]> granicus.if.org Git - postgresql/commitdiff
Don't allow LIMIT/OFFSET clause within sub-selects to be pushed to workers.
authorAmit Kapila <akapila@postgresql.org>
Fri, 14 Sep 2018 04:06:30 +0000 (09:36 +0530)
committerAmit Kapila <akapila@postgresql.org>
Fri, 14 Sep 2018 04:06:30 +0000 (09:36 +0530)
Allowing sub-select containing LIMIT/OFFSET in workers can lead to
inconsistent results at the top-level as there is no guarantee that the
row order will be fully deterministic.  The fix is to prohibit pushing
LIMIT/OFFSET within sub-selects to workers.

Reported-by: Andrew Fletcher
Bug: 15324
Author: Amit Kapila
Reviewed-by: Dilip Kumar
Backpatch-through: 9.6
Discussion: https://postgr.es/m/153417684333.10284.11356259990921828616@wrigleys.postgresql.org

src/backend/optimizer/path/allpaths.c
src/backend/optimizer/plan/planner.c
src/include/optimizer/planner.h
src/test/regress/expected/select_parallel.out
src/test/regress/sql/select_parallel.sql

index 5db1688bf0462df97fe27c4655febd1ed2e52f3a..5f74d3b36d9c7c28a4db08d15d812a3b9b203314 100644 (file)
@@ -620,7 +620,20 @@ set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel,
                         * the SubqueryScanPath as not parallel-safe.  (Note that
                         * set_subquery_pathlist() might push some of these quals down
                         * into the subquery itself, but that doesn't change anything.)
+                        *
+                        * We can't push sub-select containing LIMIT/OFFSET to workers as
+                        * there is no guarantee that the row order will be fully
+                        * deterministic, and applying LIMIT/OFFSET will lead to
+                        * inconsistent results at the top-level.  (In some cases, where
+                        * the result is ordered, we could relax this restriction.  But it
+                        * doesn't currently seem worth expending extra effort to do so.)
                         */
+                       {
+                               Query      *subquery = castNode(Query, rte->subquery);
+
+                               if (limit_needed(subquery))
+                                       return;
+                       }
                        break;
 
                case RTE_JOIN:
index 96bf0601a8eb65f2699ec2a5fad4d5f7df8d8d35..e589471fee8b9b822ef3ed568c4baa0aa0875193 100644 (file)
@@ -123,7 +123,6 @@ static void preprocess_rowmarks(PlannerInfo *root);
 static double preprocess_limit(PlannerInfo *root,
                                 double tuple_fraction,
                                 int64 *offset_est, int64 *count_est);
-static bool limit_needed(Query *parse);
 static void remove_useless_groupby_columns(PlannerInfo *root);
 static List *preprocess_groupclause(PlannerInfo *root, List *force);
 static List *extract_rollup_sets(List *groupingSets);
@@ -2870,7 +2869,7 @@ preprocess_limit(PlannerInfo *root, double tuple_fraction,
  * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
  * in preprocess_limit it's good enough to consider estimated values.
  */
-static bool
+bool
 limit_needed(Query *parse)
 {
        Node       *node;
index c090396e139aa31cf448548a40a2ba889ec63239..3e733b34ed974948f4ed8fe4ca28541e917618db 100644 (file)
@@ -47,6 +47,8 @@ extern bool is_dummy_plan(Plan *plan);
 extern RowMarkType select_rowmark_type(RangeTblEntry *rte,
                                        LockClauseStrength strength);
 
+extern bool limit_needed(Query *parse);
+
 extern void mark_partial_aggref(Aggref *agg, AggSplit aggsplit);
 
 extern Path *get_cheapest_fractional_path(RelOptInfo *rel,
index 13cdd3d5cca36f92d498f08501c62644cce18c2b..26409d39aaaf6d6cfdae1c69dad7e0fdc173bc28 100644 (file)
@@ -984,6 +984,25 @@ explain (costs off, verbose)
                                  Output: b.unique1
 (18 rows)
 
+-- LIMIT/OFFSET within sub-selects can't be pushed to workers.
+explain (costs off)
+  select * from tenk1 a where two in
+    (select two from tenk1 b where stringu1 like '%AAAA' limit 3);
+                          QUERY PLAN                           
+---------------------------------------------------------------
+ Hash Semi Join
+   Hash Cond: (a.two = b.two)
+   ->  Gather
+         Workers Planned: 4
+         ->  Parallel Seq Scan on tenk1 a
+   ->  Hash
+         ->  Limit
+               ->  Gather
+                     Workers Planned: 4
+                     ->  Parallel Seq Scan on tenk1 b
+                           Filter: (stringu1 ~~ '%AAAA'::text)
+(11 rows)
+
 -- to increase the parallel query test coverage
 SAVEPOINT settings;
 SET LOCAL force_parallel_mode = 1;
index 3d1dffe61d29cbdedf6b706ee84f865592cf7781..938c708d18fd84fba86f525af769b6d9f7c4113a 100644 (file)
@@ -368,6 +368,11 @@ explain (costs off, verbose)
     (select unique1, row_number() over() from tenk1 b);
 
 
+-- LIMIT/OFFSET within sub-selects can't be pushed to workers.
+explain (costs off)
+  select * from tenk1 a where two in
+    (select two from tenk1 b where stringu1 like '%AAAA' limit 3);
+
 -- to increase the parallel query test coverage
 SAVEPOINT settings;
 SET LOCAL force_parallel_mode = 1;