You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
postgres/contrib/pg_overexplain/expected/pg_overexplain.out

490 lines
21 KiB

-- These tests display internal details that would not be stable under
-- debug_parallel_query, so make sure that option is disabled.
SET debug_parallel_query = off;
-- Make sure that we don't print any JIT-related information, as that
-- would also make results unstable.
SET jit = off;
-- These options do not exist, so these queries should all fail.
EXPLAIN (DEBUFF) SELECT 1;
ERROR: unrecognized EXPLAIN option "debuff"
LINE 1: EXPLAIN (DEBUFF) SELECT 1;
^
EXPLAIN (DEBUG) SELECT 1;
ERROR: unrecognized EXPLAIN option "debug"
LINE 1: EXPLAIN (DEBUG) SELECT 1;
^
EXPLAIN (RANGE_TABLE) SELECT 1;
ERROR: unrecognized EXPLAIN option "range_table"
LINE 1: EXPLAIN (RANGE_TABLE) SELECT 1;
^
-- Load the module that creates the options.
LOAD 'pg_overexplain';
-- The first option still does not exist, but the others do.
EXPLAIN (DEBUFF) SELECT 1;
ERROR: unrecognized EXPLAIN option "debuff"
LINE 1: EXPLAIN (DEBUFF) SELECT 1;
^
EXPLAIN (DEBUG) SELECT 1;
QUERY PLAN
------------------------------------------
Result (cost=0.00..0.01 rows=1 width=4)
Disabled Nodes: 0
Parallel Safe: false
Plan Node ID: 0
PlannedStmt:
Command Type: select
Flags: canSetTag
Subplans Needing Rewind: none
Relation OIDs: none
Executor Parameter Types: none
Parse Location: 16 for 8 bytes
(11 rows)
EXPLAIN (RANGE_TABLE) SELECT 1;
QUERY PLAN
------------------------------------------
Result (cost=0.00..0.01 rows=1 width=4)
RTI 1 (result):
Eref: "*RESULT*" ()
(3 rows)
-- Create a partitioned table.
CREATE TABLE vegetables (id serial, name text, genus text)
PARTITION BY LIST (genus);
CREATE TABLE daucus PARTITION OF vegetables FOR VALUES IN ('daucus');
CREATE TABLE brassica PARTITION OF vegetables FOR VALUES IN ('brassica');
INSERT INTO vegetables (name, genus)
VALUES ('carrot', 'daucus'), ('bok choy', 'brassica'),
('brocooli', 'brassica'), ('cauliflower', 'brassica'),
('cabbage', 'brassica'), ('kohlrabi', 'brassica'),
('rutabaga', 'brassica'), ('turnip', 'brassica');
VACUUM ANALYZE vegetables;
-- We filter relation OIDs out of the test output in order to avoid
-- test instability. This is currently only needed for EXPLAIN (DEBUG), not
-- EXPLAIN (RANGE_TABLE). Also suppress actual row counts, which are not
-- stable (e.g. 1/8 is 0.12 on some buildfarm machines and 0.13 on others).
CREATE FUNCTION explain_filter(text) RETURNS SETOF text
LANGUAGE plpgsql AS
$$
DECLARE
ln text;
BEGIN
FOR ln IN EXECUTE $1
LOOP
ln := regexp_replace(ln, 'Relation OIDs:( \m\d+\M)+',
'Relation OIDs: NNN...', 'g');
ln := regexp_replace(ln, '<Relation-OIDs>( ?\m\d+\M)+</Relation-OIDs>',
'<Relation-OIDs>NNN...</Relation-OIDs>', 'g');
ln := regexp_replace(ln, 'actual rows=\d+\.\d+',
'actual rows=N.NN', 'g');
RETURN NEXT ln;
END LOOP;
END;
$$;
-- Test with both options together and an aggregate.
SELECT explain_filter($$
EXPLAIN (DEBUG, RANGE_TABLE, COSTS OFF)
SELECT genus, array_agg(name ORDER BY name) FROM vegetables GROUP BY genus
$$);
explain_filter
-----------------------------------------------------
GroupAggregate
Group Key: vegetables.genus
Disabled Nodes: 0
Parallel Safe: true
Plan Node ID: 0
-> Sort
Sort Key: vegetables.genus, vegetables.name
Disabled Nodes: 0
Parallel Safe: true
Plan Node ID: 1
-> Append
Disabled Nodes: 0
Parallel Safe: true
Plan Node ID: 2
Append RTIs: 1
-> Seq Scan on brassica vegetables_1
Disabled Nodes: 0
Parallel Safe: true
Plan Node ID: 3
Scan RTI: 3
-> Seq Scan on daucus vegetables_2
Disabled Nodes: 0
Parallel Safe: true
Plan Node ID: 4
Scan RTI: 4
PlannedStmt:
Command Type: select
Flags: canSetTag
Subplans Needing Rewind: none
Relation OIDs: NNN...
