Commit 9e8da0f7 taught nbtree to handle ScalarArrayOpExpr quals
natively. This works by pushing down the full context (the array keys)
to the nbtree index AM, enabling it to execute multiple primitive index
scans that the planner treats as one continuous index scan/index path.
This earlier enhancement enabled nbtree ScalarArrayOp index-only scans.
It also allowed scans with ScalarArrayOp quals to return ordered results
(with some notable restrictions, described further down).
Take this general approach a lot further: teach nbtree SAOP index scans
to decide how to execute ScalarArrayOp scans (when and where to start
the next primitive index scan) based on physical index characteristics.
This can be far more efficient. All SAOP scans will now reliably avoid
duplicative leaf page accesses (just like any other nbtree index scan).
SAOP scans whose array keys are naturally clustered together now require
far fewer index descents, since we'll reliably avoid starting a new
primitive scan just to get to a later offset from the same leaf page.
The scan's arrays now advance using binary searches for the array
element that best matches the next tuple's attribute value. Required
scan key arrays (i.e. arrays from scan keys that can terminate the scan)
ratchet forward in lockstep with the index scan. Non-required arrays
(i.e. arrays from scan keys that can only exclude non-matching tuples)
"advance" without the process ever rolling over to a higher-order array.
Naturally, only required SAOP scan keys trigger skipping over leaf pages
(non-required arrays cannot safely end or start primitive index scans).
Consequently, even index scans of a composite index with a high-order
inequality scan key (which we'll mark required) and a low-order SAOP
scan key (which we won't mark required) now avoid repeating leaf page
accesses -- that benefit isn't limited to simpler equality-only cases.
In general, all nbtree index scans now output tuples as if they were one
continuous index scan -- even scans that mix a high-order inequality
with lower-order SAOP equalities reliably output tuples in index order.
This allows us to remove a couple of special cases that were applied
when building index paths with SAOP clauses during planning.
Bugfix commit 807a40c5 taught the planner to avoid generating unsafe
path keys: path keys on a multicolumn index path, with a SAOP clause on
any attribute beyond the first/most significant attribute. These cases
are now all safe, so we go back to generating path keys without regard
for the presence of SAOP clauses (just like with any other clause type).
Affected queries can now exploit scan output order in all the usual ways
(e.g., certain "ORDER BY ... LIMIT n" queries can now terminate early).
Also undo changes from follow-up bugfix commit a4523c5a, which taught
the planner to produce alternative index paths, with path keys, but
without low-order SAOP index quals (filter quals were used instead).
We'll no longer generate these alternative paths, since they can no
longer offer any meaningful advantages over standard index qual paths.
Affected queries thereby avoid all of the disadvantages that come from
using filter quals within index scan nodes. They can avoid extra heap
page accesses from using filter quals to exclude non-matching tuples
(index quals will never have that problem). They can also skip over
irrelevant sections of the index in more cases (though only when nbtree
determines that starting another primitive scan actually makes sense).
There is a theoretical risk that removing restrictions on SAOP index
paths from the planner will break compatibility with amcanorder-based
index AMs maintained as extensions. Such an index AM could have the
same limitations around ordered SAOP scans as nbtree had up until now.
Adding a pro forma incompatibility item about the issue to the Postgres
17 release notes seems like a good idea.
Author: Peter Geoghegan <pg@bowt.ie>
Author: Matthias van de Meent <boekewurm+postgres@gmail.com>
Reviewed-By: Heikki Linnakangas <hlinnaka@iki.fi>
Reviewed-By: Matthias van de Meent <boekewurm+postgres@gmail.com>
Reviewed-By: Tomas Vondra <tomas.vondra@enterprisedb.com>
Discussion: https://postgr.es/m/CAH2-Wz=ksvN_sjcnD1+Bt-WtifRA5ok48aDYnq3pkKhxgMQpcw@mail.gmail.com
Since commit 599b33b94, this function assumed that every RTE_RELATION
RangeTblEntry would have an associated RelOptInfo. But that's not so:
we only build RelOptInfos for relations that are scanned by the query.
In particular the target of an INSERT won't have one, so that Vars
appearing in an INSERT ... RETURNING list will not have an associated
RelOptInfo. This apparently wasn't a problem before commit f7816aec2
taught examine_simple_variable() to drill down into CTEs containing
INSERT RETURNING, but it is now.
