In 5f32b29c18 I changed the creation of HashState.hashkeys to
actually use HashState as the parent (instead of HashJoinState, which
was incorrect, as they were executed below HashState), to fix the
problem of hashkeys expressions otherwise relying on slot types
appropriate for HashJoinState, rather than HashState as would be
correct. That reliance was only introduced in 12, which is why it
previously worked to use HashJoinState as the parent (although I'd be
unsurprised if there were problematic cases).
Unfortunately that's not a sufficient solution, because before this
commit, the to-be-hashed expressions referenced inner/outer as
appropriate for the HashJoin, not Hash. That didn't have obvious bad
consequences, because the slots containing the tuples were put into
ecxt_innertuple when hashing a tuple for HashState (even though Hash
doesn't have an inner plan).
There are less common cases where this can cause visible problems
however (rather than just confusion when inspecting such executor
trees). E.g. "ERROR: bogus varno: 65000", when explaining queries
containing a HashJoin where the subsidiary Hash node's hash keys
reference a subplan. While normally hashkeys aren't displayed by
EXPLAIN, if one of those expressions references a subplan, that
subplan may be printed as part of the Hash node - which then failed
because an inner plan was referenced, and Hash doesn't have that.
It seems quite possible that there's other broken cases, too.
Fix the problem by properly splitting the expression for the HashJoin
and Hash nodes at plan time, and have them reference the proper
subsidiary node. While other workarounds are possible, fixing this
correctly seems easy enough. It was a pretty ugly hack to have
ExecInitHashJoin put the expression into the already initialized
HashState, in the first place.
I decided to not just split inner/outer hashkeys inside
make_hashjoin(), but also to separate out hashoperators and
hashcollations at plan time. Otherwise we would have ended up having
two very similar loops, one at plan time and the other during executor
startup. The work seems to more appropriately belong to plan time,
anyway.
Reported-By: Nikita Glukhov, Alexander Korotkov
Author: Andres Freund
Reviewed-By: Tom Lane, in an earlier version
Discussion: https://postgr.es/m/CAPpHfdvGVegF_TKKRiBrSmatJL2dR9uwFCuR+teQ_8tEXU8mxg@mail.gmail.com
Backpatch: 12-
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
3373c7155 changed how we determine EquivalenceClasses for relations and
added an Assert to ensure all relations mentioned in each EC's ec_relids
was a RELOPT_BASEREL. However, the join removal code may remove a LEFT
JOIN and since it does not clean up EC members belonging to the removed
relations it can leave RELOPT_DEADREL rels in ec_relids.
Fix this by adjusting the Assert to allow RELOPT_DEADREL rels too.
Reported-by: sqlsmith via Andreas Seltenreich
Discussion: https://postgr.es/m/87y30r8sls.fsf@ansel.ydns.eu
Previously in order to determine which ECs a relation had members in, we
had to loop over all ECs stored in PlannerInfo's eq_classes and check if
ec_relids mentioned the relation. For the most part, this was fine, as
generally, unless queries were fairly complex, the overhead of performing
the lookup would have not been that significant. However, when queries
contained large numbers of joins and ECs, the overhead to find the set of
classes matching a given set of relations could become a significant
portion of the overall planning effort.
Here we allow a much more efficient method to access the ECs which match a
given relation or set of relations. A new Bitmapset field in RelOptInfo
now exists to store the indexes into PlannerInfo's eq_classes list which
each relation is mentioned in. This allows very fast lookups to find all
ECs belonging to a single relation. When we need to lookup ECs belonging
to a given pair of relations, we can simply bitwise-AND the Bitmapsets from
each relation and use the result to perform the lookup.
We also take the opportunity to write a new implementation of
generate_join_implied_equalities which makes use of the new indexes.
generate_join_implied_equalities_for_ecs must remain as is as it can be
given a custom list of ECs, which we can't easily determine the indexes of.
This was originally intended to fix the performance penalty of looking up
foreign keys matching a join condition which was introduced by 100340e2d.
However, we're speeding up much more than just that here.
Author: David Rowley, Tom Lane
Reviewed-by: Tom Lane, Tomas Vondra
Discussion: https://postgr.es/m/6970.1545327857@sss.pgh.pa.us
Formerly, lcons was about the same speed as lappend, but with the new
List implementation, that's not so; with a long List, data movement
imposes an O(N) cost on lcons and list_delete_first, but not lappend.
Hence, invent list_delete_last with semantics parallel to
list_delete_first (but O(1) cost), and change various places to use
lappend and list_delete_last where this can be done without much
violence to the code logic.
There are quite a few places that construct result lists using lcons not
lappend. Some have semantic rationales for that; I added comments about
it to a couple that didn't have them already. In many such places though,
I think the coding is that way only because back in the dark ages lcons
was faster than lappend. Hence, switch to lappend where this can be done
without causing semantic changes.
In ExecInitExprRec(), this results in aggregates and window functions that
are in the same plan node being executed in a different order than before.
Generally, the executions of such functions ought to be independent of
each other, so this shouldn't result in visibly different query results.
But if you push it, as one regression test case does, you can show that
the order is different. The new order seems saner; it's closer to
the order of the functions in the query text. And we never documented
or promised anything about this, anyway.
Also, in gistfinishsplit(), don't bother building a reverse-order list;
it's easy now to iterate backwards through the original list.
It'd be possible to go further towards removing uses of lcons and
list_delete_first, but it'd require more extensive logic changes,
and I'm not convinced it's worth it. Most of the remaining uses
deal with queues that probably never get long enough to be worth
sweating over. (Actually, I doubt that any of the changes in this
patch will have measurable performance effects either. But better
to have good examples than bad ones in the code base.)
Patch by me, thanks to David Rowley and Daniel Gustafsson for review.
Discussion: https://postgr.es/m/21272.1563318411@sss.pgh.pa.us
It seems worth getting rid of these functions because they require the
caller to retain a ListCell pointer into a List that it's modifying,
which is a dangerous practice with the new List implementation.
