Result Cache, added in 9eacee2e6 neglected to properly adjust the plan
references in setrefs.c. This could lead to the following error during
EXPLAIN:
ERROR: cannot decompile join alias var in plan tree
Fix that.
Bug: 17030
Reported-by: Hans Buschmann
Discussion: https://postgr.es/m/17030-5844aecae42fe223@postgresql.org
Here we add a new executor node type named "Result Cache". The planner
can include this node type in the plan to have the executor cache the
results from the inner side of parameterized nested loop joins. This
allows caching of tuples for sets of parameters so that in the event that
the node sees the same parameter values again, it can just return the
cached tuples instead of rescanning the inner side of the join all over
again. Internally, result cache uses a hash table in order to quickly
find tuples that have been previously cached.
For certain data sets, this can significantly improve the performance of
joins. The best cases for using this new node type are for join problems
where a large portion of the tuples from the inner side of the join have
no join partner on the outer side of the join. In such cases, hash join
would have to hash values that are never looked up, thus bloating the hash
table and possibly causing it to multi-batch. Merge joins would have to
skip over all of the unmatched rows. If we use a nested loop join with a
result cache, then we only cache tuples that have at least one join
partner on the outer side of the join. The benefits of using a
parameterized nested loop with a result cache increase when there are
fewer distinct values being looked up and the number of lookups of each
value is large. Also, hash probes to lookup the cache can be much faster
than the hash probe in a hash join as it's common that the result cache's
hash table is much smaller than the hash join's due to result cache only
caching useful tuples rather than all tuples from the inner side of the
join. This variation in hash probe performance is more significant when
the hash join's hash table no longer fits into the CPU's L3 cache, but the
result cache's hash table does. The apparent "random" access of hash
buckets with each hash probe can cause a poor L3 cache hit ratio for large
hash tables. Smaller hash tables generally perform better.
The hash table used for the cache limits itself to not exceeding work_mem
* hash_mem_multiplier in size. We maintain a dlist of keys for this cache
and when we're adding new tuples and realize we've exceeded the memory
budget, we evict cache entries starting with the least recently used ones
until we have enough memory to add the new tuples to the cache.
For parameterized nested loop joins, we now consider using one of these
result cache nodes in between the nested loop node and its inner node. We
determine when this might be useful based on cost, which is primarily
driven off of what the expected cache hit ratio will be. Estimating the
cache hit ratio relies on having good distinct estimates on the nested
loop's parameters.
For now, the planner will only consider using a result cache for
parameterized nested loop joins. This works for both normal joins and
also for LATERAL type joins to subqueries. It is possible to use this new
node for other uses in the future. For example, to cache results from
correlated subqueries. However, that's not done here due to some
difficulties obtaining a distinct estimation on the outer plan to
calculate the estimated cache hit ratio. Currently we plan the inner plan
before planning the outer plan so there is no good way to know if a result
cache would be useful or not since we can't estimate the number of times
the subplan will be called until the outer plan is generated.
The functionality being added here is newly introducing a dependency on
the return value of estimate_num_groups() during the join search.
Previously, during the join search, we only ever needed to perform
selectivity estimations. With this commit, we need to use
estimate_num_groups() in order to estimate what the hit ratio on the
result cache will be. In simple terms, if we expect 10 distinct values
and we expect 1000 outer rows, then we'll estimate the hit ratio to be
99%. Since cache hits are very cheap compared to scanning the underlying
nodes on the inner side of the nested loop join, then this will
significantly reduce the planner's cost for the join. However, it's
fairly easy to see here that things will go bad when estimate_num_groups()
incorrectly returns a value that's significantly lower than the actual
number of distinct values. If this happens then that may cause us to make
use of a nested loop join with a result cache instead of some other join
type, such as a merge or hash join. Our distinct estimations have been
known to be a source of trouble in the past, so the extra reliance on them
here could cause the planner to choose slower plans than it did previous
to having this feature. Distinct estimations are also fairly hard to
estimate accurately when several tables have been joined already or when a
WHERE clause filters out a set of values that are correlated to the
expressions we're estimating the number of distinct value for.
For now, the costing we perform during query planning for result caches
does put quite a bit of faith in the distinct estimations being accurate.
When these are accurate then we should generally see faster execution
times for plans containing a result cache. However, in the real world, we
may find that we need to either change the costings to put less trust in
the distinct estimations being accurate or perhaps even disable this
feature by default. There's always an element of risk when we teach the
query planner to do new tricks that it decides to use that new trick at
the wrong time and causes a regression. Users may opt to get the old
behavior by turning the feature off using the enable_resultcache GUC.
Currently, this is enabled by default. It remains to be seen if we'll
maintain that setting for the release.
Additionally, the name "Result Cache" is the best name I could think of
for this new node at the time I started writing the patch. Nobody seems
to strongly dislike the name. A few people did suggest other names but no
other name seemed to dominate in the brief discussion that there was about
names. Let's allow the beta period to see if the current name pleases
enough people. If there's some consensus on a better name, then we can
change it before the release. Please see the 2nd discussion link below
for the discussion on the "Result Cache" name.
Author: David Rowley
Reviewed-by: Andy Fan, Justin Pryzby, Zhihong Yu, Hou Zhijie
Tested-By: Konstantin Knizhnik
Discussion: https://postgr.es/m/CAApHDvrPcQyQdWERGYWx8J%2B2DLUNgXu%2BfOSbQ1UscxrunyXyrQ%40mail.gmail.com
Discussion: https://postgr.es/m/CAApHDvq=yQXr5kqhRviT2RhNKwToaWr9JAN5t+5_PzhuRJ3wvg@mail.gmail.com
This removes "Add Result Cache executor node". It seems that something
weird is going on with the tracking of cache hits and misses as
highlighted by many buildfarm animals. It's not yet clear what the
problem is as other parts of the plan indicate that the cache did work
correctly, it's just the hits and misses that were being reported as 0.
This is especially a bad time to have the buildfarm so broken, so
reverting before too many more animals go red.
Discussion: https://postgr.es/m/CAApHDvq_hydhfovm4=izgWs+C5HqEeRScjMbOgbpC-jRAeK3Yw@mail.gmail.com
Here we add a new executor node type named "Result Cache". The planner
can include this node type in the plan to have the executor cache the
results from the inner side of parameterized nested loop joins. This
allows caching of tuples for sets of parameters so that in the event that
the node sees the same parameter values again, it can just return the
cached tuples instead of rescanning the inner side of the join all over
again. Internally, result cache uses a hash table in order to quickly
find tuples that have been previously cached.
For certain data sets, this can significantly improve the performance of
joins. The best cases for using this new node type are for join problems
where a large portion of the tuples from the inner side of the join have
no join partner on the outer side of the join. In such cases, hash join
would have to hash values that are never looked up, thus bloating the hash
table and possibly causing it to multi-batch. Merge joins would have to
skip over all of the unmatched rows. If we use a nested loop join with a
result cache, then we only cache tuples that have at least one join
partner on the outer side of the join. The benefits of using a
parameterized nested loop with a result cache increase when there are
fewer distinct values being looked up and the number of lookups of each
value is large. Also, hash probes to lookup the cache can be much faster
than the hash probe in a hash join as it's common that the result cache's
hash table is much smaller than the hash join's due to result cache only
caching useful tuples rather than all tuples from the inner side of the
join. This variation in hash probe performance is more significant when
the hash join's hash table no longer fits into the CPU's L3 cache, but the
result cache's hash table does. The apparent "random" access of hash
buckets with each hash probe can cause a poor L3 cache hit ratio for large
hash tables. Smaller hash tables generally perform better.