To fix, add a fallback code path that gets the userid to use directly
from the RTEPermissionInfo associated with the RTE. (Sadly, we must
have two code paths, because not every RTE has a RTEPermissionInfo
either.)
Per report from Alexander Lakhin. No back-patch, since the case is
apparently unreachable before f7816aec2.
Discussion: https://postgr.es/m/608a4886-6c60-0f9e-97d5-591256bd4150@gmail.com
If we have DECHIST statistics about the argument expression, use
the average number of distinct elements as the array length estimate.
(It'd be better to use the average total number of elements, but
that is not currently calculated by compute_array_stats(), and
it's unclear that it'd be worth extra effort to get.)
To do this, we have to change the signature of estimate_array_length
to pass the "root" pointer. While at it, also change its result
type to "double". That's probably not really necessary, but it
avoids any risk of overflow of the value extracted from DECHIST.
All existing callers are going to use the result in a "double"
calculation anyway.
Paul Jungwirth, reviewed by Jian He and myself
Discussion: https://postgr.es/m/CA+renyUnM2d+SmrxKpDuAdpiq6FOM=FByvi6aS6yi__qyf6j9A@mail.gmail.com
00b41463c adjusted Bitmapset so that an empty set is always represented
as NULL. This makes checking for empty sets far cheaper than it used
to be.
There were various places in the code where we'd call bms_membership()
to handle the 3 possible BMS_Membership values. For the BMS_SINGLETON
case, we'd also call bms_singleton_member() to find the single set member.
This can now be done in a more optimal way by first checking if the set is
NULL and then not bothering with bms_membership() and simply call
bms_get_singleton_member() instead to find the single member. This
function will return false if there are multiple members in the set.
Here we also tidy up some logic in examine_variable() for the single
member case. There's now no need to call bms_is_member() as we've
already established that we're working with a singleton Bitmapset, so we
can just check if varRelid matches the singleton member.
Reviewed-by: Richard Guo
Discussion: https://postgr.es/m/CAApHDvqW+CxNPcY245GaWiuqkkqgTudtG2ncGvvSjGn2wdTZLA@mail.gmail.com
examine_simple_variable() left this as an unimplemented case years
ago, with the result that plans for queries involving un-flattened
CTEs might be much stupider than necessary. It's not hard to extend
the existing logic for RTE_SUBQUERY cases to also be able to drill
down into CTEs, so let's do that.
There was some discussion of whether this patch breaks the idea
of a MATERIALIZED CTE being an optimization fence. We concluded
it's okay, because we already allow the outer planner level to
see the estimated width and rowcount of the CTE result, and
letting it see column statistics too seems fairly equivalent.
Basically, what we expect of the optimization fence is that the
outer query should not affect the plan chosen for the CTE query.
Once that plan is chosen, it's okay for the outer planner level
to make use of whatever information we have about it.
Jian Guo and Tom Lane, per complaint from Hans Buschmann
Discussion: https://postgr.es/m/4504e67078d648cdac3651b2960da6e7@nidsa.net
This adds support for infinity to the interval data type, using the
same input/output representation as the other date/time data types
that support infinity. This allows various arithmetic operations on
infinite dates, timestamps and intervals.
The new values are represented by setting all fields of the interval
to INT32/64_MIN for -infinity, and INT32/64_MAX for +infinity. This
ensures that they compare as less/greater than all other interval
values, without the need for any special-case comparison code.
Note that, since those 2 values were formerly accepted as legal finite
intervals, pg_upgrade and dump/restore from an old database will turn
them from finite to infinite intervals. That seems OK, since those
exact values should be extremely rare in practice, and they are
outside the documented range supported by the interval type, which
gives us a certain amount of leeway.
Bump catalog version.
Joseph Koshakow, Jian He, and Ashutosh Bapat, reviewed by me.
Discussion: https://postgr.es/m/CAAvxfHea4%2BsPybKK7agDYOMo9N-Z3J6ZXf3BOM79pFsFNcRjwA%40mail.gmail.com
Traditionally we used the same Var struct to represent the value
of a table column everywhere in parse and plan trees. This choice
predates our support for SQL outer joins, and it's really a pretty
bad idea with outer joins, because the Var's value can depend on
where it is in the tree: it might go to NULL above an outer join.
So expression nodes that are equal() per equalfuncs.c might not
represent the same value, which is a huge correctness hazard for
the planner.