(The only other List-modifying function that takes a ListCell pointer
as input is list_delete_cell, which nowadays is preferentially used
via the constrained API foreach_delete_current.)
There was only one remaining caller of these functions after commit
2f5b8eb5a, and that was some fairly ugly GEQO code that can be much
more clearly expressed using a list-index variable and list_insert_nth.
Hence, rewrite that code, and remove the functions.
Discussion: https://postgr.es/m/26193.1563228600@sss.pgh.pa.us
In the wake of commit 1cff1b95a, the obvious way to sort a List
is to apply qsort() directly to the array of ListCells. list_qsort
was building an intermediate array of pointers-to-ListCells, which
we no longer need, but getting rid of it forces an API change:
the comparator functions need to do one less level of indirection.
Since we're having to touch the callers anyway, let's do two additional
changes: sort the given list in-place rather than making a copy (as
none of the existing callers have any use for the copying behavior),
and rename list_qsort to list_sort. It was argued that the old name
exposes more about the implementation than it should, which I find
pretty questionable, but a better reason to rename it is to be sure
we get the attention of any external callers about the need to fix
their comparator functions.
While we're at it, change four existing callers of qsort() to use
list_sort instead; previously, they all had local reinventions
of list_qsort, ie build-an-array-from-a-List-and-qsort-it.
(There are some other places where changing to list_sort perhaps
would be worthwhile, but they're less obviously wins.)
Discussion: https://postgr.es/m/29361.1563220190@sss.pgh.pa.us
This is numbered take 7, and addresses a set of issues around:
- Fixes for typos and incorrect reference names.
- Removal of unneeded comments.
- Removal of unreferenced functions and structures.
- Fixes regarding variable name consistency.
Author: Alexander Lakhin
Discussion: https://postgr.es/m/10bfd4ac-3e7c-40ab-2b2e-355ed15495e8@gmail.com
Originally, Postgres Lists were a more or less exact reimplementation of
Lisp lists, which consist of chains of separately-allocated cons cells,
each having a value and a next-cell link. We'd hacked that once before
(commit d0b4399d8) to add a separate List header, but the data was still
in cons cells. That makes some operations -- notably list_nth() -- O(N),
and it's bulky because of the next-cell pointers and per-cell palloc
overhead, and it's very cache-unfriendly if the cons cells end up
scattered around rather than being adjacent.
In this rewrite, we still have List headers, but the data is in a
resizable array of values, with no next-cell links. Now we need at
most two palloc's per List, and often only one, since we can allocate
some values in the same palloc call as the List header. (Of course,
extending an existing List may require repalloc's to enlarge the array.
But this involves just O(log N) allocations not O(N).)
Of course this is not without downsides. The key difficulty is that
addition or deletion of a list entry may now cause other entries to
move, which it did not before.
For example, that breaks foreach() and sister macros, which historically
used a pointer to the current cons-cell as loop state. We can repair
those macros transparently by making their actual loop state be an
integer list index; the exposed "ListCell *" pointer is no longer state
carried across loop iterations, but is just a derived value. (In
practice, modern compilers can optimize things back to having just one
loop state value, at least for simple cases with inline loop bodies.)
In principle, this is a semantics change for cases where the loop body
inserts or deletes list entries ahead of the current loop index; but
I found no such cases in the Postgres code.
The change is not at all transparent for code that doesn't use foreach()
but chases lists "by hand" using lnext(). The largest share of such
code in the backend is in loops that were maintaining "prev" and "next"
variables in addition to the current-cell pointer, in order to delete
list cells efficiently using list_delete_cell(). However, we no longer
need a previous-cell pointer to delete a list cell efficiently. Keeping
a next-cell pointer doesn't work, as explained above, but we can improve
matters by changing such code to use a regular foreach() loop and then
using the new macro foreach_delete_current() to delete the current cell.
(This macro knows how to update the associated foreach loop's state so
that no cells will be missed in the traversal.)
There remains a nontrivial risk of code assuming that a ListCell *
pointer will remain good over an operation that could now move the list
contents. To help catch such errors, list.c can be compiled with a new
define symbol DEBUG_LIST_MEMORY_USAGE that forcibly moves list contents
whenever that could possibly happen. This makes list operations
significantly more expensive so it's not normally turned on (though it
is on by default if USE_VALGRIND is on).
There are two notable API differences from the previous code:
* lnext() now requires the List's header pointer in addition to the
current cell's address.
* list_delete_cell() no longer requires a previous-cell argument.
These changes are somewhat unfortunate, but on the other hand code using
either function needs inspection to see if it is assuming anything
it shouldn't, so it's not all bad.
Programmers should be aware of these significant performance changes:
* list_nth() and related functions are now O(1); so there's no
major access-speed difference between a list and an array.
* Inserting or deleting a list element now takes time proportional to
the distance to the end of the list, due to moving the array elements.
(However, it typically *doesn't* require palloc or pfree, so except in
long lists it's probably still faster than before.) Notably, lcons()
used to be about the same cost as lappend(), but that's no longer true
if the list is long. Code that uses lcons() and list_delete_first()
to maintain a stack might usefully be rewritten to push and pop at the
end of the list rather than the beginning.
* There are now list_insert_nth...() and list_delete_nth...() functions
that add or remove a list cell identified by index. These have the
data-movement penalty explained above, but there's no search penalty.
* list_concat() and variants now copy the second list's data into
storage belonging to the first list, so there is no longer any
sharing of cells between the input lists. The second argument is
now declared "const List *" to reflect that it isn't changed.
This patch just does the minimum needed to get the new implementation
in place and fix bugs exposed by the regression tests. As suggested
by the foregoing, there's a fair amount of followup work remaining to
do.
Also, the ENABLE_LIST_COMPAT macros are finally removed in this
commit. Code using those should have been gone a dozen years ago.
Patch by me; thanks to David Rowley, Jesper Pedersen, and others
for review.
Discussion: https://postgr.es/m/11587.1550975080@sss.pgh.pa.us
d4c3a156c added code to remove columns that were not part of a table's
PRIMARY KEY constraint from the GROUP BY clause when all the primary key
columns were present in the group by. This is fine to do since we know
that there will only be one row per group coming from this relation.
