Do not execute subplans multiple times with cursors

Before this commit, we let AdaptiveExecutorPreExecutorRun()
to be effective multiple times on every FETCH on cursors.
That does not affect the correctness of the query results,
but adds significant overhead.
pull/4331/head
Onder Kalaci 2020-11-20 10:29:59 +01:00
parent b0ddbbd33a
commit c433c66f2b
6 changed files with 350 additions and 1 deletions

View File

@ -660,6 +660,16 @@ static void SetAttributeInputMetadata(DistributedExecution *execution,
void
AdaptiveExecutorPreExecutorRun(CitusScanState *scanState)
{
if (scanState->finishedPreScan)
{
/*
* Cursors (and hence RETURN QUERY syntax in pl/pgsql functions)
* may trigger AdaptiveExecutorPreExecutorRun() on every fetch
* operation. Though, we should only execute PreScan once.
*/
return;
}
DistributedPlan *distributedPlan = scanState->distributedPlan;
/*
@ -670,6 +680,8 @@ AdaptiveExecutorPreExecutorRun(CitusScanState *scanState)
LockPartitionsForDistributedPlan(distributedPlan);
ExecuteSubPlans(distributedPlan);
scanState->finishedPreScan = true;
}

View File

@ -558,6 +558,9 @@ AdaptiveExecutorCreateScan(CustomScan *scan)
scanState->customScanState.methods = &AdaptiveExecutorCustomExecMethods;
scanState->PreExecScan = &CitusPreExecScan;
scanState->finishedPreScan = false;
scanState->finishedRemoteScan = false;
return (Node *) scanState;
}
@ -578,6 +581,9 @@ NonPushableInsertSelectCreateScan(CustomScan *scan)
scanState->customScanState.methods =
&NonPushableInsertSelectCustomExecMethods;
scanState->finishedPreScan = false;
scanState->finishedRemoteScan = false;
return (Node *) scanState;
}

View File

@ -20,6 +20,7 @@ typedef struct CitusScanState
CustomScanState customScanState; /* underlying custom scan node */
/* function that gets called before postgres starts its execution */
bool finishedPreScan; /* flag to check if the pre scan is finished */
void (*PreExecScan)(struct CitusScanState *scanState);
DistributedPlan *distributedPlan; /* distributed execution plan */