To improve this, decorate Var nodes with a bitmapset showing
which outer joins (identified by RTE indexes) may have nulled
them at the point in the parse tree where the Var appears.
This allows us to trust that equal() Vars represent the same value.
A certain amount of klugery is still needed to cope with cases
where we re-order two outer joins, but it's possible to make it
work without sacrificing that core principle. PlaceHolderVars
receive similar decoration for the same reason.
In the planner, we include these outer join bitmapsets into the relids
that an expression is considered to depend on, and in consequence also
add outer-join relids to the relids of join RelOptInfos. This allows
us to correctly perceive whether an expression can be calculated above
or below a particular outer join.
This change affects FDWs that want to plan foreign joins. They *must*
follow suit when labeling foreign joins in order to match with the
core planner, but for many purposes (if postgres_fdw is any guide)
they'd prefer to consider only base relations within the join.
To support both requirements, redefine ForeignScan.fs_relids as
base+OJ relids, and add a new field fs_base_relids that's set up by
the core planner.
Large though it is, this commit just does the minimum necessary to
install the new mechanisms and get check-world passing again.
Follow-up patches will perform some cleanup. (The README additions
and comments mention some stuff that will appear in the follow-up.)
Patch by me; thanks to Richard Guo for review.
Discussion: https://postgr.es/m/830269.1656693747@sss.pgh.pa.us
RelOptInfo.userid is the same for all relations in a given inheritance
tree, so the code in examine_variable() and example_simple_variable()
that repeats the ACL checks on the root parent rel instead of a given
leaf child relations need not recompute userid too.
Author: Amit Langote <amitlangote09@gmail.com>
Reported-by: Justin Pryzby <pryzby@telsasoft.com>
Discussion: https://postgr.es/m/20221210201753.GA27893@telsasoft.com
This allows left join removals and unique joins to work with partitioned
tables. The planner just lacked sufficient proofs that a given join
would not cause any row duplication. Unique indexes currently serve as
that proof, so have get_relation_info() populate the indexlist for
partitioned tables too.
Author: Arne Roland
Reviewed-by: Alvaro Herrera, Zhihong Yu, Amit Langote, David Rowley
Discussion: https://postgr.es/m/c3b2408b7a39433b8230bbcd02e9f302@index.de
GIN index scans were not taking any descent CPU-based cost into account. That
made them look cheaper than other types of indexes when they shouldn't be.
We use the same heuristic as for btree indexes, but multiply it by the number
of searched entries.
Additionally, the CPU cost for the tree was based largely on a
genericcostestimate. For a GIN index, we should not charge index quals per
tuple, but per entry. On top of this, charge cpu_index_tuple_cost per actual
tuple.
This should fix the cases where a GIN index is preferred over a btree and
the ones where a memoize node is not added on top of the GIN index scan
because it seemed too cheap.
We don't packpatch this to evade unexpected plan changes in stable versions.
Discussion: https://postgr.es/m/CABs3KGQnOkyQ42-zKQqiE7M0Ks9oWDSee%3D%2BJx3-TGq%3D68xqWYw%40mail.gmail.com
Discussion: https://postgr.es/m/3188617.44csPzL39Z%40aivenronan
Author: Ronan Dunklau
Reported-By: Hung Nguyen
Reviewed-by: Tom Lane, Alexander Korotkov
Just because I'm a neatnik, and I'm currently working on
code in this area. It annoys me to not be able to pgindent
my patches without working around unrelated changes.
A future commit will move the checkAsUser field from RangeTblEntry
to a new node that, unlike RTEs, will only be created for tables
mentioned in the query but not for the inheritance child relations
added to the query by the planner. So, checkAsUser value for a
given child relation will have to be obtained by referring to that
for its ancestor mentioned in the query.
In preparation, it seems better to expand the use of RelOptInfo.userid
during planning in place of rte->checkAsUser so that there will be
fewer places to adjust for the above change.
Given that the child-to-ancestor mapping is not available during the
execution of a given "child" ForeignScan node, add a checkAsUser
field to ForeignScan to carry the child relation's RelOptInfo.userid.
Author: Amit Langote <amitlangote09@gmail.com>
Discussion: https://postgr.es/m/CA+HiwqGFCs2uq7VRKi7g+FFKbP6Ea_2_HkgZb2HPhUfaAKT3ng@mail.gmail.com
We've made multiple attempts at preventing get_actual_variable_range
from taking an unreasonable amount of time (3ca930fc3, fccebe421).