However, the logic failed to consider inheritance parent relations. These
can have child relations without a primary key, but even if they did, they
could duplicate one of the parent's rows or one from another child
relation. In this case, those additional GROUP BY columns are required.
Fix this by disabling the optimization for inheritance parent tables.
In v11 and beyond, partitioned tables are fine since partitions cannot
overlap and before v11 partitioned tables could not have a primary key.
Reported-by: Manuel Rigger
Discussion: http://postgr.es/m/CA+u7OA7VLKf_vEr6kLF3MnWSA9LToJYncgpNX2tQ-oWzYCBQAw@mail.gmail.com
Backpatch-through: 9.6
The logic in reorder_grouping_sets to order grouping set elements to
match a pre-specified sort ordering was defective, resulting in
unnecessary sort nodes (though the query output would still be
correct). Repair, simplifying the code a little, and add a test.
Per report from Richard Guo, though I didn't use their patch. Original
bug seems to have been my fault.
Backpatch back to 9.5 where grouping sets were introduced.
Discussion: https://postgr.es/m/CAN_9JTzyjGcUjiBHxLsgqfk7PkdLGXiM=pwM+=ph2LsWw0WO1A@mail.gmail.com
WHERE EXISTS (...) queries cannot be executed by Parallel Hash Join
with jointype JOIN_UNIQUE_INNER, because there is no way to make a
partial plan totally unique. The consequence of allowing such plans
was duplicate results from some EXISTS queries.
Back-patch to 11. Bug #15857.
Author: Thomas Munro
Reviewed-by: Tom Lane
Reported-by: Vladimir Kriukov
Discussion: https://postgr.es/m/15857-d1ba2a64bce0795e%40postgresql.org
Since extended statistic got introduced in PostgreSQL 10, there was a
single catalog pg_statistic_ext storing both the definitions and built
statistic. That's however problematic when a user is supposed to have
access only to the definitions, but not to user data.
Consider for example pg_dump on a database with RLS enabled - if the
pg_statistic_ext catalog respects RLS (which it should, if it contains
user data), pg_dump would not see any records and the result would not
define any extended statistics. That would be a surprising behavior.
Until now this was not a pressing issue, because the existing types of
extended statistic (functional dependencies and ndistinct coefficients)
do not include any user data directly. This changed with introduction
of MCV lists, which do include most common combinations of values.
The easiest way to fix this is to split the pg_statistic_ext catalog
into two - one for definitions, one for the built statistic values.
The new catalog is called pg_statistic_ext_data, and we're maintaining
a 1:1 relationship with the old catalog - either there are matching
records in both catalogs, or neither of them.
Bumped CATVERSION due to changing system catalog definitions.
Author: Dean Rasheed, with improvements by me
Reviewed-by: Dean Rasheed, John Naylor
Discussion: https://postgr.es/m/CAEZATCUhT9rt7Ui%3DVdx4N%3D%3DVV5XOK5dsXfnGgVOz_JhAicB%3DZA%40mail.gmail.com
If an EquivalenceClass member expression includes variables from
multiple appendrels, then instead of producing one substituted
expression per child relation as intended, we'd create additional
child expressions for combinations of children of different appendrels.
This happened because the child expressions generated while considering
the first appendrel were taken as sources during substitution of the
second appendrel, and so on. The extra expressions are useless, and are
harmless unless there are too many of them --- but if you have several
appendrels with a thousand or so members each, it gets bad fast.
To fix, consider only original (non-em_is_child) EC members as candidates
to be expanded. This requires the ability to substitute directly from a
top parent relation's Vars to those of an indirect descendant relation,
but we already have that in adjust_appendrel_attrs_multilevel().
Per bug #15847 from Feike Steenbergen. This is a longstanding misbehavior,
but it's only worth worrying about when there are more appendrel children
than we've historically considered wise to use. So I'm not going to take
the risk of back-patching this.
Discussion: https://postgr.es/m/15847-ea3734094bf8ae61@postgresql.org
This is still using the 2.0 version of pg_bsd_indent.
I thought it would be good to commit this separately,
so as to document the differences between 2.0 and 2.1 behavior.
Discussion: https://postgr.es/m/16296.1558103386@sss.pgh.pa.us
create_merge_append_plan failed to honor the CP_EXACT_TLIST flag:
it would generate the expected targetlist but then it felt free to
add resjunk sort targets to it. This demonstrably leads to assertion
failures in v11 and HEAD, and it's probably just accidental that we
don't see the same in older branches. I've not looked into whether
there would be any real-world consequences in non-assert builds.
In HEAD, create_append_plan has sprouted the same problem, so fix
that too (although we do not have any test cases that seem able to
reach that bug). This is an oversight in commit 3fc6e2d7f which
invented the CP_EXACT_TLIST flag, so back-patch to 9.6 where that
came in.
convert_subquery_pathkeys would create pathkeys for subquery output
values if they match any EquivalenceClass known in the outer query
and are available in the subquery's syntactic targetlist. However,
the second part of that condition is wrong, because such values might
not appear in the subquery relation's reltarget list, which would
mean that they couldn't be accessed above the level of the subquery
scan. We must check that they appear in the reltarget list, instead.
This can lead to dropping knowledge about the subquery's sort
ordering, but I believe it's okay, because any sort key that the
outer query actually has any interest in would appear in the
reltarget list.
This second issue is of very long standing, but right now there's no
evidence that it causes observable problems before 9.6, so I refrained
from back-patching further than that. We can revisit that choice if
somebody finds a way to make it cause problems in older branches.
(Developing useful test cases for these issues is really problematic;
fixing convert_subquery_pathkeys removes the only known way to exhibit
the create_merge_append_plan bug, and neither of the test cases added
by this patch causes a problem in all branches, even when considering
the issues separately.)
The second issue explains bug #15795 from Suresh Kumar R ("could not
find pathkey item to sort" with nested DISTINCT queries). I stumbled
across the first issue while investigating that.