The hash table used for the cache limits itself to not exceeding work_mem
* hash_mem_multiplier in size. We maintain a dlist of keys for this cache
and when we're adding new tuples and realize we've exceeded the memory
budget, we evict cache entries starting with the least recently used ones
until we have enough memory to add the new tuples to the cache.
For parameterized nested loop joins, we now consider using one of these
result cache nodes in between the nested loop node and its inner node. We
determine when this might be useful based on cost, which is primarily
driven off of what the expected cache hit ratio will be. Estimating the
cache hit ratio relies on having good distinct estimates on the nested
loop's parameters.
For now, the planner will only consider using a result cache for
parameterized nested loop joins. This works for both normal joins and
also for LATERAL type joins to subqueries. It is possible to use this new
node for other uses in the future. For example, to cache results from
correlated subqueries. However, that's not done here due to some
difficulties obtaining a distinct estimation on the outer plan to
calculate the estimated cache hit ratio. Currently we plan the inner plan
before planning the outer plan so there is no good way to know if a result
cache would be useful or not since we can't estimate the number of times
the subplan will be called until the outer plan is generated.
The functionality being added here is newly introducing a dependency on
the return value of estimate_num_groups() during the join search.
Previously, during the join search, we only ever needed to perform
selectivity estimations. With this commit, we need to use
estimate_num_groups() in order to estimate what the hit ratio on the
result cache will be. In simple terms, if we expect 10 distinct values
and we expect 1000 outer rows, then we'll estimate the hit ratio to be
99%. Since cache hits are very cheap compared to scanning the underlying
nodes on the inner side of the nested loop join, then this will
significantly reduce the planner's cost for the join. However, it's
fairly easy to see here that things will go bad when estimate_num_groups()
incorrectly returns a value that's significantly lower than the actual
number of distinct values. If this happens then that may cause us to make
use of a nested loop join with a result cache instead of some other join
type, such as a merge or hash join. Our distinct estimations have been
known to be a source of trouble in the past, so the extra reliance on them
here could cause the planner to choose slower plans than it did previous
to having this feature. Distinct estimations are also fairly hard to
estimate accurately when several tables have been joined already or when a
WHERE clause filters out a set of values that are correlated to the
expressions we're estimating the number of distinct value for.
For now, the costing we perform during query planning for result caches
does put quite a bit of faith in the distinct estimations being accurate.
When these are accurate then we should generally see faster execution
times for plans containing a result cache. However, in the real world, we
may find that we need to either change the costings to put less trust in
the distinct estimations being accurate or perhaps even disable this
feature by default. There's always an element of risk when we teach the
query planner to do new tricks that it decides to use that new trick at
the wrong time and causes a regression. Users may opt to get the old
behavior by turning the feature off using the enable_resultcache GUC.
Currently, this is enabled by default. It remains to be seen if we'll
maintain that setting for the release.
Additionally, the name "Result Cache" is the best name I could think of
for this new node at the time I started writing the patch. Nobody seems
to strongly dislike the name. A few people did suggest other names but no
other name seemed to dominate in the brief discussion that there was about
names. Let's allow the beta period to see if the current name pleases
enough people. If there's some consensus on a better name, then we can
change it before the release. Please see the 2nd discussion link below
for the discussion on the "Result Cache" name.
Author: David Rowley
Reviewed-by: Andy Fan, Justin Pryzby, Zhihong Yu
Tested-By: Konstantin Knizhnik
Discussion: https://postgr.es/m/CAApHDvrPcQyQdWERGYWx8J%2B2DLUNgXu%2BfOSbQ1UscxrunyXyrQ%40mail.gmail.com
Discussion: https://postgr.es/m/CAApHDvq=yQXr5kqhRviT2RhNKwToaWr9JAN5t+5_PzhuRJ3wvg@mail.gmail.com
This allows something like
SELECT ... FROM t1 JOIN t2 USING (a, b, c) AS x
where x has the columns a, b, c and unlike a regular alias it does not
hide the range variables of the tables being joined t1 and t2.
Per SQL:2016 feature F404 "Range variable for common column names".
Reviewed-by: Vik Fearing <vik.fearing@2ndquadrant.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://www.postgresql.org/message-id/flat/454638cf-d563-ab76-a585-2564428062af@2ndquadrant.com
Previously, pull_varnos() took the relids of a PlaceHolderVar as being
equal to the relids in its contents, but that fails to account for the
possibility that we have to postpone evaluation of the PHV due to outer
joins. This could result in a malformed plan. The known cases end up
triggering the "failed to assign all NestLoopParams to plan nodes"
sanity check in createplan.c, but other symptoms may be possible.
The right value to use is the join level we actually intend to evaluate
the PHV at. We can get that from the ph_eval_at field of the associated
PlaceHolderInfo. However, there are some places that call pull_varnos()
before the PlaceHolderInfos have been created; in that case, fall back
to the conservative assumption that the PHV will be evaluated at its
syntactic level. (In principle this might result in missing some legal
optimization, but I'm not aware of any cases where it's an issue in
practice.) Things are also a bit ticklish for calls occurring during
deconstruct_jointree(), but AFAICS the ph_eval_at fields should have
reached their final values by the time we need them.
The main problem in making this work is that pull_varnos() has no
way to get at the PlaceHolderInfos. We can fix that easily, if a
bit tediously, in HEAD by passing it the planner "root" pointer.
In the back branches that'd cause an unacceptable API/ABI break for
extensions, so leave the existing entry points alone and add new ones
with the additional parameter. (If an old entry point is called and
encounters a PHV, it'll fall back to using the syntactic level,
again possibly missing some valid optimization.)
Back-patch to v12. The computation is surely also wrong before that,
but it appears that we cannot reach a bad plan thanks to join order
restrictions imposed on the subquery that the PlaceHolderVar came from.
The error only became reachable when commit 4be058fe9 allowed trivial
subqueries to be collapsed out completely, eliminating their join order
restrictions.
Per report from Stephan Springl.
Discussion: https://postgr.es/m/171041.1610849523@sss.pgh.pa.us
Many older tests where written in a style like
SELECT '' AS two, i.* FROM INT2_TBL
where the first column indicated the number of expected result rows.
This has gotten increasingly out of date, as the test data fixtures
have expanded, so a lot of these were wrong and misleading. Moreover,
this style isn't really necessary, since the psql output already shows
the number of result rows.
To clean this up, remove all those extra columns.
Discussion: https://www.postgresql.org/message-id/flat/1a25312b-2686-380d-3c67-7a69094a999f%40enterprisedb.com
Commit 4be058fe9 forgot that the append_rel_list would already be
populated at the time we remove useless result RTEs, and it might contain
PlaceHolderVars that need to be adjusted like the ones in the main parse
tree. This could lead to "no relation entry for relid N" failures later
on, when the planner tries to do something with an unadjusted PHV.
Per report from Tom Ellis. Back-patch to v12 where the bug came in.
