PL/pgSQL internals
Pavel Stěhule
Content
● Function management
● Language handlers
● PL/pgSQL function life cycle
● PL/pgSQL architecture
– expression parsing
– expression evaluating
Function management
● described in pg_proc
● indirect call by FunctionCallInvoke(fcinfo)
typedef struct FmgrInfo
{
PGFunction fn_addr; /* pointer to function or handler to be called */
Oid fn_oid; /* OID of function (NOT of handler, if any) */
short fn_nargs; /* 0..FUNC_MAX_ARGS, or -1 if variable arg
* count */
bool fn_strict; /* function is "strict" (NULL in => NULL out) */
bool fn_retset; /* function returns a set */
unsigned char fn_stats; /* collect stats if track_functions > this */
void *fn_extra; /* extra space for use by handler */
MemoryContext fn_mcxt; /* memory context to store fn_extra in */
fmNodePtr fn_expr; /* expression parse tree for call, or NULL */
} FmgrInfo;
#define FunctionCallInvoke(fcinfo) ((* (fcinfo)->flinfo->fn_addr) (fcinfo))
Language handlers
● Call function
– execute code, translate arguments and result (from/to) PG types
● Validator function
– validate record in pg_proc
● Inline function
– execute string
–
postgres=#
– \h CREATE LANGUAGE
Command: CREATE LANGUAGE
Description: define a new procedural language
Syntax:
CREATE [ OR REPLACE ] [ PROCEDURAL ] LANGUAGE name
CREATE [ OR REPLACE ] [ TRUSTED ] [ PROCEDURAL ] LANGUAGE name
HANDLER call_handler [ INLINE inline_handler ] [ VALIDATOR valfunction ]
Predefined language handlers
postgres=# select * from pg_language ;
lanname | lanowner | lanispl | lanpltrusted | lanplcallfoid | laninline | lanvalidator | lanacl
----------+----------+---------+--------------+---------------+-----------+--------------+--------
internal | 10 | f | f | 0 | 0 | 2246 |
c | 10 | f | f | 0 | 0 | 2247 |
sql | 10 | f | t | 0 | 0 | 2248 |
plpgsql | 10 | t | t | 12654 | 12655 | 12656 |
(4 rows)
postgres=# \sf plpgsql_call_handler
CREATE OR REPLACE FUNCTION pg_catalog.plpgsql_call_handler()
RETURNS language_handler
LANGUAGE c
AS '$libdir/plpgsql', $function$plpgsql_call_handler$function$
postgres=# \sf plpgsql_validator
CREATE OR REPLACE FUNCTION pg_catalog.plpgsql_validator(oid)
RETURNS void
LANGUAGE c
STRICT
AS '$libdir/plpgsql', $function$plpgsql_validator$function$
Lifecycle
registration
CREATE
FUNCTION
pg_proc
VALIDATOR
/DEV/NULL
COMPILATION
Lifecycle
invocation
EXEC COMPILATION
NO
found
in cache
CALL
HANDLER PLPGSQL
YES SESSION
CACHE
(code)
EXECUTION
Notes about PL/pgSQL
● has no own expression unit
– expressions are not cheap
● shares data types with PostgreSQL
● run in PostgreSQL session process
– processing query result is not expensive
● no interprocess communication
● no data types conversions
● uses late I/O casting
– expensive - when result is not of same type as
target, then CAST based on IOfunc is invoked
PL/pgSQL architecture
PL/pgSQL
SQL
parser/executor
● parsing embedded queries
● expressions
PL/pgSQL
parser/executor
● PL/pgSQL statements
SPI interface
● Query execution
● Cursors
#option dump
CREATE OR REPLACE FUNCTION Hello(msg text)
RETURNS text AS $$
#option dump
BEGIN
RETURN 'Hello, ' || msg;
END;
$$ LANGUAGE plpgsql IMMUTABLE;
/* from postgresql log */
Execution tree of successfully compiled PL/pgSQL function hello(text):