But there's still an issue for the very first planning attempt after
deletion of a large number of extremal-valued tuples. While that
planning attempt will set "killed" bits on the tuples it visits and
thereby reduce effort for next time, there's still a lot of work it
has to do to visit the heap and then set those bits. It's (usually?)
not worth it to do that much work at plan time to have a slightly
better estimate, especially in a context like this where the table
contents are known to be mutating rapidly.
Therefore, let's bound the amount of work to be done by giving up
after we've visited 100 heap pages. Giving up just means we'll
fall back on the extremal value recorded in pg_statistic, so it
shouldn't mean that planner estimates suddenly become worthless.
Note that this means we'll still gradually whittle down the problem
by setting a few more index "killed" bits in each planning attempt;
so eventually we'll reach a good state (barring further deletions),
even in the absence of VACUUM.
Simon Riggs, per a complaint from Jakub Wartak (with cosmetic
adjustments by me). Back-patch to all supported branches.
Discussion: https://postgr.es/m/CAKZiRmznOwi0oaV=4PHOCM4ygcH4MgSvt8=5cu_vNCfc8FSUug@mail.gmail.com
eqjoinsel() currently makes use of MCV stats only when we have such
stats for both sides of the clause. As coded, though, it would
fetch those stats even when they're present for just one side.
This can be a bit expensive with high statistics targets, leading
to wasted effort in common cases such as joining a unique column
to a non-unique column. So it seems worth the trouble to do a quick
pre-check to confirm that both sides have MCVs before fetching either.
Also, tweak the API spec for get_attstatsslot() to document the
method we're using here.
David Geier, Tomas Vondra, Tom Lane
Discussion: https://postgr.es/m/b9846ca0-5f1c-9b26-5881-aad3f42b07f0@gmail.com
Some cases would result in "cache lookup failed for statistics object",
due to trying to fetch inherited statistics when only non-inherited
ones are available or vice versa.
Richard Guo and Justin Pryzby
Discussion: https://postgr.es/m/20221030170520.GM16921@telsasoft.com
This reverts commit db0d67db24 and
several follow-on fixes. The idea of making a cost-based choice
of the order of the sorting columns is not fundamentally unsound,
but it requires cost information and data statistics that we don't
really have. For example, relying on procost to distinguish the
relative costs of different sort comparators is pretty pointless
so long as most such comparator functions are labeled with cost 1.0.
Moreover, estimating the number of comparisons done by Quicksort
requires more than just an estimate of the number of distinct values
in the input: you also need some idea of the sizes of the larger
groups, if you want an estimate that's good to better than a factor of
three or so. That's data that's often unknown or not very reliable.
Worse, to arrive at estimates of the number of calls made to the
lower-order-column comparison functions, the code needs to make
estimates of the numbers of distinct values of multiple columns,
which are necessarily even less trustworthy than per-column stats.
Even if all the inputs are perfectly reliable, the cost algorithm
as-implemented cannot offer useful information about how to order
sorting columns beyond the point at which the average group size
is estimated to drop to 1.
Close inspection of the code added by db0d67db2 shows that there
are also multiple small bugs. These could have been fixed, but
there's not much point if we don't trust the estimates to be
accurate in-principle.
Finally, the changes in cost_sort's behavior made for very large
changes (often a factor of 2 or so) in the cost estimates for all
sorting operations, not only those for multi-column GROUP BY.
That naturally changes plan choices in many situations, and there's
precious little evidence to show that the changes are for the better.
Given the above doubts about whether the new estimates are really
trustworthy, it's hard to summon much confidence that these changes
are better on the average.
Since we're hard up against the release deadline for v15, let's
revert these changes for now. We can always try again later.
Note: in v15, I left T_PathKeyInfo in place in nodes.h even though
it's unreferenced. Removing it would be an ABI break, and it seems
a bit late in the release cycle for that.
Discussion: https://postgr.es/m/TYAPR01MB586665EB5FB2C3807E893941F5579@TYAPR01MB5866.jpnprd01.prod.outlook.com
Fix selfuncs.h cpluspluscheck complaint, without reintroducing a
parameter name inconsistency (restore the original declaration names,
and then make corresponding function definitions consistent with that).
Oversight in commit a601366a.
Author: Peter Geoghegan <pg@bowt.ie>
Reported-By: Andres Freund <andres@anarazel.de>
guc.c has grown to be one of our largest .c files, making it
a bottleneck for compilation. It's also acquired a bunch of
knowledge that'd be better kept elsewhere, because of our not
very good habit of putting variable-specific check hooks here.