Discussion: https://postgr.es/m/20201205173056.GF30712@cloudinit-builder
If a PlaceHolderVar is to be evaluated at a join relation, but
its value is only needed there and not at higher levels, we neglected
to update the joinrel's direct_lateral_relids to include the PHV's
source rel. This causes problems because join_is_legal() then won't
allow joining the joinrel to the PHV's source rel at all, leading
to "failed to build any N-way joins" planner failures.
Per report from Andreas Seltenreich. Back-patch to 9.5
where the problem originated.
Discussion: https://postgr.es/m/87blfgqa4t.fsf@aurora.ydns.eu
Add hash functions for the record type as well as a hash operator
family and operator class for the record type. This enables all the
hash functionality for the record type such as hash-based plans for
UNION/INTERSECT/EXCEPT DISTINCT, recursive queries using UNION
DISTINCT, hash joins, and hash partitioning.
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://www.postgresql.org/message-id/flat/38eccd35-4e2d-6767-1b3c-dada1eac3124%402ndquadrant.com
get_foreign_key_join_selectivity() looks for join clauses that equate
the two sides of the FK constraint. However, if we have a query like
"WHERE fktab.a = pktab.a and fktab.a = 1", it won't find any such join
clause, because equivclass.c replaces the given clauses with "fktab.a
= 1 and pktab.a = 1", which can be enforced at the scan level, leaving
nothing to be done for column "a" at the join level.
We can fix that expectation without much trouble, but then a new problem
arises: applying the foreign-key-based selectivity rule produces a
rowcount underestimate, because we're effectively double-counting the
selectivity of the "fktab.a = 1" clause. So we have to cancel that
selectivity out of the estimate.
To fix, refactor process_implied_equality() so that it can pass back the
new RestrictInfo to its callers in equivclass.c, allowing the generated
"fktab.a = 1" clause to be saved in the EquivalenceClass's ec_derives
list. Then it's not much trouble to dig out the relevant RestrictInfo
when we need to adjust an FK selectivity estimate. (While at it, we
can also remove the expensive use of initialize_mergeclause_eclasses()
to set up the new RestrictInfo's left_ec and right_ec pointers.
The equivclass.c code can set those basically for free.)
This seems like clearly a bug fix, but I'm hesitant to back-patch it,
first because there's some API/ABI risk for extensions and second because
we're usually loath to destabilize plan choices in stable branches.
Per report from Sigrid Ehrenreich.
Discussion: https://postgr.es/m/1019549.1603770457@sss.pgh.pa.us
Discussion: https://postgr.es/m/AM6PR02MB5287A0ADD936C1FA80973E72AB190@AM6PR02MB5287.eurprd02.prod.outlook.com
Add test coverage for the nbtutils.c routines concerned with IndexScans
that have native ScalarArrayOpExpr quals. The ScalarArrayOpExpr
specialized mark and restore routines, and the "find extreme element"
routine now have some test coverage.
These functions are probably infrequently exercised by real world
queries, so having some coverage seems like a good idea. The mark and
restore routines were originally added by a bugfix that came several
weeks after the first stable release of Postgres 9.2 (see commit
70bc583319).
The RTE_RESULT simplification logic added by commit 4be058fe9 had a
flaw: it would collapse out a RTE_RESULT that is due to compute a
PlaceHolderVar, and reassign the PHV to the parent join level, even if
another input relation of the join contained a lateral reference to
the PHV. That can't work because the PHV would be computed too late.
In practice it led to failures of internal sanity checks later in
planning (either assertion failures or errors such as "failed to
construct the join relation").
To fix, add code to check for the presence of such PHVs in relevant
portions of the query tree. Notably, this required refactoring
range_table_walker so that a caller could ask to walk individual RTEs
not the whole list. (It might be a good idea to refactor
range_table_mutator in the same way, if only to keep those functions
looking similar; but I didn't do so here as it wasn't necessary for
the bug fix.)
This exercise also taught me that find_dependent_phvs(), as it stood,
could only safely be used on the entire Query, not on subtrees.
Adjust its API to reflect that; which in passing allows it to have
a fast path for the common case of no PHVs anywhere.
Per report from Will Leinweber. Back-patch to v12 where the bug
was introduced.
Discussion: https://postgr.es/m/CALLb-4xJMd4GZt2YCecMC95H-PafuWNKcmps4HLRx2NHNBfB4g@mail.gmail.com
This patch causes EXPLAIN to always assign a separate table alias to the
parent RTE of an append relation (inheritance set); before, such RTEs
were ignored if not actually scanned by the plan. Since the child RTEs
now always have that same alias to start with (cf. commit 55a1954da),
the net effect is that the parent RTE usually gets the alias used or
implied by the query text, and the children all get that alias with "_N"
appended. (The exception to "usually" is if there are duplicate aliases
in different subtrees of the original query; then some of those original
RTEs will also have "_N" appended.)
This results in more uniform output for partitioned-table plans than
we had before: the partitioned table itself gets the original alias,
and all child tables have aliases with "_N", rather than the previous
behavior where one of the children would get an alias without "_N".
The reason for giving the parent RTE an alias, even if it isn't scanned
by the plan, is that we now use the parent's alias to qualify Vars that
refer to an appendrel output column and appear above the Append or
MergeAppend that computes the appendrel. But below the append, Vars
refer to some one of the child relations, and are displayed that way.
This seems clearer than the old behavior where a Var that could carry
values from any child relation was displayed as if it referred to only
one of them.
While at it, change ruleutils.c so that the code paths used by EXPLAIN
deal in Plan trees not PlanState trees. This effectively reverts a
decision made in commit 1cc29fe7c, which seemed like a good idea at
the time to make ruleutils.c consistent with explain.c. However,
it's problematic because we'd really like to allow executor startup
pruning to remove all the children of an append node when possible,
leaving no child PlanState to resolve Vars against. (That's not done
here, but will be in the next patch.) This requires different handling
of subplans and initplans than before, but is otherwise a pretty
straightforward change.
Discussion: https://postgr.es/m/001001d4f44b$2a2cca50$7e865ef0$@lab.ntt.co.jp
Fix an oversight in commit 7266d0997: as it stood, the code failed
when a function-in-FROM returns composite and can be simplified
to a composite constant.
For the moment, just test for composite result and abandon pullup
if we see one. To make it actually work, we'd have to decompose
the composite constant into per-column constants; which is surely
do-able, but I'm not convinced it's worth the code space.
Per report from Raúl Marín Rodríguez.
Discussion: https://postgr.es/m/CAM6_UM4isP+buRA5sWodO_MUEgutms-KDfnkwGmryc5DGj9XuQ@mail.gmail.com
This allows simplification of the plan tree in some common usage
patterns: we can get rid of a join to the function RTE.
In principle we could pull up any immutable expression, but restricting
it to Consts avoids the risk that multiple evaluations of the expression
might cost more than we can save. (Possibly this could be improved in
future --- but we've more or less promised people that putting a function
in FROM guarantees single evaluation, so we'd have to tread carefully.)
To do this, we need to rearrange when eval_const_expressions()
happens for expressions in function RTEs. I moved it to
inline_set_returning_functions(), which already has to iterate over
every function RTE, and in consequence renamed that function to
preprocess_function_rtes(). A useful consequence is that
inline_set_returning_function() no longer has to do this for itself,
simplifying that code.