Discussion: https://postgr.es/m/15795-fadb56c8e44ee73c@postgresql.org
In commit 7012b132d0, which added support for aggregate pushdown in
postgres_fdw, the expense of evaluating the final scan/join target
computed by make_group_input_target() was not accounted for at all in
costing aggregate pushdown paths with local statistics. The right fix
for this would be to have a separate upper stage to adjust the final
scan/join relation (see comments for apply_scanjoin_target_to_paths());
but for now, fix by adding the tlist eval cost when costing aggregate
pushdown paths with local statistics.
Apply this to HEAD only to avoid destabilizing existing plan choices.
Author: Etsuro Fujita
Reviewed-By: Antonin Houska
Discussion: https://postgr.es/m/5C66A056.60007%40lab.ntt.co.jp
Project style is to check the success of SearchSysCacheN and friends
by applying HeapTupleIsValid to the result. A tiny minority of calls
creatively did it differently. Bring them into line with the rest.
This is just cosmetic, since HeapTupleIsValid is indeed just a null
check at the moment ... but that may not be true forever, and in any
case it puts a mental burden on readers who may wonder why these
call sites are not like the rest.
Back-patch to v11 just to keep the branches in sync. (The bulk of these
errors seem to have originated in v11 or v12, though a few are old.)
Per searching to see if anyplace else had made the same error
repaired in 62148c352.
The interaction of these parameters was a bit confused/confusing,
and in fact v11 entirely misses the opportunity to apply partition
constraints when a partition is accessed directly (rather than
indirectly from its parent).
In HEAD, establish the principle that enable_partition_pruning controls
partition pruning and nothing else. When accessing a partition via its
parent, we do partition pruning (if enabled by enable_partition_pruning)
and then there is no need to consider partition constraints in the
constraint_exclusion logic. When accessing a partition directly, its
partition constraints are applied by the constraint_exclusion logic,
only if constraint_exclusion = on.
In v11, we can't have such a clean division of these GUCs' effects,
partly because we don't want to break compatibility too much in a
released branch, and partly because the clean coding requires
inheritance_planner to have applied partition pruning to a partitioned
target table, which it doesn't in v11. However, we can tweak things
enough to cover the missed case, which seems like a good idea since
it's potentially a performance regression from v10. This patch keeps
v11's previous behavior in which enable_partition_pruning overrides
constraint_exclusion for an inherited target table, though.
In HEAD, also teach relation_excluded_by_constraints that it's okay to use
inheritable constraints when trying to prune a traditional inheritance
tree. This might not be thought worthy of effort given that that feature
is semi-deprecated now, but we have enough infrastructure that it only
takes a couple more lines of code to do it correctly.
Amit Langote and Tom Lane
Discussion: https://postgr.es/m/9813f079-f16b-61c8-9ab7-4363cab28d80@lab.ntt.co.jp
Discussion: https://postgr.es/m/29069.1555970894@sss.pgh.pa.us
This has to be prevented because inlining would result in multiple
self-references, which we don't support (and in fact that's disallowed
by the SQL spec, see statements about linearly vs. nonlinearly
recursive queries). Bug fix for commit 608b167f9.
Per report from Yaroslav Schekin (via Andrew Gierth)
Discussion: https://postgr.es/m/87wolmg60q.fsf@news-spur.riddles.org.uk
join_is_legal() needs to reject forming certain outer joins in cases
where that would lead the planner down a blind alley. However, it
mistakenly supposed that the way to handle full joins was to treat them
as applying the same constraints as for left joins, only to both sides.
That doesn't work, as shown in bug #15741 from Anthony Skorski: given
a lateral reference out of a join that's fully enclosed by a full join,
the code would fail to believe that any join ordering is legal, resulting
in errors like "failed to build any N-way joins".
However, we don't really need to consider full joins at all for this
purpose, because we effectively force them to be evaluated in syntactic
order, and that order is always legal for lateral references. Hence,
get rid of this broken logic for full joins and just ignore them instead.
This seems to have been an oversight in commit 7e19db0c0.
Back-patch to all supported branches, as that was.
Discussion: https://postgr.es/m/15741-276f1f464b3f40eb@postgresql.org
Before those commits, partitioning-related code in the executor could
assume that ModifyTableState.resultRelInfo[] contains only leaf partitions.
However, now a fully-pruned update results in a dummy ModifyTable that
references the root partitioned table, and that breaks some stuff.
In v11, this led to an assertion or core dump in the tuple routing code.
Fix by disabling tuple routing, since we don't need that anyway.
(I chose to do that in HEAD as well for safety, even though the problem
doesn't manifest in HEAD as it stands.)
In v10, this confused ExecInitModifyTable's decision about whether it
needed to close the root table. But we can get rid of that altogether
by being smarter about where to find the root table.
Note that since the referenced commits haven't shipped yet, this
isn't fixing any bug the field has seen.
Amit Langote, per a report from me
Discussion: https://postgr.es/m/20710.1554582479@sss.pgh.pa.us
If we need ordered output from a scan of a partitioned table, but
the ordering matches the partition ordering, then we don't need to
use a MergeAppend to combine the pre-ordered per-partition scan
results: a plain Append will produce the same results. This
both saves useless comparison work inside the MergeAppend proper,
and allows us to start returning tuples after istarting up just
the first child node not all of them.
However, all is not peaches and cream, because if some of the
child nodes have high startup costs then there will be big
discontinuities in the tuples-returned-versus-elapsed-time curve.
The planner's cost model cannot handle that (yet, anyway).
If we model the Append's startup cost as being just the first
child's startup cost, we may drastically underestimate the cost
of fetching slightly more tuples than are available from the first
child. Since we've had bad experiences with over-optimistic choices
of "fast start" plans for ORDER BY LIMIT queries, that seems scary.
As a klugy workaround, set the startup cost estimate for an ordered
Append to be the sum of its children's startup costs (as MergeAppend
would). This doesn't really describe reality, but it's less likely
to cause a bad plan choice than an underestimated startup cost would.