Function's data area:
entry 0: VAR $1 type text (typoid 25) atttypmod -1
entry 1: VAR found type bool (typoid 16) atttypmod -1
Function's statements:
3:BLOCK <<*unnamed*>>
4: RETURN 'SELECT 'Hello, ' || msg'
END -- *unnamed*
End of execution tree of function hello(text)
"WHILE LOOP" example
CREATE OR REPLACE FUNCTION foo(a int)
RETURNS int AS $$
#option dump
DECLARE
s int := 0;
i int := 1;
BEGIN
WHILE i <= a
LOOP
s := s + i;
i := i + 1;
END LOOP;
RETURN i;
END;
$$ LANGUAGE plpgsql IMMUTABLE;
"WHILE LOOP" dump
Execution tree of successfully compiled PL/pgSQL function foo(integer):
Function's data area:
entry 0: VAR $1 type int4 (typoid 23) atttypmod -1
entry 1: VAR found type bool (typoid 16) atttypmod -1
entry 2: VAR s type int4 (typoid 23) atttypmod -1
DEFAULT 'SELECT 0'
entry 3: VAR i type int4 (typoid 23) atttypmod -1
DEFAULT 'SELECT 1'
Function's statements:
6:BLOCK <<*unnamed*>>
7: WHILE 'SELECT i <= a'
9: ASSIGN var 2 := 'SELECT s + i'
10: ASSIGN var 3 := 'SELECT i + 1'
ENDWHILE
12: RETURN 'SELECT i'
END -- *unnamed*
End of execution tree of function foo(integer)
PLpgSQL architecture
● it is glue for SQL statements
– basic control structures
● IF, WHILE, FOR, BEGIN
– nested variables stack
● assign statement, references to variables
● it is very simple interpret of abstract syntax tree
– PL/pgSQL parser skips expressions
– every node type has exec handler
PLpgSQL architecture
● it is glue for SQL statements
– basic control structures
● IF, WHILE, FOR, BEGIN
– nested variables stack
● assign statement, references to variables
● it is very simple interpret of abstract syntax
tree
– every node type has exec handler
– Execution ~ iteration over nodes (via handler
invocations)
Good to know
● There is no compilation to byte code
● There is no JIT
● There is no any optimization
● Function is compiled to AST when it is first called in
session – source code is in readable form in pg_proc
(compilation is very simple and then relative fast)
● It is just relative simple, relative fast glue of SQL
statements binary compatible with PostgreSQL
(speed is comparable with other simple interprets)
– Reduce network, protocol, ... overheads
– Possible use faster interprets for different task (Perl)
WHILE statement
Syntax
<<label>> WHILE expression
LOOP
{statements}
END LOOP
Basic structure
typedef struct
{ /* WHILE cond LOOP statement */
int cmd_type;
int lineno;
char *label;
PLpgSQL_expr *cond;
List *body; /* List of statements */
} PLpgSQL_stmt_while;
WHILE statemement
parser
proc_stmt : pl_block ';'
{ $$ = $1; }
| stmt_assign { $$ = $1; }
| stmt_while { $$ = $1; }
....
stmt_while : opt_block_label K_WHILE expr_until_loop loop_body
{
PLpgSQL_stmt_while *new;
new = palloc0(sizeof(PLpgSQL_stmt_while));
new->cmd_type = PLPGSQL_STMT_WHILE;
new->lineno = plpgsql_location_to_lineno(@2);
new->label = $1;
new->cond = $3;
new->body = $4.stmts;
check_labels($1, $4.end_label, $4.end_label_location);
plpgsql_ns_pop();
$$ = (PLpgSQL_stmt *)new;
}
;
loop_body : proc_sect K_END K_LOOP opt_label ';'
{
$$.stmts = $1;
$$.end_label = $4;
$$.end_label_location = @4;
}
;
WHILE statemement
parser
proc_stmt : pl_block ';'
{ $$ = $1; }
| stmt_assign { $$ = $1; }
| stmt_while { $$ = $1; }
....