Executor Parameter Types: none
Fix regression with location calculation of nested statements The statement location calculated for some nested query cases was wrong when multiple queries are sent as a single string, these being separated by semicolons. As pointed by Sami Imseih, the location calculation was incorrect when the last query of nested statement with multiple queries does **NOT** finish with a semicolon for the last statement. In this case, the statement length tracked by RawStmt is 0, which is equivalent to say that the string should be used until its end. The code previously discarded this case entirely, causing the location to remain at 0, the same as pointing at the beginning of the string. This caused pg_stat_statements to store incorrect query strings. This issue has been introduced in 499edb09741b. I have looked at the diffs generated by pgaudit back then, and noticed the difference generated for this nested query case, but I have missed the point that it was an actual regression with an existing case. A test case is added in pg_stat_statements to provide some coverage, restoring the pre-17 behavior for the calculation of the query locations. Special thanks to David Steele, who, through an analysis of the test diffs generated by pgaudit with the new v18 logic, has poked me about the fact that my original analysis of the matter was wrong. The test output of pg_overexplain is updated to reflect the new logic, as the new locations refer to the beginning of the argument passed to the function explain_filter(). When the module was introduced in 8d5ceb113e3f, which was after 499edb09741b (for the new calculation method), the locations of the test were not actually right: the plan generated for the query string given in input of the function pointed to the top-level query, not the nested one. Reported-by: David Steele <david@pgbackrest.org> Author: Michael Paquier <michael@paquier.xyz> Reviewed-by: Anthonin Bonnefoy <anthonin.bonnefoy@datadoghq.com> Reviewed-by: Jian He <jian.universality@gmail.com> Reviewed-by: Sami Imseih <samimseih@gmail.com> Reviewed-by: David Steele <david@pgbackrest.org> Discussion: https://postgr.es/m/844a3b38-bbf1-4fb2-9fd6-f58c35c09917@pgbackrest.org
7 months ago
Parse Location: 41 to end
RTI 1 (relation, inherited, in-from-clause):
Eref: vegetables (id, name, genus)
Relation: vegetables
Relation Kind: partitioned_table
Relation Lock Mode: AccessShareLock
Permission Info Index: 1
RTI 2 (group):
Eref: "*GROUP*" (genus)
RTI 3 (relation, in-from-clause):
Alias: vegetables (id, name, genus)
Eref: vegetables (id, name, genus)
Relation: brassica
Relation Kind: relation
Relation Lock Mode: AccessShareLock
RTI 4 (relation, in-from-clause):
Alias: vegetables (id, name, genus)
Eref: vegetables (id, name, genus)
Relation: daucus
Relation Kind: relation
Relation Lock Mode: AccessShareLock
Unprunable RTIs: 1 3 4
(53 rows)
-- Test a different output format.
