • The rangetables contain entries for the tables t1 and t2.

  • The targetlists contain one variable that points to attribute b of the rangetable entry for table t2.

  • The qualification expressions compare the attributes a of both ranges for equality.

  • The jointrees show a simple join between t1 and t2.

The consequence is, that both parsetrees result in similar execution plans. They are both joins over the two tables. For the UPDATE the missing columns from t1 are added to the targetlist by the planner and the final parsetree will read as
    UPDATE t1 SET a = t1.a, b = t2.b WHERE t1.a = t2.a;
and thus the executor run over the join will produce exactly the same result set as a
    SELECT t1.a, t2.b FROM t1, t2 WHERE t1.a = t2.a;
will do. But there is a little problem in UPDATE. The executor does not care what the results from the join it is doing are meant for. It just produces a result set of rows. The difference that one is a SELECT command and the other is an UPDATE is handled in the caller of the executor. The caller still knows (looking at the parsetree) that this is an UPDATE, and he knows that this result should go into table t1. But which of the rows that are there has to be replaced by the new row?

To resolve this problem, another entry is added to the targetlist in UPDATE (and also in DELETE) statements: the current tuple ID (ctid). This is a system attribute containing the file block number and position in the block for the row. Knowing the table, the ctid can be used to retrieve the original t1 row to be updated. After adding the ctid to the targetlist, the query actually looks like

    SELECT t1.a, t2.b, t1.ctid FROM t1, t2 WHERE t1.a = t2.a;
Now another detail of Postgres enters the stage. At this moment, table rows aren't overwritten and this is why ABORT TRANSACTION is fast. In an UPDATE, the new result row is inserted into the table (after stripping ctid) and in the tuple header of the row that ctid pointed to the cmax and xmax entries are set to the current command counter and current transaction ID. Thus the old row is hidden and after the transaction commited the vacuum cleaner can really move it out.

Knowing all that, we can simply apply view rules in absolutely the same way to any command. There is no difference.

17.2.4. The Power of Views in Postgres

The above demonstrates how the rule system incorporates view definitions into the original parsetree. In the second example a simple SELECT from one view created a final parsetree that is a join of 4 tables (unit is used twice with different names).

17.2.4.1. Benefits

The benefit of implementing views with the rule system is, that the planner has all the information about which tables have to be scanned plus the relationships between these tables plus the restrictive qualifications from the views plus the qualifications from the original query in one single parsetree. And this is still the situation when the original query is already a join over views. Now the planner has to decide which is the best path to execute the query. The more information the planner has, the better this decision can be. And the rule system as implemented in Postgres ensures, that this is all information available about the query up to now.

17.2.5. What about updating a view?

What happens if a view is named as the target relation for an INSERT, UPDATE, or DELETE? After doing the substitutions described above, we will have a querytree in which the resultrelation points at a subquery rangetable entry. This will not work, so the rewriter throws an error if it sees it has produced such a thing.

To change this we can define rules that modify the behaviour of non-SELECT queries. This is the topic of the next section.