Hence, split it up along these lines:
* guc.c itself retains just the core GUC housekeeping mechanisms.
* New file guc_funcs.c contains the SET/SHOW interfaces and some
SQL-accessible functions for GUC manipulation.
* New file guc_tables.c contains the data arrays that define the
built-in GUC variables, along with some already-exported constant
tables.
* GUC check/assign/show hook functions are moved to the variable's
home module, whenever that's clearly identifiable. A few hard-
to-classify hooks ended up in commands/variable.c, which was
already a home for miscellaneous GUC hook functions.
To avoid cluttering a lot more header files with #include "guc.h",
I also invented a new header file utils/guc_hooks.h and put all
the GUC hook functions' declarations there, regardless of their
originating module. That allowed removal of #include "guc.h"
from some existing headers. The fallout from that (hopefully
all caught here) demonstrates clearly why such inclusions are
best minimized: there are a lot of files that, for example,
were getting array.h at two or more levels of remove, despite
not having any connection at all to GUCs in themselves.
There is some very minor code beautification here, such as
renaming a couple of inconsistently-named hook functions
and improving some comments. But mostly this just moves
code from point A to point B and deals with the ensuing
needs for #include adjustments and exporting a few functions
that previously weren't exported.
Patch by me, per a suggestion from Andres Freund; thanks also
to Michael Paquier for the idea to invent guc_funcs.c.
Discussion: https://postgr.es/m/587607.1662836699@sss.pgh.pa.us
The standard way to check for list emptiness is to compare the
List pointer to NIL; our list code goes out of its way to ensure
that that is the only representation of an empty list. (An
acceptable alternative is a plain boolean test for non-null
pointer, but explicit mention of NIL is usually preferable.)
Various places didn't get that memo and expressed the condition
with list_length(), which might not be so bad except that there
were such a variety of ways to check it exactly: equal to zero,
less than or equal to zero, less than one, yadda yadda. In the
name of code readability, let's standardize all those spellings
as "list == NIL" or "list != NIL". (There's probably some
microscopic efficiency gain too, though few of these look to be
at all performance-critical.)
A very small number of cases were left as-is because they seemed
more consistent with other adjacent list_length tests that way.
Peter Smith, with bikeshedding from a number of us
Discussion: https://postgr.es/m/CAHut+PtQYe+ENX5KrONMfugf0q6NHg4hR5dAhqEXEc2eefFeig@mail.gmail.com
The patch that added regcollation doesn't seem to have been too
thorough about supporting it everywhere that other reg* types
are supported. Fix that. (The find_expr_references omission
is moderately serious, since it could result in missing expression
dependencies. The others are less exciting.)
Noted while fixing bug #17483. Back-patch to v13 where
regcollation was added.
Discussion: https://postgr.es/m/1423433.1652722406@sss.pgh.pa.us
This replaces all MemSet() calls with struct initialization where that
is easily and obviously possible. (For example, some cases have to
worry about padding bits, so I left those.)
(The same could be done with appropriate memset() calls, but this
patch is part of an effort to phase out MemSet(), so it doesn't touch
memset() calls.)
Reviewed-by: Ranier Vilela <ranier.vf@gmail.com>
Reviewed-by: Alvaro Herrera <alvherre@alvh.no-ip.org>
Discussion: https://www.postgresql.org/message-id/9847b13c-b785-f4e2-75c3-12ec77a3b05c@enterprisedb.com
When evaluating a query with a multi-column GROUP BY clause using sort,
the cost may be heavily dependent on the order in which the keys are
compared when building the groups. Grouping does not imply any ordering,
so we're allowed to compare the keys in arbitrary order, and a Hash Agg
leverages this. But for Group Agg, we simply compared keys in the order
as specified in the query. This commit explores alternative ordering of
the keys, trying to find a cheaper one.
In principle, we might generate grouping paths for all permutations of
the keys, and leave the rest to the optimizer. But that might get very
expensive, so we try to pick only a couple interesting orderings based
on both local and global information.
When planning the grouping path, we explore statistics (number of
distinct values, cost of the comparison function) for the keys and
reorder them to minimize comparison costs. Intuitively, it may be better
to perform more expensive comparisons (for complex data types etc.)
last, because maybe the cheaper comparisons will be enough. Similarly,
the higher the cardinality of a key, the lower the probability we’ll
need to compare more keys. The patch generates and costs various
orderings, picking the cheapest ones.