SELECT explain_filter($$
EXPLAIN (DEBUG, RANGE_TABLE, FORMAT XML, COSTS OFF)
SELECT genus, array_agg(name ORDER BY name) FROM vegetables GROUP BY genus
$$);
explain_filter
---------------------------------------------------------------------
<explain xmlns="http://www.postgresql.org/2009/explain"> +
<Query> +
<Plan> +
<Node-Type>Aggregate</Node-Type> +
<Strategy>Sorted</Strategy> +
<Partial-Mode>Simple</Partial-Mode> +
<Parallel-Aware>false</Parallel-Aware> +
<Async-Capable>false</Async-Capable> +
<Disabled>false</Disabled> +
<Group-Key> +
<Item>vegetables.genus</Item> +
</Group-Key> +
<Disabled-Nodes>0</Disabled-Nodes> +
<Parallel-Safe>true</Parallel-Safe> +
<Plan-Node-ID>0</Plan-Node-ID> +
<extParam>none</extParam> +
<allParam>none</allParam> +
<Plans> +
<Plan> +
<Node-Type>Sort</Node-Type> +
<Parent-Relationship>Outer</Parent-Relationship> +
<Parallel-Aware>false</Parallel-Aware> +
<Async-Capable>false</Async-Capable> +
<Disabled>false</Disabled> +
<Sort-Key> +
<Item>vegetables.genus</Item> +
<Item>vegetables.name</Item> +
</Sort-Key> +
<Disabled-Nodes>0</Disabled-Nodes> +
<Parallel-Safe>true</Parallel-Safe> +
<Plan-Node-ID>1</Plan-Node-ID> +
<extParam>none</extParam> +
<allParam>none</allParam> +
<Plans> +
<Plan> +
<Node-Type>Append</Node-Type> +
<Parent-Relationship>Outer</Parent-Relationship> +
<Parallel-Aware>false</Parallel-Aware> +
<Async-Capable>false</Async-Capable> +
<Disabled>false</Disabled> +
<Disabled-Nodes>0</Disabled-Nodes> +
<Parallel-Safe>true</Parallel-Safe> +
<Plan-Node-ID>2</Plan-Node-ID> +
<extParam>none</extParam> +
<allParam>none</allParam> +
<Append-RTIs>1</Append-RTIs> +
<Subplans-Removed>0</Subplans-Removed> +
<Plans> +
<Plan> +
<Node-Type>Seq Scan</Node-Type> +
<Parent-Relationship>Member</Parent-Relationship>+
<Parallel-Aware>false</Parallel-Aware> +
<Async-Capable>false</Async-Capable> +
<Relation-Name>brassica</Relation-Name> +
<Alias>vegetables_1</Alias> +
<Disabled>false</Disabled> +
<Disabled-Nodes>0</Disabled-Nodes> +
<Parallel-Safe>true</Parallel-Safe> +
<Plan-Node-ID>3</Plan-Node-ID> +
<extParam>none</extParam> +
<allParam>none</allParam> +
<Scan-RTI>3</Scan-RTI> +
</Plan> +
<Plan> +
<Node-Type>Seq Scan</Node-Type> +
<Parent-Relationship>Member</Parent-Relationship>+
<Parallel-Aware>false</Parallel-Aware> +
<Async-Capable>false</Async-Capable> +
<Relation-Name>daucus</Relation-Name> +
<Alias>vegetables_2</Alias> +
<Disabled>false</Disabled> +
<Disabled-Nodes>0</Disabled-Nodes> +
<Parallel-Safe>true</Parallel-Safe> +
<Plan-Node-ID>4</Plan-Node-ID> +
<extParam>none</extParam> +
<allParam>none</allParam> +
<Scan-RTI>4</Scan-RTI> +
</Plan> +
</Plans> +
</Plan> +
</Plans> +
</Plan> +
</Plans> +
</Plan> +
<PlannedStmt> +
<Command-Type>select</Command-Type> +
<Flags>canSetTag</Flags> +
<Subplans-Needing-Rewind>none</Subplans-Needing-Rewind> +
<Relation-OIDs>NNN...</Relation-OIDs> +
<Executor-Parameter-Types>none</Executor-Parameter-Types> +
Fix regression with location calculation of nested statements The statement location calculated for some nested query cases was wrong when multiple queries are sent as a single string, these being separated by semicolons. As pointed by Sami Imseih, the location calculation was incorrect when the last query of nested statement with multiple queries does **NOT** finish with a semicolon for the last statement. In this case, the statement length tracked by RawStmt is 0, which is equivalent to say that the string should be used until its end. The code previously discarded this case entirely, causing the location to remain at 0, the same as pointing at the beginning of the string. This caused pg_stat_statements to store incorrect query strings. This issue has been introduced in 499edb09741b. I have looked at the diffs generated by pgaudit back then, and noticed the difference generated for this nested query case, but I have missed the point that it was an actual regression with an existing case. A test case is added in pg_stat_statements to provide some coverage, restoring the pre-17 behavior for