The PostgreSQL 9.0 Reference Manual - Volume 2 - Programming Guide by The PostgreSQL Global Development Group Paperback (6"x9"), 478 pages ISBN 9781906966065 RRP £14.95 ($19.95) Sales of this book support the PostgreSQL project! Get a printed copy>>>

PostgreSQL 9.0 Reference Manual Volume 2 Buy the book here! >>> support free software

9.10.2 Plan Caching

The PL/pgSQL interpreter parses the function's source text and produces an internal binary instruction tree the first time the function is called (within each session). The instruction tree fully translates the PL/pgSQL statement structure, but individual SQL expressions and SQL commands used in the function are not translated immediately.

As each expression and SQL command is first executed in the function, the PL/pgSQL interpreter creates a prepared execution plan (using the SPI manager's SPI_prepare and SPI_saveplan functions). Subsequent visits to that expression or command reuse the prepared plan. Thus, a function with conditional code that contains many statements for which execution plans might be required will only prepare and save those plans that are really used during the lifetime of the database connection. This can substantially reduce the total amount of time required to parse and generate execution plans for the statements in a PL/pgSQL function. A disadvantage is that errors in a specific expression or command cannot be detected until that part of the function is reached in execution. (Trivial syntax errors will be detected during the initial parsing pass, but anything deeper will not be detected until execution.)

A saved plan will be re-planned automatically if there is any schema change to any table used in the query, or if any user-defined function used in the query is redefined. This makes the re-use of prepared plans transparent in most cases, but there are corner cases where a stale plan might be re-used. An example is that dropping and re-creating a user-defined operator won't affect already-cached plans; they'll continue to call the original operator's underlying function, if that has not been changed. When necessary, the cache can be flushed by starting a fresh database session.

Because PL/pgSQL saves execution plans in this way, SQL commands that appear directly in a PL/pgSQL function must refer to the same tables and columns on every execution; that is, you cannot use a parameter as the name of a table or column in an SQL command. To get around this restriction, you can construct dynamic commands using the PL/pgSQL EXECUTE statement--at the price of constructing a new execution plan on every execution.

Another important point is that the prepared plans are parameterized to allow the values of PL/pgSQL variables to change from one use to the next, as discussed in detail above. Sometimes this means that a plan is less efficient than it would be if generated for a specific variable value. As an example, consider

SELECT * INTO myrec FROM dictionary WHERE word LIKE search_term;

where search_term is a PL/pgSQL variable. The cached plan for this query will never use an index on word, since the planner cannot assume that the LIKE pattern will be left-anchored at run time. To use an index the query must be planned with a specific constant LIKE pattern provided. This is another situation where EXECUTE can be used to force a new plan to be generated for each execution.

The mutable nature of record variables presents another problem in this connection. When fields of a record variable are used in expressions or statements, the data types of the fields must not change from one call of the function to the next, since each expression will be planned using the data type that is present when the expression is first reached. EXECUTE can be used to get around this problem when necessary.

If the same function is used as a trigger for more than one table, PL/pgSQL prepares and caches plans independently for each such table--that is, there is a cache for each trigger function and table combination, not just for each function. This alleviates some of the problems with varying data types; for instance, a trigger function will be able to work successfully with a column named key even if it happens to have different types in different tables.

Likewise, functions having polymorphic argument types have a separate plan cache for each combination of actual argument types they have been invoked for, so that data type differences do not cause unexpected failures.

Plan caching can sometimes have surprising effects on the interpretation of time-sensitive values. For example there is a difference between what these two functions do:

CREATE FUNCTION logfunc1(logtxt text) RETURNS void AS $$
    BEGIN
        INSERT INTO logtable VALUES (logtxt, 'now');
    END;
$$ LANGUAGE plpgsql;

and:

CREATE FUNCTION logfunc2(logtxt text) RETURNS void AS $$
    DECLARE
        curtime timestamp;
    BEGIN
        curtime := 'now';
        INSERT INTO logtable VALUES (logtxt, curtime);
    END;
$$ LANGUAGE plpgsql;

In the case of logfunc1, the PostgreSQL main parser knows when preparing the plan for the INSERT that the string 'now' should be interpreted as timestamp, because the target column of logtable is of that type. Thus, 'now' will be converted to a constant when the INSERT is planned, and then used in all invocations of logfunc1 during the lifetime of the session. Needless to say, this isn't what the programmer wanted.

In the case of logfunc2, the PostgreSQL main parser does not know what type 'now' should become and therefore it returns a data value of type text containing the string now. During the ensuing assignment to the local variable curtime, the PL/pgSQL interpreter casts this string to the timestamp type by calling the text_out and timestamp_in functions for the conversion. So, the computed time stamp is updated on each execution as the programmer expects.