The ordering of group keys may interact with other parts of the query,
some of which may not be known while planning the grouping. E.g. there
may be an explicit ORDER BY clause, or some other ordering-dependent
operation, higher up in the query, and using the same ordering may allow
using either incremental sort or even eliminate the sort entirely.
The patch generates orderings and picks those minimizing the comparison
cost (for various pathkeys), and then adds orderings that might be
useful for operations higher up in the plan (ORDER BY, etc.). Finally,
it always keeps the ordering specified in the query, on the assumption
the user might have additional insights.
This introduces a new GUC enable_group_by_reordering, so that the
optimization may be disabled if needed.
The original patch was proposed by Teodor Sigaev, and later improved and
reworked by Dmitry Dolgov. Reviews by a number of people, including me,
Andrey Lepikhov, Claudio Freire, Ibrar Ahmed and Zhihong Yu.
Author: Dmitry Dolgov, Teodor Sigaev, Tomas Vondra
Reviewed-by: Tomas Vondra, Andrey Lepikhov, Claudio Freire, Ibrar Ahmed, Zhihong Yu
Discussion: https://postgr.es/m/7c79e6a5-8597-74e8-0671-1c39d124c9d6%40sigaev.ru
Discussion: https://postgr.es/m/CA%2Bq6zcW_4o2NC0zutLkOJPsFt80megSpX_dVRo6GK9PC-Jx_Ag%40mail.gmail.com
Add pg_statistic_ext_data.stxdinherit flag, so that for each extended
statistics definition we can store two versions of data - one for the
relation alone, one for the whole inheritance tree. This is analogous to
pg_statistic.stainherit, but we failed to include such flag in catalogs
for extended statistics, and we had to work around it (see commits
859b3003de, 36c4bc6e72 and 20b9fa308e).
This changes the relationship between the two catalogs storing extended
statistics objects (pg_statistic_ext and pg_statistic_ext_data). Until
now, there was a simple 1:1 mapping - for each definition there was one
pg_statistic_ext_data row, and this row was inserted while creating the
statistics (and then updated during ANALYZE). With the stxdinherit flag,
we don't know how many rows there will be (child relations may be added
after the statistics object is defined), so there may be up to two rows.
We could make CREATE STATISTICS to always create both rows, but that
seems wasteful - without partitioning we only need stxdinherit=false
rows, and declaratively partitioned tables need only stxdinherit=true.
So we no longer initialize pg_statistic_ext_data in CREATE STATISTICS,
and instead make that a responsibility of ANALYZE. Which is what we do
for regular statistics too.
Patch by me, with extensive improvements and fixes by Justin Pryzby.
Author: Tomas Vondra, Justin Pryzby
Reviewed-by: Tomas Vondra, Justin Pryzby
Discussion: https://postgr.es/m/20210923212624.GI831%40telsasoft.com
Commit 859b3003de disabled building of extended stats for inheritance
trees, to prevent updating the same catalog row twice. While that
resolved the issue, it also means there are no extended stats for
declaratively partitioned tables, because there are no data in the
non-leaf relations.
That also means declaratively partitioned tables were not affected by
the issue 859b3003de addressed, which means this is a regression
affecting queries that calculate estimates for the whole inheritance
tree as a whole (which includes e.g. GROUP BY queries).
But because partitioned tables are empty, we can invert the condition
and build statistics only for the case with inheritance, without losing
anything. And we can consider them when calculating estimates.
It may be necessary to run ANALYZE on partitioned tables, to collect
proper statistics. For declarative partitioning there should no prior
statistics, and it might take time before autoanalyze is triggered. For
tables partitioned by inheritance the statistics may include data from
child relations (if built 859b3003de), contradicting the current code.
Report and patch by Justin Pryzby, minor fixes and cleanup by me.
Backpatch all the way back to PostgreSQL 10, where extended statistics
were introduced (same as 859b3003de).
Author: Justin Pryzby
Reported-by: Justin Pryzby
Backpatch-through: 10
Discussion: https://postgr.es/m/20210923212624.GI831%40telsasoft.com
Since commit 859b3003de we only build extended statistics for individual
relations, ignoring the child relations. This resolved the issue with
updating catalog tuple twice, but we still tried to use the statistics
when calculating estimates for the whole inheritance tree. When the
relations contain very distinct data, it may produce bogus estimates.