the calculation of the query locations. Special thanks to David Steele, who, through an analysis of the test diffs generated by pgaudit with the new v18 logic, has poked me about the fact that my original analysis of the matter was wrong. The test output of pg_overexplain is updated to reflect the new logic, as the new locations refer to the beginning of the argument passed to the function explain_filter(). When the module was introduced in 8d5ceb113e3f, which was after 499edb09741b (for the new calculation method), the locations of the test were not actually right: the plan generated for the query string given in input of the function pointed to the top-level query, not the nested one. Reported-by: David Steele <david@pgbackrest.org> Author: Michael Paquier <michael@paquier.xyz> Reviewed-by: Anthonin Bonnefoy <anthonin.bonnefoy@datadoghq.com> Reviewed-by: Jian He <jian.universality@gmail.com> Reviewed-by: Sami Imseih <samimseih@gmail.com> Reviewed-by: David Steele <david@pgbackrest.org> Discussion: https://postgr.es/m/844a3b38-bbf1-4fb2-9fd6-f58c35c09917@pgbackrest.org
7 months ago
<Parse-Location>53 to end</Parse-Location> +
</PlannedStmt> +
<Range-Table> +
<Range-Table-Entry> +
<RTI>1</RTI> +
<Kind>relation</Kind> +
<Inherited>true</Inherited> +
<In-From-Clause>true</In-From-Clause> +
<Eref>vegetables (id, name, genus)</Eref> +
<Relation>vegetables</Relation> +
<Relation-Kind>partitioned_table</Relation-Kind> +
<Relation-Lock-Mode>AccessShareLock</Relation-Lock-Mode> +
<Permission-Info-Index>1</Permission-Info-Index> +
<Security-Barrier>false</Security-Barrier> +
<Lateral>false</Lateral> +
</Range-Table-Entry> +
<Range-Table-Entry> +
<RTI>2</RTI> +
<Kind>group</Kind> +
<Inherited>false</Inherited> +
<In-From-Clause>false</In-From-Clause> +
<Eref>"*GROUP*" (genus)</Eref> +
<Security-Barrier>false</Security-Barrier> +
<Lateral>false</Lateral> +
</Range-Table-Entry> +
<Range-Table-Entry> +
<RTI>3</RTI> +
<Kind>relation</Kind> +
<Inherited>false</Inherited> +
<In-From-Clause>true</In-From-Clause> +
<Alias>vegetables (id, name, genus)</Alias> +
<Eref>vegetables (id, name, genus)</Eref> +
<Relation>brassica</Relation> +
<Relation-Kind>relation</Relation-Kind> +
<Relation-Lock-Mode>AccessShareLock</Relation-Lock-Mode> +
<Security-Barrier>false</Security-Barrier> +
<Lateral>false</Lateral> +
</Range-Table-Entry> +
<Range-Table-Entry> +
<RTI>4</RTI> +
<Kind>relation</Kind> +
<Inherited>false</Inherited> +
<In-From-Clause>true</In-From-Clause> +
<Alias>vegetables (id, name, genus)</Alias> +
<Eref>vegetables (id, name, genus)</Eref> +
<Relation>daucus</Relation> +
<Relation-Kind>relation</Relation-Kind> +
<Relation-Lock-Mode>AccessShareLock</Relation-Lock-Mode> +
<Security-Barrier>false</Security-Barrier> +
<Lateral>false</Lateral> +
</Range-Table-Entry> +
<Unprunable-RTIs>1 3 4</Unprunable-RTIs> +
<Result-RTIs>none</Result-RTIs> +
</Range-Table> +
</Query> +
</explain>
(1 row)
-- Test just the DEBUG option. Verify that it shows information about
-- disabled nodes, parallel safety, and the parallelModeNeeded flag.
SET enable_seqscan = false;
SET debug_parallel_query = true;
SELECT explain_filter($$
EXPLAIN (DEBUG, COSTS OFF)
SELECT genus, array_agg(name ORDER BY name) FROM vegetables GROUP BY genus
$$);
explain_filter
-----------------------------------------------------------
Gather
Workers Planned: 1
Single Copy: true
Disabled Nodes: 0
Parallel Safe: false
Plan Node ID: 0
-> GroupAggregate
Group Key: vegetables.genus
Disabled Nodes: 2
Parallel Safe: true
Plan Node ID: 1
-> Sort
Sort Key: vegetables.genus, vegetables.name
Disabled Nodes: 2
Parallel Safe: true
Plan Node ID: 2
-> Append
Disabled Nodes: 2
Parallel Safe: true
Plan Node ID: 3
-> Seq Scan on brassica vegetables_1
Disabled: true
Disabled Nodes: 1
Parallel Safe: true
Plan Node ID: 4
-> Seq Scan on daucus vegetables_2
Disabled: true
Disabled Nodes: 1
Parallel Safe: true
Plan Node ID: 5
PlannedStmt:
Command Type: select
Flags: canSetTag, parallelModeNeeded
Subplans Needing Rewind: none
Relation OIDs: NNN...