In practice, the cases where we really care about this optimization
will have child plans that are IndexScans with zero startup cost,
so that the overly conservative estimate is still just zero.
David Rowley, reviewed by Julien Rouhaud and Antonin Houska
Discussion: https://postgr.es/m/CAKJS1f-hAqhPLRk_RaSFTgYxd=Tz5hA7kQ2h4-DhJufQk8TGuw@mail.gmail.com
The assertions added by commit b04aeb0a0 exposed that there are some
code paths wherein the executor will try to open an index without
holding any lock on it. We do have some lock on the index's table,
so it seems likely that there's no fatal problem with this (for
instance, the index couldn't get dropped from under us). Still,
it's bad practice and we should fix it.
To do so, remove the optimizations in ExecInitIndexScan and friends
that tried to avoid taking a lock on an index belonging to a target
relation, and just take the lock always. In non-bug cases, this
will result in no additional shared-memory access, since we'll find
in the local lock table that we already have a lock of the desired
type; hence, no significant performance degradation should occur.
Also, adjust the planner and executor so that the type of lock taken
on an index is always identical to the type of lock taken for its table,
by relying on the recently added RangeTblEntry.rellockmode field.
This avoids some corner cases where that might not have been true
before (possibly resulting in extra locking overhead), and prevents
future maintenance issues from having multiple bits of logic that
all needed to be in sync. In addition, this change removes all core
calls to ExecRelationIsTargetRelation, which avoids a possible O(N^2)
startup penalty for queries with large numbers of target relations.
(We'd probably remove that function altogether, were it not that we
advertise it as something that FDWs might want to use.)
Also adjust some places in selfuncs.c to not take any lock on indexes
they are transiently opening, since we can assume that plancat.c
did that already.
In passing, change gin_clean_pending_list() to take RowExclusiveLock
not AccessShareLock on its target index. Although it's not clear that
that's actually a bug, it seemed very strange for a function that's
explicitly going to modify the index to use only AccessShareLock.
David Rowley, reviewed by Julien Rouhaud and Amit Langote,
a bit of further tweaking by me
Discussion: https://postgr.es/m/19465.1541636036@sss.pgh.pa.us
The upper-planner pathification allows FDWs to arrange to push down
different types of upper-stage operations to the remote side. This
commit teaches postgres_fdw to do it for the (FINAL, NULL) upperrel,
which is responsible for doing LockRows, LIMIT, and/or ModifyTable.
This provides the ability for postgres_fdw to handle SELECT commands
so that it 1) skips the LockRows step (if any) (note that this is
safe since it performs early locking) and 2) pushes down the LIMIT
and/or OFFSET restrictions (if any) to the remote side. This doesn't
handle the INSERT/UPDATE/DELETE cases.
Author: Etsuro Fujita
Reviewed-By: Antonin Houska and Jeff Janes
Discussion: https://postgr.es/m/87pnz1aby9.fsf@news-spur.riddles.org.uk
This moves bitmap heap scan support to below an optional tableam
callback. It's optional as the whole concept of bitmap heapscans is
fairly block specific.
This basically moves the work previously done in bitgetpage() into the
new scan_bitmap_next_block callback, and the direct poking into the
buffer done in BitmapHeapNext() into the new scan_bitmap_next_tuple()
callback.
The abstraction is currently somewhat leaky because
nodeBitmapHeapscan.c's prefetching and visibilitymap based logic
remains - it's likely that we'll later have to move more into the
AM. But it's not trivial to do so without introducing a significant
amount of code duplication between the AMs, so that's a project for
later.
Note that now nodeBitmapHeapscan.c and the associated node types are a
bit misnamed. But it's not clear whether renaming wouldn't be a cure
worse than the disease. Either way, that'd be best done in a separate
commit.
Author: Andres Freund
Reviewed-By: Robert Haas (in an older version)
Discussion: https://postgr.es/m/20180703070645.wchpu5muyto5n647@alap3.anarazel.de
We can set this up once and for all in subquery_planner's initial survey
of the flattened rangetable, rather than incrementally adjusting it in
build_simple_rel. The previous approach made it rather hard to reason
about exactly when the value would be available, and we were definitely
using it in some places before the final value was computed.
Noted while fooling around with Amit Langote's patch to delay creation
of inheritance child rels. That didn't break this code, but it made it
even more fragile, IMO.
This just moves the table/matview[/toast] determination of relation
size to a callback, and uses a copy of the existing logic to implement
that callback for heap.
It probably would make sense to also move the index specific logic
into a callback, so the metapage handling (and probably more) can be
index specific. But that's a separate task.
Author: Andres Freund
Discussion: https://postgr.es/m/20180703070645.wchpu5muyto5n647@alap3.anarazel.de
Previously, the planner created RangeTblEntry and RelOptInfo structs
for every partition of a partitioned table, even though many of them
might later be deemed uninteresting thanks to partition pruning logic.
This incurred significant overhead when there are many partitions.
Arrange to postpone creation of these data structures until after
we've processed the query enough to identify restriction quals for
the partitioned table, and then apply partition pruning before not
after creation of each partition's data structures. In this way
we need not open the partition relations at all for partitions that
the planner has no real interest in.
For queries that can be proven at plan time to access only a small
number of partitions, this patch improves the practical maximum
number of partitions from under 100 to perhaps a few thousand.
Amit Langote, reviewed at various times by Dilip Kumar, Jesper Pedersen,
Yoshikazu Imai, and David Rowley
Discussion: https://postgr.es/m/9d7c5112-cb99-6a47-d3be-cf1ee6862a1d@lab.ntt.co.jp
While trying to plan a partitionwise join, we may be faced with cases
where one or both input partitions for a particular segment of the join
have been pruned away. In HEAD and v11, this is problematic because
earlier processing didn't bother to make a pruned RelOptInfo fully
valid. With an upcoming patch to make partition pruning more efficient,
this'll be even more problematic because said RelOptInfo won't exist at
all.