View File

@ -0,0 +1,218 @@
CREATE SCHEMA cursors;
SET search_path TO cursors;
CREATE TABLE distributed_table (key int, value text);
SELECT create_distributed_table('distributed_table', 'key');
create_distributed_table
---------------------------------------------------------------------
(1 row)
-- load some data, but not very small amounts because RETURN QUERY in plpgsql
-- hard codes the cursor fetch to 50 rows on PG 12, though they might increase
-- it sometime in the future, so be mindful
INSERT INTO distributed_table SELECT i % 10, i::text FROM generate_series(0, 1000) i;
CREATE OR REPLACE FUNCTION simple_cursor_on_dist_table(cursor_name refcursor) RETURNS refcursor AS '
BEGIN
OPEN $1 FOR SELECT DISTINCT key FROM distributed_table ORDER BY 1;
RETURN $1;
END;
' LANGUAGE plpgsql;
CREATE OR REPLACE FUNCTION cursor_with_intermediate_result_on_dist_table(cursor_name refcursor) RETURNS refcursor AS '
BEGIN
OPEN $1 FOR
WITH cte_1 AS (SELECT * FROM distributed_table OFFSET 0)
SELECT DISTINCT key FROM distributed_table WHERE value in (SELECT value FROM cte_1) ORDER BY 1;
RETURN $1;
END;
' LANGUAGE plpgsql;
CREATE OR REPLACE FUNCTION cursor_with_intermediate_result_on_dist_table_with_param(cursor_name refcursor, filter text) RETURNS refcursor AS '
BEGIN
OPEN $1 FOR
WITH cte_1 AS (SELECT * FROM distributed_table WHERE value < $2 OFFSET 0)
SELECT DISTINCT key FROM distributed_table WHERE value in (SELECT value FROM cte_1) ORDER BY 1;
RETURN $1;
END;
' LANGUAGE plpgsql;
-- pretty basic query with cursors
-- Citus should plan/execute once and pull
-- the results to coordinator, then serve it
-- from the coordinator
BEGIN;
SELECT simple_cursor_on_dist_table('cursor_1');
simple_cursor_on_dist_table
---------------------------------------------------------------------
cursor_1
(1 row)
SET LOCAL citus.log_intermediate_results TO ON;
SET LOCAL client_min_messages TO DEBUG1;
FETCH 5 IN cursor_1;
key
---------------------------------------------------------------------
0
1
2
3
4
(5 rows)
FETCH 50 IN cursor_1;
key
---------------------------------------------------------------------
5
6
7
8
9
(5 rows)
FETCH ALL IN cursor_1;
key
---------------------------------------------------------------------
(0 rows)
COMMIT;
BEGIN;
SELECT cursor_with_intermediate_result_on_dist_table('cursor_1');
cursor_with_intermediate_result_on_dist_table
---------------------------------------------------------------------
cursor_1
(1 row)
-- multiple FETCH commands should not trigger re-running the subplans
SET LOCAL citus.log_intermediate_results TO ON;
SET LOCAL client_min_messages TO DEBUG1;
FETCH 5 IN cursor_1;
DEBUG: Subplan XXX_1 will be sent to localhost:xxxxx
DEBUG: Subplan XXX_1 will be sent to localhost:xxxxx
key
---------------------------------------------------------------------
0
1
2
3
4
(5 rows)
FETCH 1 IN cursor_1;
key
---------------------------------------------------------------------
5
(1 row)
FETCH ALL IN cursor_1;
key
---------------------------------------------------------------------
6
7
8
9
(4 rows)
FETCH 5 IN cursor_1;
key
---------------------------------------------------------------------
(0 rows)
COMMIT;
BEGIN;
SELECT cursor_with_intermediate_result_on_dist_table_with_param('cursor_1', '600');
cursor_with_intermediate_result_on_dist_table_with_param
---------------------------------------------------------------------
cursor_1
(1 row)
-- multiple FETCH commands should not trigger re-running the subplans
-- also test with parameters
SET LOCAL citus.log_intermediate_results TO ON;
SET LOCAL client_min_messages TO DEBUG1;
FETCH 1 IN cursor_1;
DEBUG: Subplan XXX_1 will be sent to localhost:xxxxx
DEBUG: Subplan XXX_1 will be sent to localhost:xxxxx
key
---------------------------------------------------------------------
0
(1 row)
FETCH 1 IN cursor_1;
key
---------------------------------------------------------------------
1
(1 row)
FETCH 1 IN cursor_1;
key
---------------------------------------------------------------------
2
(1 row)
FETCH 1 IN cursor_1;
key
---------------------------------------------------------------------
3
(1 row)
FETCH 1 IN cursor_1;
key
---------------------------------------------------------------------
4
(1 row)
FETCH 1 IN cursor_1;
key
---------------------------------------------------------------------
5
(1 row)
FETCH ALL IN cursor_1;
key
---------------------------------------------------------------------
6
7
8
9
(4 rows)
COMMIT;
CREATE OR REPLACE FUNCTION value_counter() RETURNS TABLE(counter text) LANGUAGE PLPGSQL AS $function$
BEGIN
return query
WITH cte AS
(SELECT dt.value
FROM distributed_table dt
WHERE dt.value in
(SELECT value
FROM distributed_table p
GROUP BY p.value
HAVING count(*) > 0))
SELECT * FROM cte;
END;
$function$ ;
SET citus.log_intermediate_results TO ON;
SET client_min_messages TO DEBUG1;
\set VERBOSITY terse
SELECT count(*) from (SELECT value_counter()) as foo;
DEBUG: CTE cte is going to be inlined via distributed planning
DEBUG: generating subplan XXX_1 for subquery SELECT value FROM cursors.distributed_table p GROUP BY value HAVING (count(*) OPERATOR(pg_catalog.>) 0)
DEBUG: Plan XXX query after replacing subqueries and CTEs: SELECT value FROM (SELECT dt.value FROM cursors.distributed_table dt WHERE (dt.value OPERATOR(pg_catalog.=) ANY (SELECT intermediate_result.value FROM read_intermediate_result('XXX_1'::text, 'binary'::citus_copy_format) intermediate_result(value text)))) cte
DEBUG: Subplan XXX_1 will be sent to localhost:xxxxx
DEBUG: Subplan XXX_1 will be sent to localhost:xxxxx
count
---------------------------------------------------------------------
1001
(1 row)
BEGIN;
SELECT count(*) from (SELECT value_counter()) as foo;
DEBUG: Subplan XXX_1 will be sent to localhost:xxxxx
DEBUG: Subplan XXX_1 will be sent to localhost:xxxxx
count
---------------------------------------------------------------------
1001
(1 row)
COMMIT;
-- suppress NOTICEs
SET client_min_messages TO ERROR;
DROP SCHEMA cursors CASCADE;