stmt_while : opt_block_label K_WHILE expr_until_loop loop_body
{
PLpgSQL_stmt_while *new;
new = palloc0(sizeof(PLpgSQL_stmt_while));
new->cmd_type = PLPGSQL_STMT_WHILE;
new->lineno = plpgsql_location_to_lineno(@2);
new->label = $1;
new->cond = $3;
new->body = $4.stmts;
check_labels($1, $4.end_label, $4.end_label_location);
plpgsql_ns_pop();
$$ = (PLpgSQL_stmt *)new;
}
;
loop_body : proc_sect K_END K_LOOP opt_label ';'
{
$$.stmts = $1;
$$.end_label = $4;
$$.end_label_location = @4;
}
;
WHILE statement
executor/main switch
static int
exec_stmt(PLpgSQL_execstate *estate, PLpgSQL_stmt *stmt)
{
PLpgSQL_stmt *save_estmt;
int rc = -1;
save_estmt = estate->err_stmt;
estate->err_stmt = stmt;
/* Let the plugin know that we are about to execute this statement */
if (*plugin_ptr && (*plugin_ptr)->stmt_beg)
((*plugin_ptr)->stmt_beg) (estate, stmt);
CHECK_FOR_INTERRUPTS();
switch ((enum PLpgSQL_stmt_types) stmt->cmd_type)
{
case PLPGSQL_STMT_BLOCK:
rc = exec_stmt_block(estate, (PLpgSQL_stmt_block *) stmt);
break;
case PLPGSQL_STMT_ASSIGN:
rc = exec_stmt_assign(estate, (PLpgSQL_stmt_assign *) stmt);
break;
case PLPGSQL_STMT_PERFORM:
rc = exec_stmt_perform(estate, (PLpgSQL_stmt_perform *) stmt);
break;
case PLPGSQL_STMT_GETDIAG:
rc = exec_stmt_getdiag(estate, (PLpgSQL_stmt_getdiag *) stmt);
break;
WHILE statement
node handler
static int
exec_stmt_while(PLpgSQL_execstate *estate, PLpgSQL_stmt_while *stmt)
{
for (;;)
{
int rc;
bool value;
bool isnull;
value = exec_eval_boolean(estate, stmt->cond, &isnull);
exec_eval_cleanup(estate);
if (isnull || !value)
break;
rc = exec_stmts(estate, stmt->body);
switch (rc)
{
case PLPGSQL_RC_OK:
break;
case PLPGSQL_RC_EXIT:
return PLPGSQL_RC_OK;
case PLPGSQL_RC_CONTINUE:
break;
case PLPGSQL_RC_RETURN:
return rc;
default:
elog(ERROR, "unrecognized rc: %d", rc);
}
}
return PLPGSQL_RC_OK;
}
Known issues
● Collisions of PL/pgSQL and SQL identifiers
– solved in 9.0 (smart parameter placeholder
positioning)
● Suboptimal plan for some queries
– solved in 9.2 (prepared statements optimization)
– in older version - using DYNAMIC SQL
● Late complete check of expressions
– it is feature
● + don't need to solve dependencies
● - some errors are detected only in run-time
– missing columns, wrong identifiers
Identifiers collisions
● dumb algorithm (8.4 and older)
● use placeholder $n everywhere, where is varname in
query string
– collisions cannot be detected (strange errors)
– positioning placeholder on wrong position (strange run-time
errors)
CREATE OR REPLACE FUNCTION foo(a integer)
RETURNS int AS $$
DECLARE x int;
BEGIN
SELECT a FROM mytab WHERE mytab.a = a INTO x;
Identifiers collisions
● smart algorithm (9.0 and higher)
– callback functions from PostgreSQL parser
● p_pre_columnref_hook
● p_post_columnref_hook
– called when PostgreSQL parser process column references - raise error or
return placeholder node
● p_paramref_hook
/*
* plpgsql_parser_setup set up parser hooks for dynamic parameters
*
* Note: this routine, and the hook functions it prepares for, are logically
* part of plpgsql parsing. But they actually run during function execution,
* when we are ready to evaluate a SQL query or expression that has not
* previously been parsed and planned.
