From 6f94531b0cf83b1401aca31c71412e633c566a63 Mon Sep 17 00:00:00 2001 From: Tom Lane Date: Thu, 28 Jan 2021 13:41:55 -0500 Subject: [PATCH] Fix hash partition pruning with asymmetric partition sets. MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit perform_pruning_combine_step() was not taught about the number of partition indexes used in hash partitioning; more embarrassingly, get_matching_hash_bounds() also had it wrong. These errors are masked in the common case where all the partitions have the same modulus and no partition is missing. However, with missing or unequal-size partitions, we could erroneously prune some partitions that need to be scanned, leading to silently wrong query answers. While a minimal-footprint fix for this could be to export get_partition_bound_num_indexes and make the incorrect functions use it, I'm of the opinion that that function should never have existed in the first place. It's not reasonable data structure design that PartitionBoundInfoData lacks any explicit record of the length of its indexes[] array. Perhaps that was all right when it could always be assumed equal to ndatums, but something should have been done about it as soon as that stopped being true. Putting in an explicit "nindexes" field makes both partition_bounds_equal() and partition_bounds_copy() simpler, safer, and faster than before, and removes explicit knowledge of the number-of-partition-indexes rules from some other places too. This change also makes get_hash_partition_greatest_modulus obsolete. I left that in place in case any external code uses it, but no core code does anymore. Per bug #16840 from Michał Albrycht. Back-patch to v11 where the hash partitioning code came in. (In the back branches, add the new field at the end of PartitionBoundInfoData to minimize ABI risks.) Discussion: https://postgr.es/m/16840-571a22976f829ad4@postgresql.org --- src/backend/executor/execPartition.c | 4 +- src/backend/partitioning/partbounds.c | 111 +++++------------- src/backend/partitioning/partprune.c | 30 ++--- src/backend/utils/cache/partcache.c | 3 + src/include/partitioning/partbounds.h | 29 +++-- src/test/regress/expected/partition_prune.out | 62 +++++++--- src/test/regress/sql/partition_prune.sql | 27 +++-- 7 files changed, 125 insertions(+), 141 deletions(-) diff --git a/src/backend/executor/execPartition.c b/src/backend/executor/execPartition.c index 306f695de1..dc58fd0a82 100644 --- a/src/backend/executor/execPartition.c +++ b/src/backend/executor/execPartition.c @@ -1159,15 +1159,13 @@ get_partition_for_tuple(Relation relation, Datum *values, bool *isnull) { case PARTITION_STRATEGY_HASH: { - int greatest_modulus; uint64 rowHash; - greatest_modulus = get_hash_partition_greatest_modulus(boundinfo); rowHash = compute_partition_hash_value(key->partnatts, key->partsupfunc, values, isnull); - part_index = boundinfo->indexes[rowHash % greatest_modulus]; + part_index = boundinfo->indexes[rowHash % boundinfo->nindexes]; } break; diff --git a/src/backend/partitioning/partbounds.c b/src/backend/partitioning/partbounds.c index e6f3068d66..6103b96db6 100644 --- a/src/backend/partitioning/partbounds.c +++ b/src/backend/partitioning/partbounds.c @@ -36,7 +36,6 @@ #include "utils/ruleutils.h" #include "utils/syscache.h" -static int get_partition_bound_num_indexes(PartitionBoundInfo b); static Expr *make_partition_op_expr(PartitionKey key, int keynum, uint16 strategy, Expr *arg1, Expr *arg2); static Oid get_partition_operator(PartitionKey key, int col, @@ -112,45 +111,41 @@ partition_bounds_equal(int partnatts, int16 *parttyplen, bool *parttypbyval, if (b1->ndatums != b2->ndatums) return false; + if (b1->nindexes != b2->nindexes) + return false; + if (b1->null_index != b2->null_index) return false; if (b1->default_index != b2->default_index) return false; + /* For all partition strategies, the indexes[] arrays have to match */ + for (i = 0; i < b1->nindexes; i++) + { + if (b1->indexes[i] != b2->indexes[i]) + return false; + } + + /* Finally, compare the datums[] arrays */ if (b1->strategy == PARTITION_STRATEGY_HASH) { - int greatest_modulus = get_hash_partition_greatest_modulus(b1); - - /* - * If two hash partitioned tables have