The existing code attempts to deal with this by retroactively making the
RelOptInfo fully valid, but that causes crashes under GEQO because join
planning is done in a short-lived memory context. In v11 we could
probably have fixed this by switching to the planner's main context
while fixing up the RelOptInfo, but that idea doesn't scale well to the
upcoming patch. It would be better not to mess with the base-relation
data structures during join planning, anyway --- that's just a recipe
for order-of-operations bugs.
In many cases, though, we don't actually need the child RelOptInfo,
because if the input is certainly empty then the join segment's result
is certainly empty, so we can skip making a join plan altogether. (The
existing code ultimately arrives at the same conclusion, but only after
doing a lot more work.) This approach works except when the pruned-away
partition is on the nullable side of a LEFT, ANTI, or FULL join, and the
other side isn't pruned. But in those cases the existing code leaves a
lot to be desired anyway --- the correct output is just the result of
the unpruned side of the join, but we were emitting a useless outer join
against a dummy Result. Pending somebody writing code to handle that
more nicely, let's just abandon the partitionwise-join optimization in
such cases.
When the modified code skips making a join plan, it doesn't make a
join RelOptInfo either; this requires some upper-level code to
cope with nulls in part_rels[] arrays. We would have had to have
that anyway after the upcoming patch.
Back-patch to v11 since the crash is demonstrable there.
Discussion: https://postgr.es/m/8305.1553884377@sss.pgh.pa.us
This is an SQL-standard feature that allows creating columns that are
computed from expressions rather than assigned, similar to a view or
materialized view but on a column basis.
This implements one kind of generated column: stored (computed on
write). Another kind, virtual (computed on read), is planned for the
future, and some room is left for it.
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Reviewed-by: Pavel Stehule <pavel.stehule@gmail.com>
Discussion: https://www.postgresql.org/message-id/flat/b151f851-4019-bdb1-699e-ebab07d2f40a@2ndquadrant.com
Introduce a third extended statistic type, supported by the CREATE
STATISTICS command - MCV lists, a generalization of the statistic
already built and used for individual columns.
Compared to the already supported types (n-distinct coefficients and
functional dependencies), MCV lists are more complex, include column
values and allow estimation of much wider range of common clauses
(equality and inequality conditions, IS NULL, IS NOT NULL etc.).
Similarly to the other types, a new pseudo-type (pg_mcv_list) is used.
Author: Tomas Vondra
Reviewed-by: Dean Rasheed, David Rowley, Mark Dilger, Alvaro Herrera
Discussion: https://postgr.es/m/dfdac334-9cf2-2597-fb27-f0fb3753f435@2ndquadrant.com
In the dim past, the planner kept the fully-processed version of the query
targetlist (the result of preprocess_targetlist) in grouping_planner's
local variable "tlist", and only grudgingly passed it to individual other
routines as needed. Later we discovered a need to still have it available
after grouping_planner finishes, and invented the root->processed_tlist
field for that purpose, but it wasn't used internally to grouping_planner;
the tlist was still being passed around separately in the same places as
before.
Now comes a proposed patch to allow appendrel expansion to add entries
to the processed tlist, well after preprocess_targetlist has finished
its work. To avoid having to pass around the tlist explicitly, it's
proposed to allow appendrel expansion to modify root->processed_tlist.
That makes aliasing the tlist with assorted parameters and local
variables really scary. It would accidentally work as long as the
tlist is initially nonempty, because then the List header won't move
around, but it's not exactly hard to think of ways for that to break.
Aliased values are poor programming practice anyway.
Hence, get rid of local variables and parameters that can be identified
with root->processed_tlist, in favor of just using that field directly.
And adjust comments to match. (Some of the new comments speak as though
it's already possible for appendrel expansion to modify the tlist; that's
not true yet, but will happen in a later patch.)
Discussion: https://postgr.es/m/9d7c5112-cb99-6a47-d3be-cf1ee6862a1d@lab.ntt.co.jp
Up to now, otherrel RelOptInfos were built at the same time as baserel
RelOptInfos, thanks to recursion in build_simple_rel(). However,
nothing in query_planner's preprocessing cares at all about otherrels,
only baserels, so we don't really need to build them until just before
we enter make_one_rel. This has two benefits:
* create_lateral_join_info did a lot of extra work to propagate
lateral-reference information from parents to the correct children.
But if we delay creation of the children till after that, it's
trivial (and much harder to break, too).
* Since we have all the restriction quals correctly assigned to
parent appendrels by this point, it'll be possible to do plan-time
pruning and never make child RelOptInfos at all for partitions that
can be pruned away. That's not done here, but will be later on.
Amit Langote, reviewed at various times by Dilip Kumar, Jesper Pedersen,
Yoshikazu Imai, and David Rowley
Discussion: https://postgr.es/m/9d7c5112-cb99-6a47-d3be-cf1ee6862a1d@lab.ntt.co.jp
We've been creating duplicate RTEs for partitioned tables just
because we do so for regular inheritance parent tables. But unlike
regular-inheritance parents which are themselves regular tables
and thus need to be scanned, partitioned tables don't need the
extra RTE.
This makes the conditions for building a child RTE the same as those
for building an AppendRelInfo, allowing minor simplification in
expand_single_inheritance_child. Since the planner's actual processing
is driven off the AppendRelInfo list, nothing much changes beyond that,
we just have one fewer useless RTE entry.
Amit Langote, reviewed and hacked a bit by me
Discussion: https://postgr.es/m/9d7c5112-cb99-6a47-d3be-cf1ee6862a1d@lab.ntt.co.jp
If there's only one child relation, the Append or MergeAppend isn't
doing anything useful, and can be elided. It does have a purpose
during planning though, which is to serve as a buffer between parent
and child Var numbering. Therefore we keep it all the way through
to setrefs.c, and get rid of it only after fixing references in the
plan level(s) above it. This works largely the same as setrefs.c's
ancient hack to get rid of no-op SubqueryScan nodes, and can even
share some code with that.