In passing, break out pull_up_simple_subquery's code that knows where
everything that needs pullup_replace_vars() processing is, so that
the new pull_up_constant_function() routine can share it. We'd
gotten away with one-and-a-half copies of that code so far, since
pull_up_simple_values() could assume that a lot of cases didn't apply
to it --- but I don't think pull_up_constant_function() can make any
simplifying assumptions. Might as well make pull_up_simple_values()
use it too.
(Possibly this refactoring should go further: maybe we could share
some of the code to fill in the pullup_replace_vars_context struct?
For now, I left it that the callers fill that completely.)
Note: the one existing test case that this patch changes has to be
changed because inlining its function RTEs would destroy the point
of the test, namely to check join order.
Alexander Kuzmenkov and Aleksandr Parfenov, reviewed by
Antonin Houska and Anastasia Lubennikova, and whacked around
some more by me
Discussion: https://postgr.es/m/402356c32eeb93d4fed01f66d6c7fe2d@postgrespro.ru
The point of this change is to increase the potential for parallelism
while running the core regression tests. Most people these days are
using parallel testing modes on multi-core machines, so we might as
well try a bit harder to keep multiple cores busy. Hence, a test that
runs much longer than others in its parallel group is a candidate to
be sub-divided.
In this patch, create_index.sql and join.sql are split up.
I haven't changed the content of the tests in any way, just
moved them.
I moved create_index.sql's SP-GiST-related tests into a new script
create_index_spgist, and moved its btree multilevel page deletion test
over to the existing script btree_index. (btree_index is a more natural
home for that test, and it's shorter than others in its parallel group,
so this doesn't hurt total runtime of that group.) There might be
room for more aggressive splitting of create_index, but this is enough
to improve matters considerably.
Likewise, I moved join.sql's "exercises for the hash join code" into
a new file join_hash. Those exercises contributed three-quarters of
the script's runtime. Which might well be excessive ... but for the
moment, I'm satisfied with shoving them into a different parallel
group, where they can share runtime with the roughly-equally-lengthy
gist test.
(Note for anybody following along at home: there are interesting
interactions between the runtimes of create_index and anything running
in parallel with it, because the tests of CREATE INDEX CONCURRENTLY
in that file will repeatedly block waiting for concurrent transactions
to commit. As committed in this patch, create_index and
create_index_spgist have roughly equal runtimes, but that's mostly an
artifact of forced synchronization of the CONCURRENTLY tests; when run
serially, create_index is much faster. A followup patch will reduce
the runtime of create_index_spgist and thereby also create_index.)
Discussion: https://postgr.es/m/735.1554935715@sss.pgh.pa.us
The assertions added by commits 34ea1ab7f et al found another problem:
set_dummy_rel_pathlist and mark_dummy_rel were failing to label
the dummy paths they create with the correct outer_relids, in case
the relation is necessarily parameterized due to having lateral
references in its tlist. It's likely that this has no user-visible
consequences in production builds, at the moment; but still an assertion
failure is a bad thing, so back-patch the fix.
Per bug #15694 from Roman Zharkov (via Alexander Lakhin)
and an independent report by Tushar Ahuja.
Discussion: https://postgr.es/m/15694-74f2ca97e7044f7f@postgresql.org
Discussion: https://postgr.es/m/7d72ab20-c725-3ce2-f99d-4e64dd8a0de6@enterprisedb.com
There isn't any good reason for this test to be a self-join rather
than a join between separate tables, except that it saved a couple
of SQL commands for setup. A proposed patch to optimize away
self-joins breaks the test, so adjust it to avoid that happening.
Discussion: https://postgr.es/m/64486b0b-0404-e39e-322d-0801154901f3@postgrespro.ru
In place of three separate but interrelated lists (indexclauses,
indexquals, and indexqualcols), an IndexPath now has one list
"indexclauses" of IndexClause nodes. This holds basically the same
information as before, but in a more useful format: in particular, there
is now a clear connection between an indexclause (an original restriction
clause from WHERE or JOIN/ON) and the indexquals (directly usable index
conditions) derived from it.
We also change the ground rules a bit by mandating that clause commutation,
if needed, be done up-front so that what is stored in the indexquals list
is always directly usable as an index condition. This gets rid of repeated
re-determination of which side of the clause is the indexkey during costing
and plan generation, as well as repeated lookups of the commutator
operator. To minimize the added up-front cost, the typical case of
commuting a plain OpExpr is handled by a new special-purpose function
commute_restrictinfo(). For RowCompareExprs, generating the new clause
properly commuted to begin with is not really any more complex than before,
it's just different --- and we can save doing that work twice, as the
pretty-klugy original implementation did.
Tracking the connection between original and derived clauses lets us
also track explicitly whether the derived clauses are an exact or lossy
translation of the original. This provides a cheap solution to getting
rid of unnecessary rechecks of boolean index clauses, which previously
seemed like it'd be more expensive than it was worth.
Another pleasant (IMO) side-effect is that EXPLAIN now always shows
index clauses with the indexkey on the left; this seems less confusing.
This commit leaves expand_indexqual_conditions() and some related
functions in a slightly messy state. I didn't bother to change them
any more than minimally necessary to work with the new data structure,
because all that code is going to be refactored out of existence in
a follow-on patch.
Discussion: https://postgr.es/m/22182.1549124950@sss.pgh.pa.us
The fact that "SELECT expression" has no base relations has long been a
thorn in the side of the planner. It makes it hard to flatten a sub-query
that looks like that, or is a trivial VALUES() item, because the planner
generally uses relid sets to identify sub-relations, and such a sub-query
would have an empty relid set if we flattened it. prepjointree.c contains
some baroque logic that works around this in certain special cases --- but
there is a much better answer. We can replace an empty FROM clause with a
dummy RTE that acts like a table of one row and no columns, and then there
are no such corner cases to worry about. Instead we need some logic to
get rid of useless dummy RTEs, but that's simpler and covers more cases
than what was there before.
For really trivial cases, where the query is just "SELECT expression" and
nothing else, there's a hazard that adding the extra RTE makes for a
noticeable slowdown; even though it's not much processing, there's not
that much for the planner to do overall. However testing says that the
penalty is very small, close to the noise level. In more complex queries,
this is able to find optimizations that we could not find before.
The new RTE type is called RTE_RESULT, since the "scan" plan type it
gives rise to is a Result node (the same plan we produced for a "SELECT
expression" query before). To avoid confusion, rename the old ResultPath
path type to GroupResultPath, reflecting that it's only used in degenerate
grouping cases where we know the query produces just one grouped row.
(It wouldn't work to unify the two cases, because there are different
rules about where the associated quals live during query_planner.)
Note: although this touches readfuncs.c, I don't think a catversion
bump is required, because the added case can't occur in stored rules,
only plans.
Patch by me, reviewed by David Rowley and Mark Dilger
Discussion: https://postgr.es/m/15944.1521127664@sss.pgh.pa.us
This prevents failures in cases where we pull up a constant or var-free
expression from a subquery and put it into a full join's qual. That can
result in not recognizing the qual as containing a mergejoin-able or
hashjoin-able condition. A PHV prevents the problem because it is still
recognized as belonging to the side of the join the subquery is in.
I'm not very sure about the net effect of this change on plan quality.