Executor Parameter Types: none
Fix regression with location calculation of nested statements The statement location calculated for some nested query cases was wrong when multiple queries are sent as a single string, these being separated by semicolons. As pointed by Sami Imseih, the location calculation was incorrect when the last query of nested statement with multiple queries does **NOT** finish with a semicolon for the last statement. In this case, the statement length tracked by RawStmt is 0, which is equivalent to say that the string should be used until its end. The code previously discarded this case entirely, causing the location to remain at 0, the same as pointing at the beginning of the string. This caused pg_stat_statements to store incorrect query strings. This issue has been introduced in 499edb09741b. I have looked at the diffs generated by pgaudit back then, and noticed the difference generated for this nested query case, but I have missed the point that it was an actual regression with an existing case. A test case is added in pg_stat_statements to provide some coverage, restoring the pre-17 behavior for the calculation of the query locations. Special thanks to David Steele, who, through an analysis of the test diffs generated by pgaudit with the new v18 logic, has poked me about the fact that my original analysis of the matter was wrong. The test output of pg_overexplain is updated to reflect the new logic, as the new locations refer to the beginning of the argument passed to the function explain_filter(). When the module was introduced in 8d5ceb113e3f, which was after 499edb09741b (for the new calculation method), the locations of the test were not actually right: the plan generated for the query string given in input of the function pointed to the top-level query, not the nested one. Reported-by: David Steele <david@pgbackrest.org> Author: Michael Paquier <michael@paquier.xyz> Reviewed-by: Anthonin Bonnefoy <anthonin.bonnefoy@datadoghq.com> Reviewed-by: Jian He <jian.universality@gmail.com> Reviewed-by: Sami Imseih <samimseih@gmail.com> Reviewed-by: David Steele <david@pgbackrest.org> Discussion: https://postgr.es/m/844a3b38-bbf1-4fb2-9fd6-f58c35c09917@pgbackrest.org
7 months ago
Parse Location: 28 to end
(37 rows)
SET debug_parallel_query = false;
RESET enable_seqscan;
-- Test the DEBUG option with a non-SELECT query, and also verify that the
-- hasReturning flag is shown.
SELECT explain_filter($$
EXPLAIN (DEBUG, COSTS OFF)
INSERT INTO vegetables (name, genus)
VALUES ('Brotero''s carrot', 'brassica') RETURNING id
$$);
explain_filter
----------------------------------
Insert on vegetables
Disabled Nodes: 0
Parallel Safe: false
Plan Node ID: 0
-> Result
Disabled Nodes: 0
Parallel Safe: false
Plan Node ID: 1
PlannedStmt:
Command Type: insert
Flags: hasReturning, canSetTag
Subplans Needing Rewind: none
Relation OIDs: NNN...
Executor Parameter Types: 0
Fix regression with location calculation of nested statements The statement location calculated for some nested query cases was wrong when multiple queries are sent as a single string, these being separated by semicolons. As pointed by Sami Imseih, the location calculation was incorrect when the last query of nested statement with multiple queries does **NOT** finish with a semicolon for the last statement. In this case, the statement length tracked by RawStmt is 0, which is equivalent to say that the string should be used until its end. The code previously discarded this case entirely, causing the location to remain at 0, the same as pointing at the beginning of the string. This caused pg_stat_statements to store incorrect query strings. This issue has been introduced in 499edb09741b. I have looked at the diffs generated by pgaudit back then, and noticed the difference generated for this nested query case, but I have missed the point that it was an actual regression with an existing case. A test case is added in pg_stat_statements to provide some coverage, restoring the pre-17 behavior for the calculation of the query locations. Special thanks to David Steele, who, through an analysis of the test diffs generated by pgaudit with the new v18 logic, has poked me about the fact that my original analysis of the matter was wrong. The test output of pg_overexplain is updated to reflect the new logic, as the new locations refer to the beginning of the argument passed to the function explain_filter(). When the module was introduced in 8d5ceb113e3f, which was after 499edb09741b (for the new calculation method), the locations of the test were not actually right: the plan generated for the query string given in input of the function pointed to the top-level query, not the nested one. Reported-by: David Steele <david@pgbackrest.org> Author: Michael Paquier <michael@paquier.xyz> Reviewed-by: Anthonin Bonnefoy <anthonin.bonnefoy@datadoghq.com> Reviewed-by: Jian He <jian.universality@gmail.com> Reviewed-by: Sami Imseih <samimseih@gmail.com> Reviewed-by: David Steele <david@pgbackrest.org> Discussion: https://postgr.es/m/844a3b38-bbf1-4fb2-9fd6-f58c35c09917@pgbackrest.org
7 months ago
Parse Location: 28 to end
(15 rows)
-- Create an index, and then attempt to force a nested loop with inner index
-- scan so that we can see parameter-related information. Also, let's try
-- actually running the query, but try to suppress potentially variable output.