Note the change to make setrefs.c use apply_tlist_labeling rather than
ad-hoc code. This has the effect of propagating the child's resjunk
and ressortgroupref labels, which formerly weren't propagated when
removing a SubqueryScan. Doing that is demonstrably necessary for
the [Merge]Append cases, and seems harmless for SubqueryScan, if only
because trivial_subqueryscan is afraid to collapse cases where the
resjunk marking differs. (I suspect that restriction could now be
removed, though it's unclear that it'd make any new matches possible,
since the outer query can't have references to a child resjunk column.)
David Rowley, reviewed by Alvaro Herrera and Tomas Vondra
Discussion: https://postgr.es/m/CAKJS1f_7u8ATyJ1JGTMHFoKDvZdeF-iEBhs+sM_SXowOr9cArg@mail.gmail.com
I (tgl) remain dubious that it's a good idea for PartitionDirectory
to hold a pin on a relcache entry throughout planning, rather than
copying the data or using some kind of refcount scheme. However, it's
certainly the responsibility of the PartitionDirectory code to ensure
that what it's handing back is a stable data structure, not that of
its caller. So this is a pretty clear oversight in commit 898e5e329,
and one that can cost a lot of performance when there are many
partitions.
Amit Langote (extracted from a much larger patch set)
Discussion: https://postgr.es/m/CA+TgmoY3bRmGB6-DUnoVy5fJoreiBJ43rwMrQRCdPXuKt4Ykaw@mail.gmail.com
Discussion: https://postgr.es/m/9d7c5112-cb99-6a47-d3be-cf1ee6862a1d@lab.ntt.co.jp
This adds a flag "deterministic" to collations. If that is false,
such a collation disables various optimizations that assume that
strings are equal only if they are byte-wise equal. That then allows
use cases such as case-insensitive or accent-insensitive comparisons
or handling of strings with different Unicode normal forms.
This functionality is only supported with the ICU provider. At least
glibc doesn't appear to have any locales that work in a
nondeterministic way, so it's not worth supporting this for the libc
provider.
The term "deterministic comparison" in this context is from Unicode
Technical Standard #10
(https://unicode.org/reports/tr10/#Deterministic_Comparison).
This patch makes changes in three areas:
- CREATE COLLATION DDL changes and system catalog changes to support
this new flag.
- Many executor nodes and auxiliary code are extended to track
collations. Previously, this code would just throw away collation
information, because the eventually-called user-defined functions
didn't use it since they only cared about equality, which didn't
need collation information.
- String data type functions that do equality comparisons and hashing
are changed to take the (non-)deterministic flag into account. For
comparison, this just means skipping various shortcuts and tie
breakers that use byte-wise comparison. For hashing, we first need
to convert the input string to a canonical "sort key" using the ICU
analogue of strxfrm().
Reviewed-by: Daniel Verite <daniel@manitou-mail.org>
Reviewed-by: Peter Geoghegan <pg@bowt.ie>
Discussion: https://www.postgresql.org/message-id/flat/1ccc668f-4cbc-0bef-af67-450b47cdfee7@2ndquadrant.com
Previously, the SERIALIZABLE isolation level prevented parallel query
from being used. Allow the two features to be used together by
sharing the leader's SERIALIZABLEXACT with parallel workers.
An extra per-SERIALIZABLEXACT LWLock is introduced to make it safe to
share, and new logic is introduced to coordinate the early release
of the SERIALIZABLEXACT required for the SXACT_FLAG_RO_SAFE
optimization, as follows:
The first backend to observe the SXACT_FLAG_RO_SAFE flag (set by
some other transaction) will 'partially release' the SERIALIZABLEXACT,
meaning that the conflicts and locks it holds are released, but the
SERIALIZABLEXACT itself will remain active because other backends
might still have a pointer to it.
Whenever any backend notices the SXACT_FLAG_RO_SAFE flag, it clears
its own MySerializableXact variable and frees local resources so that
it can skip SSI checks for the rest of the transaction. In the
special case of the leader process, it transfers the SERIALIZABLEXACT
to a new variable SavedSerializableXact, so that it can be completely
released at the end of the transaction after all workers have exited.
Remove the serializable_okay flag added to CreateParallelContext() by
commit 9da0cc35, because it's now redundant.
Author: Thomas Munro
Reviewed-by: Haribabu Kommi, Robert Haas, Masahiko Sawada, Kevin Grittner
Discussion: https://postgr.es/m/CAEepm=0gXGYhtrVDWOTHS8SQQy_=S9xo+8oCxGLWZAOoeJ=yzQ@mail.gmail.com
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
In commit 960df2a971 ("Correctly assess parallel-safety of tlists when
SRFs are used."), the testing of scan/join target was done incorrectly,
which caused a plan-quality problem. Backpatch through to v11 where
the aforementioned commit went in, since this is a regression from v10.
Author: Etsuro Fujita
Reviewed-by: Robert Haas and Tom Lane
Discussion: https://postgr.es/m/5C75303E.8020303@lab.ntt.co.jp
When we introduced separate ProjectSetPath nodes for application of
set-returning functions in v10, we inadvertently broke some cases where
we're supposed to recognize that the result of a subquery is known to be
empty (contain zero rows). That's because IS_DUMMY_REL was just looking
for a childless AppendPath without allowing for a ProjectSetPath being
possibly stuck on top. In itself, this didn't do anything much worse
than produce slightly worse plans for some corner cases.
Then in v11, commit 11cf92f6e rearranged things to allow the scan/join
targetlist to be applied directly to partial paths before they get
gathered. But it inserted a short-circuit path for dummy relations
that was a little too short: it failed to insert a ProjectSetPath node
at all for a targetlist containing set-returning functions, resulting in
bogus "set-valued function called in context that cannot accept a set"
errors, as reported in bug #15669 from Madelaine Thibaut.
The best way to fix this mess seems to be to reimplement IS_DUMMY_REL
so that it drills down through any ProjectSetPath nodes that might be
there (and it seems like we'd better allow for ProjectionPath as well).
While we're at it, make it look at rel->pathlist not cheapest_total_path,
so that it gives the right answer independently of whether set_cheapest
has been done lately. That dependency looks pretty shaky in the context
of code like apply_scanjoin_target_to_paths, and even if it's not broken
today it'd certainly bite us at some point. (Nastily, unsafe use of the
old coding would almost always work; the hazard comes down to possibly
looking through a dangling pointer, and only once in a blue moon would
you find something there that resulted in the wrong answer.)