In "typical" cases where the join keys are Vars, nothing changes.
In an affected case, the PHV-wrapped expression is less likely to be seen
as equal to PHV-less instances below the join, but more likely to be seen
as equal to similar expressions above the join, so it may end up being a
wash. In the one existing case where there's any visible change in a
regression-test plan, it amounts to referencing a lower computation of a
COALESCE result instead of recomputing it, which seems like a win.
Given my uncertainty about that and the lack of field complaints,
no back-patch, even though this is a very ancient problem.
Discussion: https://postgr.es/m/32090.1539378124@sss.pgh.pa.us
canonicalize_ec_expression() is supposed to agree with coerce_type() as to
whether a RelabelType should be inserted to make a subexpression be valid
input for the operators of a given opclass. However, it did the wrong
thing with named-composite-type inputs to record_eq(): it put in a
RelabelType to RECORDOID, which the parser doesn't. In some cases this was
harmless because all code paths involving a particular equivalence class
did the same thing, but in other cases this would result in failing to
recognize a composite-type expression as being a member of an equivalence
class that it actually is a member of. The most obvious bad effect was to
fail to recognize that an index on a composite column could provide the
sort order needed for a mergejoin on that column, as reported by Teodor
Sigaev. I think there might be other, subtler, cases that result in
misoptimization. It also seems possible that an unwanted RelabelType
would sometimes get into an emitted plan --- but because record_eq and
friends don't examine the declared type of their input expressions, that
would not create any visible problems.
To fix, just treat RECORDOID as if it were a polymorphic type, which in
some sense it is. We might want to consider formalizing that a bit more
someday, but for the moment this seems to be the only place where an
IsPolymorphicType() test ought to include RECORDOID as well.
This has been broken for a long time, so back-patch to all supported
branches.
Discussion: https://postgr.es/m/a6b22369-e3bf-4d49-f59d-0c41d3551e81@sigaev.ru
In some cases a clause attached to an outer join can be pushed down into
the outer join's RHS even though the clause is not degenerate --- this
can happen if we choose to make a parameterized path for the RHS. If
the clause ends up attached to a lower outer join, we'd misclassify it
as being a "join filter" not a plain "filter" condition at that node,
leading to wrong query results.
To fix, teach extract_actual_join_clauses to examine each join clause's
required_relids, not just its is_pushed_down flag. (The latter now
seems vestigial, or at least in need of rethinking, but we won't do
anything so invasive as redefining it in a bug-fix patch.)
This has been wrong since we introduced parameterized paths in 9.2,
though it's evidently hard to hit given the lack of previous reports.
The test case used here involves a lateral function call, and I think
that a lateral reference may be required to get the planner to select
a broken plan; though I wouldn't swear to that. In any case, even if
LATERAL is needed to trigger the bug, it still affects all supported
branches, so back-patch to all.
Per report from Andreas Karlsson. Thanks to Andrew Gierth for
preliminary investigation.
Discussion: https://postgr.es/m/f8128b11-c5bf-3539-48cd-234178b2314d@proxel.se
Many of the objects we create during the regression tests are put in the
public schema, so that using the same names in different regression tests
creates a hazard of test failures if any two such scripts run concurrently.
This patch cleans up a bunch of latent hazards of that sort, as well as two
live hazards.
The current situation in this regard is far worse than it was a year or two
back, because practically all of the partitioning-related test cases have
reused table names with enthusiasm. I despaired of cleaning up that mess
within the five most-affected tests (create_table, alter_table, insert,
update, inherit); fortunately those don't run concurrently.
Other than partitioning problems, most of the issues boil down to using
names like "foo", "bar", "tmp", etc, without thought for the fact that
other test scripts might use similar names concurrently. I've made an
effort to make all such names more specific.
One of the live hazards was that commit 7421f4b8 caused with.sql to
create a table named "test", conflicting with a similarly-named table
in alter_table.sql; this was exposed in the buildfarm recently.
The other one was that join.sql and transactions.sql both create tables
named "foo" and "bar"; but join.sql's uses of those names date back
only to December or so.
Since commit 7421f4b8 was back-patched to v10, back-patch a minimal
fix for that problem. The rest of this is just future-proofing.
Discussion: https://postgr.es/m/4627.1521070268@sss.pgh.pa.us
Given overlapping or partially redundant join clauses, for example
t1 JOIN t2 ON t1.a = t2.x AND t1.b = t2.x
the planner's EquivalenceClass machinery will ordinarily refactor the
clauses as "t1.a = t1.b AND t1.a = t2.x", so that join processing doesn't
see multiple references to the same EquivalenceClass in a list of join
equality clauses. However, if the join is outer, it's incorrect to derive
a restriction clause on the outer side from the join conditions, so the
clause refactoring does not happen and we end up with overlapping join
conditions. The code that attempted to deal with such cases had several
subtle bugs, which could result in "left and right pathkeys do not match in
mergejoin" or "outer pathkeys do not match mergeclauses" planner errors,
if the selected join plan type was a mergejoin. (It does not appear that
any actually incorrect plan could have been emitted.)
The core of the problem really was failure to recognize that the outer and
inner relations' pathkeys have different relationships to the mergeclause
list. A join's mergeclause list is constructed by reference to the outer
pathkeys, so it will always be ordered the same as the outer pathkeys, but
this cannot be presumed true for the inner pathkeys. If the inner sides of
the mergeclauses contain multiple references to the same EquivalenceClass
({t2.x} in the above example) then a simplistic rendering of the required
inner sort order is like "ORDER BY t2.x, t2.x", but the pathkey machinery
recognizes that the second sort column is redundant and throws it away.
The mergejoin planning code failed to account for that behavior properly.
One error was to try to generate cut-down versions of the mergeclause list
from cut-down versions of the inner pathkeys in the same way as the initial
construction of the mergeclause list from the outer pathkeys was done; this
could lead to choosing a mergeclause list that fails to match the outer
pathkeys. The other problem was that the pathkey cross-checking code in
create_mergejoin_plan treated the inner and outer pathkey lists
identically, whereas actually the expectations for them must be different.
That led to false "pathkeys do not match" failures in some cases, and in
principle could have led to failure to detect bogus plans in other cases,
though there is no indication that such bogus plans could be generated.
Reported by Alexander Kuzmenkov, who also reviewed this patch. This has
been broken for years (back to around 8.3 according to my testing), so
back-patch to all supported branches.
Discussion: https://postgr.es/m/5dad9160-4632-0e47-e120-8e2082000c01@postgrespro.ru
If we're inside a lateral subquery, there may be no unparameterized paths
for a particular child relation of an appendrel, in which case we *must*
be able to create similarly-parameterized paths for each other child
relation, else the planner will fail with "could not devise a query plan
for the given query". This means that there are situations where we'd
better be able to reparameterize at least one path for each child.
This calls into question the assumption in reparameterize_path() that
it can just punt if it feels like it. However, the only case that is
known broken right now is where the child is itself an appendrel so that
all its paths are AppendPaths. (I think possibly I disregarded that in
the original coding on the theory that nested appendrels would get folded
together --- but that only happens *after* reparameterize_path(), so it's
not excused from handling a child AppendPath.) Given that this code's been
like this since 9.3 when LATERAL was introduced, it seems likely we'd have
heard of other cases by now if there were a larger problem.