CREATE INDEX ON vegetables (id);
ANALYZE vegetables;
SET enable_hashjoin = false;
SET enable_material = false;
SET enable_mergejoin = false;
SET enable_seqscan = false;
SELECT explain_filter($$
EXPLAIN (BUFFERS OFF, COSTS OFF, SUMMARY OFF, TIMING OFF, ANALYZE, DEBUG)
SELECT * FROM vegetables v1, vegetables v2 WHERE v1.id = v2.id;
$$);
explain_filter
------------------------------------------------------------------------------------------
Nested Loop (actual rows=N.NN loops=1)
Disabled Nodes: 0
Parallel Safe: true
Plan Node ID: 0
-> Append (actual rows=N.NN loops=1)
Disabled Nodes: 0
Parallel Safe: true
Plan Node ID: 1
-> Index Scan using brassica_id_idx on brassica v1_1 (actual rows=N.NN loops=1)
Index Searches: 1
Disabled Nodes: 0
Parallel Safe: true
Plan Node ID: 2
-> Index Scan using daucus_id_idx on daucus v1_2 (actual rows=N.NN loops=1)
Index Searches: 1
Disabled Nodes: 0
Parallel Safe: true
Plan Node ID: 3
-> Append (actual rows=N.NN loops=8)
Disabled Nodes: 0
Parallel Safe: true
Plan Node ID: 4
extParam: 0
allParam: 0
-> Index Scan using brassica_id_idx on brassica v2_1 (actual rows=N.NN loops=8)
Index Cond: (id = v1.id)
Index Searches: 8
Disabled Nodes: 0
Parallel Safe: true
Plan Node ID: 5
extParam: 0
allParam: 0
-> Index Scan using daucus_id_idx on daucus v2_2 (actual rows=N.NN loops=8)
Index Cond: (id = v1.id)
Index Searches: 8
Disabled Nodes: 0
Parallel Safe: true
Plan Node ID: 6
extParam: 0
allParam: 0
PlannedStmt:
Command Type: select
Flags: canSetTag
Subplans Needing Rewind: none
Relation OIDs: NNN...
Executor Parameter Types: 23
Parse Location: 75 for 62 bytes
(47 rows)
RESET enable_hashjoin;
RESET enable_material;
RESET enable_mergejoin;
RESET enable_seqscan;
-- Test the RANGE_TABLE option with a case that allows partition pruning.
EXPLAIN (RANGE_TABLE, COSTS OFF)
SELECT * FROM vegetables WHERE genus = 'daucus';
QUERY PLAN
----------------------------------------------
Seq Scan on daucus vegetables
Filter: (genus = 'daucus'::text)
Scan RTI: 2
RTI 1 (relation, inherited, in-from-clause):
Eref: vegetables (id, name, genus)
Relation: vegetables
Relation Kind: partitioned_table
Relation Lock Mode: AccessShareLock
Permission Info Index: 1
RTI 2 (relation, in-from-clause):
Alias: vegetables (id, name, genus)
Eref: vegetables (id, name, genus)
Relation: daucus
Relation Kind: relation
Relation Lock Mode: AccessShareLock
Unprunable RTIs: 1 2
(16 rows)
-- Also test a case that involves a write.
EXPLAIN (RANGE_TABLE, COSTS OFF)
INSERT INTO vegetables (name, genus) VALUES ('broccoflower', 'brassica');
QUERY PLAN
----------------------------------------
Insert on vegetables
Nominal RTI: 1
Exclude Relation RTI: 0
-> Result
RTI 1 (relation):
Eref: vegetables (id, name, genus)
Relation: vegetables
Relation Kind: partitioned_table
Relation Lock Mode: RowExclusiveLock
Permission Info Index: 1
RTI 2 (result):
Eref: "*RESULT*" ()
Unprunable RTIs: 1
Result RTIs: 1
(14 rows)