It now looks like it was a mistake for IS_DUMMY_REL to be a macro: if
there are any extensions using it, they'll continue to use the old
inadequate logic until they're recompiled, after which they'll fail
to load into server versions predating this fix. Hopefully there are
few such extensions.
Having fixed IS_DUMMY_REL, the special path for dummy rels in
apply_scanjoin_target_to_paths is unnecessary as well as being wrong,
so we can just drop it.
Also change a few places that were testing for partitioned-ness of a
planner relation but not using IS_PARTITIONED_REL for the purpose; that
seems unsafe as well as inconsistent, plus it required an ugly hack in
apply_scanjoin_target_to_paths.
In passing, save a few cycles in apply_scanjoin_target_to_paths by
skipping processing of pre-existing paths for partitioned rels,
and do some cosmetic cleanup and comment adjustment in that function.
I renamed IS_DUMMY_PATH to IS_DUMMY_APPEND with the intention of breaking
any code that might be using it, since in almost every case that would
be wrong; IS_DUMMY_REL is what to be using instead.
In HEAD, also make set_dummy_rel_pathlist static (since it's no longer
used from outside allpaths.c), and delete is_dummy_plan, since it's no
longer used anywhere.
Back-patch as appropriate into v11 and v10.
Tom Lane and Julien Rouhaud
Discussion: https://postgr.es/m/15669-02fb3296cca26203@postgresql.org
We still require AccessExclusiveLock on the partition itself, because
otherwise an insert that violates the newly-imposed partition
constraint could be in progress at the same time that we're changing
that constraint; only the lock level on the parent relation is
weakened.
To make this safe, we have to cope with (at least) three separate
problems. First, relevant DDL might commit while we're in the process
of building a PartitionDesc. If so, find_inheritance_children() might
see a new partition while the RELOID system cache still has the old
partition bound cached, and even before invalidation messages have
been queued. To fix that, if we see that the pg_class tuple seems to
be missing or to have a null relpartbound, refetch the value directly
from the table. We can't get the wrong value, because DETACH PARTITION
still requires AccessExclusiveLock throughout; if we ever want to
change that, this will need more thought. In testing, I found it quite
difficult to hit even the null-relpartbound case; the race condition
is extremely tight, but the theoretical risk is there.
Second, successive calls to RelationGetPartitionDesc might not return
the same answer. The query planner will get confused if lookup up the
PartitionDesc for a particular relation does not return a consistent
answer for the entire duration of query planning. Likewise, query
execution will get confused if the same relation seems to have a
different PartitionDesc at different times. Invent a new
PartitionDirectory concept and use it to ensure consistency. This
ensures that a single invocation of either the planner or the executor
sees the same view of the PartitionDesc from beginning to end, but it
does not guarantee that the planner and the executor see the same
view. Since this allows pointers to old PartitionDesc entries to
survive even after a relcache rebuild, also postpone removing the old
PartitionDesc entry until we're certain no one is using it.
For the most part, it seems to be OK for the planner and executor to
have different views of the PartitionDesc, because the executor will
just ignore any concurrently added partitions which were unknown at
plan time; those partitions won't be part of the inheritance
expansion, but invalidation messages will trigger replanning at some
point. Normally, this happens by the time the very next command is
executed, but if the next command acquires no locks and executes a
prepared query, it can manage not to notice until a new transaction is
started. We might want to tighten that up, but it's material for a
separate patch. There would still be a small window where a query
that started just after an ATTACH PARTITION command committed might
fail to notice its results -- but only if the command starts before
the commit has been acknowledged to the user. All in all, the warts
here around serializability seem small enough to be worth accepting
for the considerable advantage of being able to add partitions without
a full table lock.
Although in general the consequences of new partitions showing up
between planning and execution are limited to the query not noticing
the new partitions, run-time partition pruning will get confused in
that case, so that's the third problem that this patch fixes.
Run-time partition pruning assumes that indexes into the PartitionDesc
are stable between planning and execution. So, add code so that if
new partitions are added between plan time and execution time, the
indexes stored in the subplan_map[] and subpart_map[] arrays within
the plan's PartitionedRelPruneInfo get adjusted accordingly. There
does not seem to be a simple way to generalize this scheme to cope
with partitions that are removed, mostly because they could then get
added back again with different bounds, but it works OK for added
partitions.
This code does not try to ensure that every backend participating in
a parallel query sees the same view of the PartitionDesc. That
currently doesn't matter, because we never pass PartitionDesc
indexes between backends. Each backend will ignore the concurrently
added partitions which it notices, and it doesn't matter if different
backends are ignoring different sets of concurrently added partitions.
If in the future that matters, for example because we allow writes in
parallel query and want all participants to do tuple routing to the same
set of partitions, the PartitionDirectory concept could be improved to
share PartitionDescs across backends. There is a draft patch to
serialize and restore PartitionDescs on the thread where this patch
was discussed, which may be a useful place to start.
Patch by me. Thanks to Alvaro Herrera, David Rowley, Simon Riggs,
Amit Langote, and Michael Paquier for discussion, and to Alvaro
Herrera for some review.
Discussion: http://postgr.es/m/CA+Tgmobt2upbSocvvDej3yzokd7AkiT+PvgFH+a9-5VV1oJNSQ@mail.gmail.com
Discussion: http://postgr.es/m/CA+TgmoZE0r9-cyA-aY6f8WFEROaDLLL7Vf81kZ8MtFCkxpeQSw@mail.gmail.com
Discussion: http://postgr.es/m/CA+TgmoY13KQZF-=HNTrt9UYWYx3_oYOQpu9ioNT49jGgiDpUEA@mail.gmail.com
Michael Paquier
Andres Freund
Tom Lane
David Rowley
Andrew Gierth
Thomas Munro
Tomas Vondra
Etsuro Fujita
Peter Eisentraut
Robert Haas