Per report from Elvis Pranskevichus. Back-patch to 9.3.
Discussion: https://postgr.es/m/5981018.zdth1YWmNy@hammer.magicstack.net
Introduce parallel-aware hash joins that appear in EXPLAIN plans as Parallel
Hash Join with Parallel Hash. While hash joins could already appear in
parallel queries, they were previously always parallel-oblivious and had a
partial subplan only on the outer side, meaning that the work of the inner
subplan was duplicated in every worker.
After this commit, the planner will consider using a partial subplan on the
inner side too, using the Parallel Hash node to divide the work over the
available CPU cores and combine its results in shared memory. If the join
needs to be split into multiple batches in order to respect work_mem, then
workers process different batches as much as possible and then work together
on the remaining batches.
The advantages of a parallel-aware hash join over a parallel-oblivious hash
join used in a parallel query are that it:
* avoids wasting memory on duplicated hash tables
* avoids wasting disk space on duplicated batch files
* divides the work of building the hash table over the CPUs
One disadvantage is that there is some communication between the participating
CPUs which might outweigh the benefits of parallelism in the case of small
hash tables. This is avoided by the planner's existing reluctance to supply
partial plans for small scans, but it may be necessary to estimate
synchronization costs in future if that situation changes. Another is that
outer batch 0 must be written to disk if multiple batches are required.
A potential future advantage of parallel-aware hash joins is that right and
full outer joins could be supported, since there is a single set of matched
bits for each hashtable, but that is not yet implemented.
A new GUC enable_parallel_hash is defined to control the feature, defaulting
to on.
Author: Thomas Munro
Reviewed-By: Andres Freund, Robert Haas
Tested-By: Rafia Sabih, Prabhat Sahu
Discussion:
https://postgr.es/m/CAEepm=2W=cOkiZxcg6qiFQP-dHUe09aqTrEMM7yJDrHMhDv_RA@mail.gmail.comhttps://postgr.es/m/CAEepm=37HKyJ4U6XOLi=JgfSHM3o6B-GaeO-6hkOmneTDkH+Uw@mail.gmail.com
If a hash join appears in a parallel query, there may be no hash table
available for explain.c to inspect even though a hash table may have
been built in other processes. This could happen either because
parallel_leader_participation was set to off or because the leader
happened to hit the end of the outer relation immediately (even though
the complete relation is not empty) and decided not to build the hash
table.
Commit bf11e7ee introduced a way for workers to exchange
instrumentation via the DSM segment for Sort nodes even though they
are not parallel-aware. This commit does the same for Hash nodes, so
that explain.c has a way to find instrumentation data from an
arbitrary participant that actually built the hash table.
Author: Thomas Munro
Reviewed-By: Andres Freund
Discussion: https://postgr.es/m/CAEepm%3D3DUQC2-z252N55eOcZBer6DPdM%3DFzrxH9dZc5vYLsjaA%40mail.gmail.com
Previously, this function estimated the selectivity as 1 minus eqjoinsel()
for the negator equality operator, regardless of join type (I think there
was an expectation that eqjoinsel would handle the join type). But
actually this is completely wrong for semijoin cases: the fraction of the
LHS that has a non-matching row is not one minus the fraction of the LHS
that has a matching row. In reality a semijoin with <> will nearly always
succeed: it can only fail when the RHS is empty, or it contains a single
distinct value that is equal to the particular LHS value, or the LHS value
is null. The only one of those things we should have much confidence in
estimating is the fraction of LHS values that are null, so let's just take
the selectivity as 1 minus outer nullfrac.
Per coding convention, antijoin should be estimated the same as semijoin.
Arguably this is a bug fix, but in view of the lack of field complaints
and the risk of destabilizing plans, no back-patch.
Thomas Munro, reviewed by Ashutosh Bapat
Discussion: https://postgr.es/m/CAEepm=270ze2hVxWkJw-5eKzc3AB4C9KpH3L2kih75R5pdSogg@mail.gmail.com
Although hash joins are already tested by many queries, these tests
systematically cover the four different states we can reach as part of
the strategy for respecting work_mem.
Author: Thomas Munro
Reviewed-By: Andres Freund
Flattening the partitioning hierarchy at this stage makes various
desirable optimizations difficult. The original use case for this
patch was partition-wise join, which wants to match up the partitions
in one partitioning hierarchy with those in another such hierarchy.
However, it now seems that it will also be useful in making partition
pruning work using the PartitionDesc rather than constraint exclusion,
because with a flattened expansion, we have no easy way to figure out
which PartitionDescs apply to which leaf tables in a multi-level
partition hierarchy.
As it turns out, we end up creating both rte->inh and !rte->inh RTEs
for each intermediate partitioned table, just as we previously did for
the root table. This seems unnecessary since the partitioned tables
have no storage and are not scanned. We might want to go back and
rejigger things so that no partitioned tables (including the parent)
need !rte->inh RTEs, but that seems to require some adjustments not
related to the core purpose of this patch.
Ashutosh Bapat, reviewed by me and by Amit Langote. Some final
adjustments by me.
Discussion: http://postgr.es/m/CAFjFpRd=1venqLL7oGU=C1dEkuvk2DJgvF+7uKbnPHaum1mvHQ@mail.gmail.com
David Rowley found that the "use the smallest per-column selectivity"
heuristic applied in some cases by get_foreign_key_join_selectivity()
was badly off if the FK columns are independent, producing estimates
much worse than we got before that code was added in 9.6.
One case where that heuristic was used was for LEFT and FULL outer joins
with the referenced rel on the outside of the join. But we should not
really need to special-case those here. eqjoinsel() never has had such a
special case; the correction is applied by calc_joinrel_size_estimate()
instead. Let's just estimate such cases like inner joins and rely on that
later adjustment. (I think there was something of a thinko here, in that
the comments seem to be thinking about the selectivity as defined for
semi/anti joins; but that shouldn't apply to left/full joins.) Add a
regression test exercising such a case to show that this is sane in
at least some cases.
The other case where we used that heuristic was for SEMI/ANTI outer joins,
either if the referenced rel was on the outside, or if it was on the inside
but was part of a join within the RHS. In either case, the FK doesn't give
us a lot of traction towards estimating the selectivity. To ensure that
we don't have regressions from what happened before 9.6, let's punt by
ignoring the FK in such cases and applying the traditional selectivity
calculation. (We might be able to improve on that later, but for now
I just want to be sure it's not worse than 9.5.)
Report and patch by David Rowley, simplified a bit by me. Back-patch
to 9.6 where this code was added.
Discussion: https://postgr.es/m/CAKJS1f8NO8oCDcxrteohG6O72uU1saEVT9qX=R8pENr5QWerXw@mail.gmail.com
When costing a nestloop with stop-at-first-inner-match semantics, and a
non-indexscan inner path, final_cost_nestloop() wants to charge the full
scan cost of the inner rel at least once, with additional scans charged
at inner_rescan_run_cost which might be less. However the logic for
doing this effectively assumed that outer_matched_rows is at least 1.
If it's zero, which is not unlikely for a small outer rel, we ended up
charging inner_run_cost plus N times inner_rescan_run_cost, as much as
double the correct charge for an outer rel with only one row that
we're betting won't be matched. (Unless the inner rel is materialized,
in which case it has very small inner_rescan_run_cost and the cost
is not so far off what it should have been.)