different greatest moduli, - * their partition schemes don't match. - */ - if (greatest_modulus != get_hash_partition_greatest_modulus(b2)) - return false; - /* * We arrange the partitions in the ascending order of their moduli * and remainders. Also every modulus is factor of next larger * modulus. Therefore we can safely store index of a given partition * in indexes array at remainder of that partition. Also entries at * (remainder + N * modulus) positions in indexes array are all same - * for (modulus, remainder) specification for any partition. Thus - * datums array from both the given bounds are same, if and only if - * their indexes array will be same. So, it suffices to compare - * indexes array. - */ - for (i = 0; i < greatest_modulus; i++) - if (b1->indexes[i] != b2->indexes[i]) - return false; - -#ifdef USE_ASSERT_CHECKING - - /* - * Nonetheless make sure that the bounds are indeed same when the + * for (modulus, remainder) specification for any partition. Thus the + * datums arrays from the given bounds are the same, if and only if + * their indexes arrays are the same. So, it suffices to compare the + * indexes arrays. + * + * Nonetheless make sure that the bounds are indeed the same when the * indexes match. Hash partition bound stores modulus and remainder * at b1->datums[i][0] and b1->datums[i][1] position respectively. */ +#ifdef USE_ASSERT_CHECKING for (i = 0; i < b1->ndatums; i++) Assert((b1->datums[i][0] == b2->datums[i][0] && b1->datums[i][1] == b2->datums[i][1])); @@ -196,15 +191,7 @@ partition_bounds_equal(int partnatts, int16 *parttyplen, bool *parttypbyval, parttypbyval[j], parttyplen[j])) return false; } - - if (b1->indexes[i] != b2->indexes[i]) - return false; } - - /* There are ndatums+1 indexes in case of range partitions */ - if (b1->strategy == PARTITION_STRATEGY_RANGE && - b1->indexes[i] != b2->indexes[i]) - return false; } return true; } @@ -220,17 +207,16 @@ partition_bounds_copy(PartitionBoundInfo src, PartitionBoundInfo dest; int i; int ndatums; + int nindexes; int partnatts; - int num_indexes; dest = (PartitionBoundInfo) palloc(sizeof(PartitionBoundInfoData)); dest->strategy = src->strategy; ndatums = dest->ndatums = src->ndatums; + nindexes = dest->nindexes = src->nindexes; partnatts = key->partnatts; - num_indexes = get_partition_bound_num_indexes(src); - /* List partitioned tables have only a single partition key. */ Assert(key->strategy != PARTITION_STRATEGY_LIST || partnatts == 1); @@ -288,8 +274,8 @@ partition_bounds_copy(PartitionBoundInfo src, } } - dest->indexes = (int *) palloc(sizeof(int) * num_indexes); - memcpy(dest->indexes, src->indexes, sizeof(int) * num_indexes); + dest->indexes = (int *) palloc(sizeof(int) * nindexes); + memcpy(dest->indexes, src->indexes, sizeof(int) * nindexes); dest->null_index = src->null_index; dest->default_index = src->default_index; @@ -389,7 +375,7 @@ check_new_partition_bound(char *relname, Relation parent, (errcode(ERRCODE_INVALID_OBJECT_DEFINITION), errmsg("every hash partition modulus must be a factor of the next larger modulus"))); - greatest_modulus = get_hash_partition_greatest_modulus(boundinfo); + greatest_modulus = boundinfo->nindexes; remainder = spec->remainder; /* @@ -756,18 +742,15 @@ check_default_partition_contents(Relation parent, Relation default_rel, /* * get_hash_partition_greatest_modulus * - * Returns the greatest modulus of the hash partition bound. The greatest - * modulus will be at the end of the datums array because hash partitions are - * arranged in the ascending order of their moduli and remainders. + * Returns the greatest modulus of the hash partition bound. + * This is no longer used in the core code, but we keep it around + * in case external modules are using it. */ int get_hash_partition_greatest_modulus(PartitionBoundInfo bound) { Assert(bound && bound->strategy == PARTITION_STRATEGY_HASH); - Assert(bound->datums && bound->ndatums > 0); - Assert(DatumGetInt32(bound->datums[bound->ndatums - 1][0]) > 0); - - return DatumGetInt32(bound->datums[bound->ndatums - 1][0]); + return bound->nindexes; } /* @@ -1115,46 +1098,6 @@ partition_hash_bsearch(PartitionBoundInfo boundinfo, return lo; } -/* - * get_partition_bound_num_indexes - * - * Returns the number of the entries in the partition bound indexes array. - */ -static int -get_partition_bound_num_indexes(PartitionBoundInfo bound) -{ - int num_indexes; - - Assert(bound); - - switch (bound->strategy) - { - case PARTITION_STRATEGY_HASH: - - /* - * The number of the entries in the indexes array is same as the - * greatest modulus. - */ - num_indexes = get_hash_partition_greatest_modulus(bound); - break; - - case PARTITION_STRATEGY_LIST: - num_indexes = bound->ndatums; - break; - - case PARTITION_STRATEGY_RANGE: - /* Range partitioned table has an extra index. */ - num_indexes = bound->ndatums + 1; - break; - - default: - elog(ERROR, "unexpected partition strategy: %d", - (int) bound->strategy); - } - - return num_indexes; -} - /* * get_partition_operator * diff --git a/src/backend/partitioning/partprune.c b/src/backend/partitioning/partprune.c index 2da9cff1d4..f3864d2992 100644 --- a/src/backend/partitioning/partprune.c +++ b/src/backend/partitioning/partprune.c @@ -769,7 +769,10 @@ get_matching_partitions(PartitionPruneContext *context, List *pruning_steps) scan_default = final_result->scan_default; while ((i = bms_next_member(final_result->bound_offsets, i)) >= 0) { - int partindex = context->boundinfo->indexes[i]; + int partindex; + + Assert(i < context->boundinfo->nindexes); + partindex = context->boundinfo->indexes[i]; if (partindex < 0) { @@ -2497,20 +2500,19 @@ get_matching_hash_bounds(PartitionPruneContext *context, for (i = 0; i < partnatts; i++) isnull[i] = bms_is_member(i, nullkeys); - greatest_modulus = get_hash_partition_greatest_modulus(boundinfo); rowHash = compute_partition_hash_value(partnatts, partsupfunc, values, isnull); + greatest_modulus = boundinfo->nindexes; if (partindices[rowHash % greatest_modulus] >= 0) result->bound_offsets = bms_make_singleton(rowHash % greatest_modulus); } else { - /* Getting here means at least one hash partition exists. */ - Assert(boundinfo->ndatums > 0); + /* Report all valid offsets into the boundinfo->indexes array. */ result->bound_offsets = bms_add_range(NULL, 0, - boundinfo->ndatums - 1); + boundinfo->nindexes - 1); } /* @@ -3371,30 +3373,20 @@ perform_pruning_combine_step(PartitionPruneContext *context, PartitionPruneStepCombine *cstep, PruneStepResult **step_results) { - ListCell *lc1; - PruneStepResult *result = NULL; + PruneStepResult *result = (PruneStepResult *) palloc0(sizeof(PruneStepResult)); bool firststep; + ListCell *lc1; /* * A combine step without any source steps is an indication to not perform * any partition pruning. Return all datum indexes in that case. */ - result = (PruneStepResult *) palloc0(sizeof(PruneStepResult)); - if (list_length(cstep->source_stepids) == 0) + if (cstep->source_stepids == NIL) { PartitionBoundInfo boundinfo = context->boundinfo; - int rangemax; - - /* - * Add all valid offsets into the boundinfo->indexes array. For range - * partitioning, boundinfo->indexes contains (boundinfo->ndatums + 1) - * valid entries; otherwise there are boundinfo->ndatums. - */ - rangemax = context->strategy == PARTITION_STRATEGY_RANGE ? - boundinfo->ndatums : boundinfo->ndatums - 1; result->bound_offsets = - bms_add_range(result->bound_offsets, 0, rangemax); + bms_add_range(NULL, 0, boundinfo->nindexes - 1); result->scan_default = partition_bound_has_default(boundinfo); result->scan_null = partition_bound_accepts_nulls(boundinfo); return result; diff --git a/src/backend/utils/cache/partcache.c b/src/backend/utils/cache/partcache.c index 331c0f2475..a22b84e6f3 100644 --- a/src/backend/utils/cache/partcache.c +++ b/src/backend/utils/cache/partcache.c @@ -622,6 +622,7 @@ RelationBuildPartitionDesc(Relation rel) /* Moduli are stored in ascending order */ int greatest_modulus = hbounds[ndatums - 1]->modulus; + boundinfo->nindexes = greatest_modulus; boundinfo->indexes = (int *) palloc(greatest_modulus * sizeof(int)); @@ -655,6 +656,7 @@ RelationBuildPartitionDesc(Relation rel) case PARTITION_STRATEGY_LIST: { + boundinfo->nindexes = ndatums; boundinfo->indexes = (int *) palloc(ndatums * sizeof(int)); /* @@ -719,6 +721,7 @@ RelationBuildPartitionDesc(Relation rel) boundinfo->kind = (PartitionRangeDatumKind **) palloc(ndatums * sizeof(PartitionRangeDatumKind *)); + boundinfo->nindexes = ndatums + 1; boundinfo->indexes = (int *) palloc((ndatums + 1) * sizeof(int)); diff --git a/src/include/partitioning/partbounds.h b/src/include/partitioning/partbounds.h index 160b319e0a..7f830c6870 100644 --- a/src/include/partitioning/partbounds.h +++ b/src/include/partitioning/partbounds.h @@ -30,7 +30,7 @@ * In the case of range partitioning, ndatums will typically be far less than * 2 * nparts, because a partition's upper bound and the next partition's lower * bound are the same in most common cases, and we only store one of them (the - * upper bound). In case of hash partitioning, ndatums will be same as the + * upper bound). In case of hash partitioning, ndatums will be the same as the * number of partitions. * * For range and list partitioned tables, datums is an array of datum-tuples @@ -46,20 +46,26 @@ * the partition key's operator classes and collations. * * In the case of list partitioning, the indexes array stores one entry for - * every datum, which is the index of the partition that accepts a given datum. - * In case of range partitioning, it stores one entry per distinct range - * datum, which is the index of the partition for which a given datum - * is an upper bound. In the case of hash partitioning, the number of the - * entries in the indexes array is same as the greatest modulus amongst all - * partitions. For a given partition key datum-tuple, the index of the - * partition which would accept that datum-tuple would be given by the entry - * pointed by remainder produced when hash value of the datum-tuple is divided - * by the greatest modulus. + * each datum-array entry, which is the index of the partition that accepts + * rows matching that datum. So nindexes == ndatums. + * + * In the case of range partitioning, the indexes array stores one entry per + * distinct range datum, which is the index of the partition for which that + * datum is an upper bound (or -1 for a "gap" that has no partition). It is + * convenient to have an extra -1 entry representing values above the last + * range datum, so nindexes == ndatums + 1. + * + * In the case of hash partitioning, the number of entries in the indexes + * array is the same as the greatest modulus amongst all partitions (which + * is a multiple of all partition moduli), so nindexes == greatest modulus. + * The indexes array is indexed according to the hash key's remainder modulo + * the greatest modulus, and it contains either the partition index accepting + * that remainder, or -1 if there is no partition for that remainder. */ typedef struct PartitionBoundInfoData { char strategy; /* hash, list or range? */ - int ndatums; /* Length of the datums following array */ + int ndatums; /* Length of the datums[] array */ Datum **datums; PartitionRangeDatumKind **kind; /* The kind of each range bound datum; * NULL for hash and list partitioned @@ -69,6 +75,7 @@ typedef struct PartitionBoundInfoData * if there isn't one */ int default_index; /* Index of the default partition; -1 if there * isn't one */ + int nindexes; /* Length of the indexes[] array */ } PartitionBoundInfoData; #define partition_bound_accepts_nulls(bi) ((bi)->null_index != -1) diff --git a/src/test/regress/expected/partition_prune.out b/src/test/regress/expected/partition_prune.out index 3dcc8845b9..28ce692eea 100644 --- a/src/test/regress/expected/partition_prune.out +++ b/src/test/regress/expected/partition_prune.out @@ -1532,26 +1532,27 @@ drop table lp, coll_pruning, rlp, mc3p, mc2p, boolpart, boolrangep, rp, coll_pru -- result on different machines. See the definitions of -- part_part_test_int4_ops and part_test_text_ops in insert.sql. -- -create table hp (a int, b text) partition by hash (a part_test_int4_ops, b part_test_text_ops); +create table hp (a int, b text, c int) + partition by hash (a part_test_int4_ops, b part_test_text_ops); create table hp0 partition of hp for values with (modulus 4, remainder 0); create table hp3 partition of hp for values with (modulus 4, remainder 3); create table hp1 partition of hp for values with (modulus 4, remainder 1); create table hp2 partition of hp for values with (modulus 