*/
void
plpgsql_parser_setup(struct ParseState *pstate, PLpgSQL_expr *expr)
{
pstate->p_pre_columnref_hook = plpgsql_pre_column_ref;
pstate->p_post_columnref_hook = plpgsql_post_column_ref;
pstate->p_paramref_hook = plpgsql_param_ref;
/* no need to use p_coerce_param_hook */
pstate->p_ref_hook_state = (void *) expr;
}
Dumb positioning
DECLARE
a int;
r record;
BEGIN
SELECT x AS a, y AS b FROM tab INTO rec;
/* RESULT */
SELECT x AS $1, y AS b FROM tab INTO rec; --- SYNTAX ERROR
Expression simple
evaluation lifecycle expressions
are evaluated
sign. faster
result is scalar
PREPARE
EVALUATE
NO PLAN
EXPRESSION has
plan
YES
TRY RUN
AS SIMPLE
EXPRESSION
EVAL FULL
QUERY
NO
PLPGSQL
has SESSION
result CACHE
YES (plan)
RETURNS
RESULT
Late IO casting
static bool
exec_eval_boolean(PLpgSQL_execstate *estate, PLpgSQL_expr *expr, bool *isNull)
{
Datum exprdatum; Oid exprtypeid;
exprdatum = exec_eval_expr(estate, expr, isNull, &exprtypeid);
exprdatum = exec_simple_cast_value(estate, exprdatum, exprtypeid, BOOLOID, -1, *isNull);
return DatumGetBool(exprdatum);
}
static Datum
exec_cast_value(PLpgSQL_execstate *estate, Datum value, Oid valtype, Oid reqtype,
FmgrInfo *reqinput, Oid reqtypioparam, int32 reqtypmod, bool isnull)
{
/* If the type of the given value isn't what's requested, convert it. */
if (valtype != reqtype || reqtypmod != -1)
{
MemoryContext oldcontext;
oldcontext = MemoryContextSwitchTo(estate->eval_econtext->ecxt_per_tuple_memory);
if (!isnull)
{
char *extval;
extval = convert_value_to_string(estate, value, valtype);
value = InputFunctionCall(reqinput, extval, reqtypioparam, reqtypmod);
}
else
{
value = InputFunctionCall(reqinput, NULL, reqtypioparam, reqtypmod);
}
MemoryContextSwitchTo(oldcontext);
}
return value;
}
PL/pgSQL expressions
● Reuse PostgreSQL parser/executor
● No redundant code
● Absolutely compatible with PostgreSQL
● Some operations are slow – array update
● Usually fast enough – bottle neck is in query
processing – it is little bit slower than Python
(1M iterations ~ 370 ms, Python ~ 256ms)
Late (IO) casting issue
● IO cast can be slow
● Possible lost of precission
● Different behave than SQL Casting
● It is not solved yet
postgres=# \sf test_assign
CREATE OR REPLACE FUNCTION public.test_assign()
RETURNS void
LANGUAGE plpgsql
AS $function$ declare x int;
BEGIN
x := 9E3/2;
END
$function$
postgres=# select test_assign();
ERROR: invalid input syntax for integer: "4500.0000000000000000"
CONTEXT: PL/pgSQL function test_assign() line 3 at assignment
Late (IO) casting issue
● IO cast can be slow
● Possible lost of precission
● Different behave than SQL Casting
postgres=# \sf test_assign
CREATE OR REPLACE FUNCTION public.test_assign()
RETURNS void
LANGUAGE plpgsql
AS $function$ declare x int;
BEGIN
x := 9E3/2;
END
$function$
postgres=# select test_assign();
ERROR: invalid input syntax for integer: "4500.0000000000000000"
CONTEXT: PL/pgSQL function test_assign() line 3 at assignment
Cached query plans
● Every query, every expression has a execution
plan
● Plans are stored in session cache, created
when query is evaluated first time
● Plans are dropped when related relations are
dropped