The upshot of this was that the planner had a tendency to select plans
that failed to make effective use of the stop-at-first-inner-match
semantics, and that might have Materialize nodes in them even when the
predicted number of executions of the Materialize subplan was only 1.
This was not so obvious before commit 9c7f5229a, because the case only
arose in connection with semi/anti joins where there's not freedom to
reverse the join order. But with the addition of unique-inner joins,
it could result in some fairly bad planning choices, as reported by
Teodor Sigaev. Indeed, some of the test cases added by that commit
have plans that look dubious on closer inspection, and are changed
by this patch.
Fix the logic to ensure that we don't charge for too many inner scans.
I chose to adjust it so that the full-freight scan cost is associated
with an unmatched outer row if possible, not a matched one, since that
seems like a better model of what would happen at runtime.
This is a longstanding bug, but given the lesser impact in back branches,
and the lack of field complaints, I won't risk a back-patch.
Discussion: https://postgr.es/m/CAKJS1f-LzkUsFxdJ_-Luy38orQ+AdEXM5o+vANR+-pHAWPSecg@mail.gmail.com
If the inner relation can be proven unique, that is it can have no more
than one matching row for any row of the outer query, then we might as
well implement the semijoin as a plain inner join, allowing substantially
more freedom to the planner. This is a form of outer join strength
reduction, but it can't be implemented in reduce_outer_joins() because
we don't have enough info about the individual relations at that stage.
Instead do it much like remove_useless_joins(): once we've built base
relations, we can make another pass over the SpecialJoinInfo list and
get rid of any entries representing reducible semijoins.
This is essentially a followon to the inner-unique patch (commit 9c7f5229a)
and makes use of the proof machinery that that patch created. We need only
minor refactoring of innerrel_is_unique's API to support this usage.
Per performance complaint from Teodor Sigaev.
Discussion: https://postgr.es/m/f994fc98-389f-4a46-d1bc-c42e05cb43ed@sigaev.ru
The inner-unique patch (commit 9c7f5229a) supposed that if we're
considering a JOIN_UNIQUE_INNER join path, we can always set inner_unique
for the join, because the inner path produced by create_unique_path should
be unique relative to the outer relation. However, that's true only if
we're considering joining to the whole outer relation --- otherwise we may
be applying only some of the join quals, and so the inner path might be
non-unique from the perspective of this join. Adjust the test to only
believe that we can set inner_unique if we have the whole semijoin LHS on
the outer side.
There is more that can be done in this area, but this commit is only
intended to provide the minimal fix needed to get correct plans.
Per report from Teodor Sigaev. Thanks to David Rowley for preliminary
investigation.
Discussion: https://postgr.es/m/f994fc98-389f-4a46-d1bc-c42e05cb43ed@sigaev.ru
If there can certainly be no more than one matching inner row for a given
outer row, then the executor can move on to the next outer row as soon as
it's found one match; there's no need to continue scanning the inner
relation for this outer row. This saves useless scanning in nestloop
and hash joins. In merge joins, it offers the opportunity to skip
mark/restore processing, because we know we have not advanced past the
first possible match for the next outer row.
Of course, the devil is in the details: the proof of uniqueness must
depend only on joinquals (not otherquals), and if we want to skip
mergejoin mark/restore then it must depend only on merge clauses.
To avoid adding more planning overhead than absolutely necessary,
the present patch errs in the conservative direction: there are cases
where inner_unique or skip_mark_restore processing could be used, but
it will not do so because it's not sure that the uniqueness proof
depended only on "safe" clauses. This could be improved later.
David Rowley, reviewed and rather heavily editorialized on by me
Discussion: https://postgr.es/m/CAApHDvqF6Sw-TK98bW48TdtFJ+3a7D2mFyZ7++=D-RyPsL76gw@mail.gmail.com
A proposed patch, also by Thomas and in the same thread, would change
the output order of these. Independent of the follow-up patches
getting committed, nailing down the order in these specific tests at
worst seems harmless.
Author: Thomas Munro
Discussion: https://postgr.es/m/CAEepm=1D4-tP7j7UAgT_j4ZX2j4Ehe1qgZQWFKBMb8F76UW5Rg@mail.gmail.com
Given a three-or-more-way equivalence class, such as X.Y = Y.Y = Z.Z,
it was possible for the planner to omit one of the quals needed to
enforce that all members of the equivalence class are actually equal.
This only happened in the case of a parameterized join node for two
of the relations, that is a plan tree like
Nested Loop
-> Scan X
-> Nested Loop
-> Scan Y
-> Scan Z
Filter: Z.Z = X.X
The eclass machinery normally expects to apply X.X = Y.Y when those
two relations are joined, but in this shape of plan tree they aren't
joined until the top node --- and, if the lower nested loop is marked
as parameterized by X, the top node will assume that the relevant eclass
condition(s) got pushed down into the lower node. On the other hand,
the scan of Z assumes that it's only responsible for constraining Z.Z
to match any one of the other eclass members. So one or another of
the required quals sometimes fell between the cracks, depending on
whether consideration of the eclass in get_joinrel_parampathinfo()
for the lower nested loop chanced to generate X.X = Y.Y or X.X = Z.Z
as the appropriate constraint there. If it generated the latter,
it'd erroneously suppose that the Z scan would take care of matters.
To fix, force X.X = Y.Y to be generated and applied at that join node
when this case occurs.
This is *extremely* hard to hit in practice, because various planner
behaviors conspire to mask the problem; starting with the fact that the
planner doesn't really like to generate a parameterized plan of the
above shape. (It might have been impossible to hit it before we
tweaked things to allow this plan shape for star-schema cases.) Many
thanks to Alexander Kirkouski for submitting a reproducible test case.
The bug can be demonstrated in all branches back to 9.2 where parameterized
paths were introduced, so back-patch that far.
Given a left join containing a full join in its righthand side, with
the left join's joinclause referencing only one side of the full join
(in a non-strict fashion, so that the full join doesn't get simplified),
the planner could fail with "failed to build any N-way joins" or related
errors. This happened because the full join was seen as overlapping the
left join's RHS, and then recent changes within join_is_legal() caused
that function to conclude that the full join couldn't validly be formed.
Rather than try to rejigger join_is_legal() yet more to allow this,
I think it's better to fix initsplan.c so that the required join order
is explicit in the SpecialJoinInfo data structure. The previous coding
there essentially ignored full joins, relying on the fact that we don't
flatten them in the joinlist data structure to preserve their ordering.
That's sufficient to prevent a wrong plan from being formed, but as this
example shows, it's not sufficient to ensure that the right plan will
be formed. We need to work a bit harder to ensure that the right plan
looks sane according to the SpecialJoinInfos.
Per bug #14105 from Vojtech Rylko. This was apparently induced by
commit 8703059c6 (though now that I've seen it, I wonder whether there
are related cases that could have failed before that); so back-patch
to all active branches. Unfortunately, that patch also went into 9.0,
so this bug is a regression that won't be fixed in that branch.
I've been saying we needed to do this for more than five years, and here it
finally is. This patch removes the ever-growing tangle of spaghetti logic
that grouping_planner() used to use to try to identify the best plan for
post-scan/join query steps. Now, there is (nearly) independent
consideration of each execution step, and entirely separate construction of
Paths to represent each of the possible ways to do that step. We choose
the best Path or set of Paths using the same add_path() logic that's been
used inside query_planner() for years.