4, remainder 2); -insert into hp values (null, null); -insert into hp values (1, null); -insert into hp values (1, 'xxx'); -insert into hp values (null, 'xxx'); -insert into hp values (2, 'xxx'); -insert into hp values (1, 'abcde'); -select tableoid::regclass, * from hp order by 1; - tableoid | a | b -----------+---+------- - hp0 | | - hp0 | 1 | xxx - hp3 | 2 | xxx - hp1 | 1 | - hp2 | | xxx - hp2 | 1 | abcde +insert into hp values (null, null, 0); +insert into hp values (1, null, 1); +insert into hp values (1, 'xxx', 2); +insert into hp values (null, 'xxx', 3); +insert into hp values (2, 'xxx', 4); +insert into hp values (1, 'abcde', 5); +select tableoid::regclass, * from hp order by c; + tableoid | a | b | c +----------+---+-------+--- + hp0 | | | 0 + hp1 | 1 | | 1 + hp0 | 1 | xxx | 2 + hp2 | | xxx | 3 + hp3 | 2 | xxx | 4 + hp2 | 1 | abcde | 5 (6 rows) -- partial keys won't prune, nor would non-equality conditions @@ -1715,6 +1716,35 @@ explain (costs off) select * from hp where (a = 1 and b = 'abcde') or (a = 2 and Filter: (((a = 1) AND (b = 'abcde'::text)) OR ((a = 2) AND (b = 'xxx'::text)) OR ((a IS NULL) AND (b IS NULL))) (7 rows) +-- test pruning when not all the partitions exist +drop table hp1; +drop table hp3; +explain (costs off) select * from hp where a = 1 and b = 'abcde'; + QUERY PLAN +--------------------------------------------------- + Append + -> Seq Scan on hp2 + Filter: ((a = 1) AND (b = 'abcde'::text)) +(3 rows) + +explain (costs off) select * from hp where a = 1 and b = 'abcde' and + (c = 2 or c = 3); + QUERY PLAN +---------------------------------------------------------------------------- + Append + -> Seq Scan on hp2 + Filter: ((a = 1) AND (b = 'abcde'::text) AND ((c = 2) OR (c = 3))) +(3 rows) + +drop table hp2; +explain (costs off) select * from hp where a = 1 and b = 'abcde' and + (c = 2 or c = 3); + QUERY PLAN +-------------------------- + Result + One-Time Filter: false +(2 rows) + drop table hp; -- -- Test runtime partition pruning diff --git a/src/test/regress/sql/partition_prune.sql b/src/test/regress/sql/partition_prune.sql index f17240a5e4..9b94ac50af 100644 --- a/src/test/regress/sql/partition_prune.sql +++ b/src/test/regress/sql/partition_prune.sql @@ -291,19 +291,20 @@ drop table lp, coll_pruning, rlp, mc3p, mc2p, boolpart, boolrangep, rp, coll_pru -- part_part_test_int4_ops and part_test_text_ops in insert.sql. -- -create table hp (a int, b text) partition by hash (a part_test_int4_ops, b part_test_text_ops); +create table hp (a int, b text, c int) + partition by hash (a part_test_int4_ops, b part_test_text_ops); create table hp0 partition of hp for values with (modulus 4, remainder 0); create table hp3 partition of hp for values with (modulus 4, remainder 3); create table hp1 partition of hp for values with (modulus 4, remainder 1); create table hp2 partition of hp for values with (modulus 4, remainder 2); -insert into hp values (null, null); -insert into hp values (1, null); -insert into hp values (1, 'xxx'); -insert into hp values (null, 'xxx'); -insert into hp values (2, 'xxx'); -insert into hp values (1, 'abcde'); -select tableoid::regclass, * from hp order by 1; +insert into hp values (null, null, 0); +insert into hp values (1, null, 1); +insert into hp values (1, 'xxx', 2); +insert into hp values (null, 'xxx', 3); +insert into hp values (2, 'xxx', 4); +insert into hp values (1, 'abcde', 5); +select tableoid::regclass, * from hp order by c; -- partial keys won't prune, nor would non-equality conditions explain (costs off) select * from hp where a = 1; @@ -324,6 +325,16 @@ explain (costs off) select * from hp where a = 2 and b = 'xxx'; explain (costs off) select * from hp where a = 1 and b = 'abcde'; explain (costs off) select * from hp where (a = 1 and b = 'abcde') or (a = 2 and b = 'xxx') or (a is null and b is null); +-- test pruning when not all the partitions exist +drop table hp1; +drop table hp3; +explain (costs off) select * from hp where a = 1 and b = 'abcde'; +explain (costs off) select * from hp where a = 1 and b = 'abcde' and + (c = 2 or c = 3); +drop table hp2; +explain (costs off) select * from hp where a = 1 and b = 'abcde' and + (c = 2 or c = 3); + drop table hp; --