Cached query plans
● Every query, every expression has a execution
plan
● Plans are stored in session cache, created
when query is evaluated first time
● Plans are dropped when related relations are
dropped
● Plans are dropped when cost is significantly
different for current parameters (9.2)
Cached plan issue (solved in 9.2)
● Index are used when should not be used
● Index are not used, but should be used
postgres=# \d omega
Table "public.omega"
Column | Type | Modifiers
--------+---------+-----------
a | integer |
Indexes:
"omega_a_idx" btree (a)
postgres=# insert into omega select 1 from generate_series(1,1000000);
INSERT 0 10000
postgres=# insert into omega select 2 from generate_series(1,1000);
INSERT 0 10
Optimization based on heuristic
(blind optimization)
postgres=# prepare x(int) as select count(*) from omega where a = $1;
PREPARE
postgres=# explain execute x(1);
QUERY PLAN
──────────────────────────────────────────────────────────────────────────────────
Aggregate (cost=17808.36..17808.37 rows=1 width=0)
-> Index Scan using omega_a_idx on omega (cost=0.00..16545.86 rows=505000 width=0)
Index Cond: (a = $1)
(3 rows)
postgres=# explain execute x(2);
QUERY PLAN
──────────────────────────────────────────────────────────────────────────────────
Aggregate (cost=17808.36..17808.37 rows=1 width=0)
-> Index Scan using omega_a_idx on omega (cost=0.00..16545.86 rows=505000 width=0)
Index Cond: (a = $1)
(3 rows)
Optimization for real value
(wait for first and recheck)
postgres=# prepare x(int) as select count(*) from omega where a = $1;
PREPARE
postgres=# explain execute x(1);
QUERY PLAN
────────────────────────────────────────────────────────────────────
Aggregate (cost=19085.83..19085.84 rows=1 width=0)
-> Seq Scan on omega (cost=0.00..16586.00 rows=999934 width=0)
Filter: (a = 1)
(3 rows)
postgres=# explain execute x(2);
QUERY PLAN
─────────────────────────────────────────────────────────────────────────────────
Aggregate (cost=318.73..318.74 rows=1 width=0)
-> Index Only Scan using omega_a_idx on omega (cost=0.00..293.57 rows=10066 width=0)
Index Cond: (a = 2)
(3 rows)
Performance tips
● In 99% SQL and built-in function and
functionality will be faster than your code
● FOR statement will be faster than WHILE
– FOR IN int
– FOR IN SELECT
● Minimalist code is usually faster
● PL/pgSQL is perfect language for data
operations (based on SQL), and worst
language for intensive mathematic op
Performance tips - examples
--bad -- bad
DECLARE v varchar; DECLARE s varchar := '';
BEGIN BEGIN
v := 'a'; IF x1 IS NULL THEN
v := v || 'b'; s := s || 'NULL,'
v := v || 'c'; ELSE
RETURN v; s := s || x1;
END; END IF;
--good IF x2 IS NULL THEN
BEGIN s := s || 'NULL, '
RETURN 'a' || 'b' || 'c'; ELSE
END; s := s || x2;
END IF;
...
-- good
DECLARE s varchar;
BEGIN
s := COALESCE(x1 || ',', 'NULL,')
|| COALESCE(x2 || ',', 'NULL,')
Pavel Stěhule
● PostgreSQL lector, consultant
● Now in GoodData performance team
● Some patches to PostgreSQL (PL/pgSQL)
– CONTINUE statement
– EXECUTE USING
– RETURN QUERY
– CASE statement in PL/pgSQL
– RAISE USING ..
– VARIADIC FUNCTIONS, DEFAULT parameters
– FOREACH IN ARRAY
– GET STACKED DIAGNOSTICS