In addition, this patch removes the old restriction that subquery_planner()
could return only a single Plan. It now returns a RelOptInfo containing a
set of Paths, just as query_planner() does, and the parent query level can
use each of those Paths as the basis of a SubqueryScanPath at its level.
This allows finding some optimizations that we missed before, wherein a
subquery was capable of returning presorted data and thereby avoiding a
sort in the parent level, making the overall cost cheaper even though
delivering sorted output was not the cheapest plan for the subquery in
isolation. (A couple of regression test outputs change in consequence of
that. However, there is very little change in visible planner behavior
overall, because the point of this patch is not to get immediate planning
benefits but to create the infrastructure for future improvements.)
There is a great deal left to do here. This patch unblocks a lot of
planner work that was basically impractical in the old code structure,
such as allowing FDWs to implement remote aggregation, or rewriting
plan_set_operations() to allow consideration of multiple implementation
orders for set operations. (The latter will likely require a full
rewrite of plan_set_operations(); what I've done here is only to fix it
to return Paths not Plans.) I have also left unfinished some localized
refactoring in createplan.c and planner.c, because it was not necessary
to get this patch to a working state.
Thanks to Robert Haas, David Rowley, and Amit Kapila for review.
Up to now, there's been an assumption that all Paths for a given relation
compute the same output column set (targetlist). However, there are good
reasons to remove that assumption. For example, an indexscan on an
expression index might be able to return the value of an expensive function
"for free". While we have the ability to generate such a plan today in
simple cases, we don't have a way to model that it's cheaper than a plan
that computes the function from scratch, nor a way to create such a plan
in join cases (where the function computation would normally happen at
the topmost join node). Also, we need this so that we can have Paths
representing post-scan/join steps, where the targetlist may well change
from one step to the next. Therefore, invent a "struct PathTarget"
representing the columns we expect a plan step to emit. It's convenient
to include the output tuple width and tlist evaluation cost in this struct,
and there will likely be additional fields in future.
While Path nodes that actually do have custom outputs will need their own
PathTargets, it will still be true that most Paths for a given relation
will compute the same tlist. To reduce the overhead added by this patch,
keep a "default PathTarget" in RelOptInfo, and allow Paths that compute
that column set to just point to their parent RelOptInfo's reltarget.
(In the patch as committed, actually every Path is like that, since we
do not yet have any cases of custom PathTargets.)
I took this opportunity to provide some more-honest costing of
PlaceHolderVar evaluation. Up to now, the assumption that "scan/join
reltargetlists have cost zero" was applied not only to Vars, where it's
reasonable, but also PlaceHolderVars where it isn't. Now, we add the eval
cost of a PlaceHolderVar's expression to the first plan level where it can
be computed, by including it in the PathTarget cost field and adding that
to the cost estimates for Paths. This isn't perfect yet but it's much
better than before, and there is a way forward to improve it more. This
costing change affects the join order chosen for a couple of the regression
tests, changing expected row ordering.
If a GROUP BY clause includes all columns of a non-deferred primary key,
as well as other columns of the same relation, those other columns are
redundant and can be dropped from the grouping; the pkey is enough to
ensure that each row of the table corresponds to a separate group.
Getting rid of the excess columns will reduce the cost of the sorting or
hashing needed to implement GROUP BY, and can indeed remove the need for
a sort step altogether.
This seems worth testing for since many query authors are not aware of
the GROUP-BY-primary-key exception to the rule about queries not being
allowed to reference non-grouped-by columns in their targetlists or
HAVING clauses. Thus, redundant GROUP BY items are not uncommon. Also,
we can make the test pretty cheap in most queries where it won't help
by not looking up a rel's primary key until we've found that at least
two of its columns are in GROUP BY.
David Rowley, reviewed by Julien Rouhaud
Commit 45f6240a8f added an assumption in ExecHashIncreaseNumBatches
and ExecHashIncreaseNumBuckets that they could find all tuples in the main
hash table by iterating over the "dense storage" introduced by that patch.
However, ExecHashRemoveNextSkewBucket continued its old practice of simply
re-linking deleted skew tuples into the main table's hashchains. Hence,
such tuples got lost during any subsequent increase in nbatch or nbuckets,
and would never get joined, as reported in bug #13908 from Seth P.
I (tgl) think that the aforesaid commit has got multiple design issues
and should be reworked rather completely; but there is no time for that
right now, so band-aid the problem by making ExecHashRemoveNextSkewBucket
physically copy deleted skew tuples into the "dense storage" arena.
The added test case is able to exhibit the problem by means of fooling the
planner with a WHERE condition that it will underestimate the selectivity
of, causing the initial nbatch estimate to be too small.
Tomas Vondra and Tom Lane. Thanks to David Johnston for initial
investigation into the bug report.
More fuzz testing by Andreas Seltenreich exposed that the planner did not
cope well with chains of lateral references. If relation X references Y
laterally, and Y references Z laterally, then we will have to scan X on the
inside of a nestloop with Z, so for all intents and purposes X is laterally
dependent on Z too. The planner did not understand this and would generate
intermediate joins that could not be used. While that was usually harmless
except for wasting some planning cycles, under the right circumstances it
would lead to "failed to build any N-way joins" or "could not devise a
query plan" planner failures.
To fix that, convert the existing per-relation lateral_relids and
lateral_referencers relid sets into their transitive closures; that is,
they now show all relations on which a rel is directly or indirectly
laterally dependent. This not only fixes the chained-reference problem
but allows some of the relevant tests to be made substantially simpler
and faster, since they can be reduced to simple bitmap manipulations
instead of searches of the LateralJoinInfo list.
Also, when a PlaceHolderVar that is due to be evaluated at a join contains
lateral references, we should treat those references as indirect lateral
dependencies of each of the join's base relations. This prevents us from
trying to join any individual base relations to the lateral reference
source before the join is formed, which again cannot work.
Andreas' testing also exposed another oversight in the "dangerous
PlaceHolderVar" test added in commit 85e5e222b1. Simply rejecting
unsafe join paths in joinpath.c is insufficient, because in some cases
we will end up rejecting *all* possible paths for a particular join, again
leading to "could not devise a query plan" failures. The restriction has
to be known also to join_is_legal and its cohort functions, so that they
will not select a join for which that will happen. I chose to move the
supporting logic into joinrels.c where the latter functions are.
Back-patch to 9.3 where LATERAL support was introduced.
It was possible for the planner to decide to join a LATERAL subquery to
the outer side of an outer join before the outer join itself is completed.
Normally that's fine because of the associativity rules, but it doesn't
work if the subquery contains a lateral reference to the inner side of the
outer join. In such a situation the outer join *must* be done first.
join_is_legal() missed this consideration and would allow the join to be
attempted, but the actual path-building code correctly decided that no
valid join path could be made, sometimes leading to planner errors such as
"failed to build any N-way joins".
Per report from Andreas Seltenreich. Back-patch to 9.3 where LATERAL
support was added.
David Rowley
Peter Eisentraut
Tom Lane
Peter Geoghegan
Andres Freund
Robert Haas
Heikki Linnakangas