View File

@ -108,7 +108,7 @@ test: ch_bench_subquery_repartition
test: multi_agg_type_conversion multi_count_type_conversion
test: multi_partition_pruning single_hash_repartition_join
test: multi_join_pruning multi_hash_pruning intermediate_result_pruning
test: multi_null_minmax_value_pruning
test: multi_null_minmax_value_pruning cursors
test: multi_query_directory_cleanup
test: multi_task_assignment_policy multi_cross_shard
test: multi_utility_statements

View File

@ -0,0 +1,112 @@
CREATE SCHEMA cursors;
SET search_path TO cursors;
CREATE TABLE distributed_table (key int, value text);
SELECT create_distributed_table('distributed_table', 'key');
-- load some data, but not very small amounts because RETURN QUERY in plpgsql
-- hard codes the cursor fetch to 50 rows on PG 12, though they might increase
-- it sometime in the future, so be mindful
INSERT INTO distributed_table SELECT i % 10, i::text FROM generate_series(0, 1000) i;
CREATE OR REPLACE FUNCTION simple_cursor_on_dist_table(cursor_name refcursor) RETURNS refcursor AS '
BEGIN
OPEN $1 FOR SELECT DISTINCT key FROM distributed_table ORDER BY 1;
RETURN $1;
END;
' LANGUAGE plpgsql;
CREATE OR REPLACE FUNCTION cursor_with_intermediate_result_on_dist_table(cursor_name refcursor) RETURNS refcursor AS '
BEGIN
OPEN $1 FOR
WITH cte_1 AS (SELECT * FROM distributed_table OFFSET 0)
SELECT DISTINCT key FROM distributed_table WHERE value in (SELECT value FROM cte_1) ORDER BY 1;
RETURN $1;
END;
' LANGUAGE plpgsql;
CREATE OR REPLACE FUNCTION cursor_with_intermediate_result_on_dist_table_with_param(cursor_name refcursor, filter text) RETURNS refcursor AS '
BEGIN
OPEN $1 FOR
WITH cte_1 AS (SELECT * FROM distributed_table WHERE value < $2 OFFSET 0)
SELECT DISTINCT key FROM distributed_table WHERE value in (SELECT value FROM cte_1) ORDER BY 1;
RETURN $1;
END;
' LANGUAGE plpgsql;
-- pretty basic query with cursors
-- Citus should plan/execute once and pull
-- the results to coordinator, then serve it
-- from the coordinator
BEGIN;
SELECT simple_cursor_on_dist_table('cursor_1');
SET LOCAL citus.log_intermediate_results TO ON;
SET LOCAL client_min_messages TO DEBUG1;
FETCH 5 IN cursor_1;
FETCH 50 IN cursor_1;
FETCH ALL IN cursor_1;
COMMIT;
BEGIN;
SELECT cursor_with_intermediate_result_on_dist_table('cursor_1');
-- multiple FETCH commands should not trigger re-running the subplans
SET LOCAL citus.log_intermediate_results TO ON;
SET LOCAL client_min_messages TO DEBUG1;
FETCH 5 IN cursor_1;
FETCH 1 IN cursor_1;
FETCH ALL IN cursor_1;
FETCH 5 IN cursor_1;
COMMIT;
BEGIN;
SELECT cursor_with_intermediate_result_on_dist_table_with_param('cursor_1', '600');
-- multiple FETCH commands should not trigger re-running the subplans
-- also test with parameters
SET LOCAL citus.log_intermediate_results TO ON;
SET LOCAL client_min_messages TO DEBUG1;
FETCH 1 IN cursor_1;
FETCH 1 IN cursor_1;
FETCH 1 IN cursor_1;
FETCH 1 IN cursor_1;
FETCH 1 IN cursor_1;
FETCH 1 IN cursor_1;
FETCH ALL IN cursor_1;
COMMIT;
CREATE OR REPLACE FUNCTION value_counter() RETURNS TABLE(counter text) LANGUAGE PLPGSQL AS $function$
BEGIN
return query
WITH cte AS
(SELECT dt.value
FROM distributed_table dt
WHERE dt.value in
(SELECT value
FROM distributed_table p
GROUP BY p.value
HAVING count(*) > 0))
SELECT * FROM cte;
END;
$function$ ;
SET citus.log_intermediate_results TO ON;
SET client_min_messages TO DEBUG1;
\set VERBOSITY terse
SELECT count(*) from (SELECT value_counter()) as foo;
BEGIN;
SELECT count(*) from (SELECT value_counter()) as foo;
COMMIT;
-- suppress NOTICEs
SET client_min_messages TO ERROR;
DROP SCHEMA cursors CASCADE;