This is roughly the same issue 427c6b5b9 addressed ~15 years ago, and we
fix it the same way - by ignoring extended statistics when calculating
estimates for the inheritance tree as a whole. We still consider
extended statistics when calculating estimates for individual child
relations, of course.
This may result in plan changes due to different estimates, but if the
old statistics were not describing the inheritance tree particularly
well it's quite likely the new plans is actually better.
Report and patch by Justin Pryzby, minor fixes and cleanup by me.
Backpatch all the way back to PostgreSQL 10, where extended statistics
were introduced (same as 859b3003de).
Author: Justin Pryzby
Reported-by: Justin Pryzby
Backpatch-through: 10
Discussion: https://postgr.es/m/20210923212624.GI831%40telsasoft.com
get_variable_range() would incautiously believe that statistics
containing only an MCV list are sufficient to derive a range estimate.
That's okay for an enum-like column that contains only MCVs, but
otherwise the estimate could be pretty bad. Make it report that the
range is indeterminate unless the MCVs plus nullfrac account for
the whole table.
I don't think this needs a dedicated test case, since a quick code
coverage check verifies that the existing regression tests traverse
all the alternatives. There is room to doubt that a future-proof
test case could be built anyway, given that the submitted example
accidentally doesn't fail before v11.
Per bug #17207 from Simon Perepelitsa. Back-patch to v10.
In principle this has been broken all along, but I'm hesitant to
make such changes in 9.6, since if anyone is unhappy with 9.6.24's
behavior there will be no second chance to fix it.
Discussion: https://postgr.es/m/17207-5265aefa79e333b4@postgresql.org
The code inconsistently used "statistic object" or "statistics" where
the correct term, as discussed, is actually "statistics object". This
improves the state of the code to be more consistent.
While on it, fix an incorrect error message introduced in a4d75c8. This
error should never happen, as the code states, but it would be
misleading.
Author: Justin Pryzby
Reviewed-by: Álvaro Herrera, Michael Paquier
Discussion: https://postgr.es/m/20210924215827.GS831@telsasoft.com
Backpatch-through: 14
This case should be disallowed, just as FOR UPDATE with a plain
GROUP BY is disallowed; FOR UPDATE only makes sense when each row
of the query result can be identified with a single table row.
However, we missed teaching CheckSelectLocking() to check
groupingSets as well as groupClause, so that it would allow
degenerate grouping sets. That resulted in a bad plan and
a null-pointer dereference in the executor.
Looking around for other instances of the same bug, the only one
I found was in examine_simple_variable(). That'd just lead to
silly estimates, but it should be fixed too.
Per private report from Yaoguang Chen.
Back-patch to all supported branches.
Also "make reformat-dat-files".
The only change worthy of note is that pgindent messed up the formatting
of launcher.c's struct LogicalRepWorkerId, which led me to notice that
that struct wasn't used at all anymore, so I just took it out.
Here we add a new output parameter to estimate_num_groups() to allow it to
inform the caller of additional, possibly useful information about the
estimation.
The new output parameter is a struct that currently contains just a single
field with a set of flags. This was done rather than having the flags as
an output parameter to allow future fields to be added without having to
change the signature of the function at a later date when we want to pass
back further information that might not be suitable to store in the flags
field.
It seems reasonable that one day in the future that the planner would want
to know more about the estimation. For example, how many individual sets
of statistics was the estimation generated from? The planner may want to
take that into account if we ever want to consider risks as well as costs
when generating plans.
For now, there's only 1 flag we set in the flags field. This is to
indicate if the estimation fell back on using the hard-coded constants in
any part of the estimation. Callers may like to change their behavior if
this is set, and this gives them the ability to do so. Callers may pass
the flag pointer as NULL if they have no interest in obtaining any
additional information about the estimate.
We're not adding any actual usages of these flags here. Some follow-up
commits will make use of this feature. Additionally, we're also not
making any changes to add support for clauselist_selectivity() and
clauselist_selectivity_ext(). However, if this is required in the future
then the same struct being added here should be fine to use as a new
output argument for those functions too.
Author: David Rowley
Discussion: https://postgr.es/m/CAApHDvqQqpk=1W-G_ds7A9CsXX3BggWj_7okinzkLVhDubQzjA@mail.gmail.com
Allow defining extended statistics on expressions, not just just on
simple column references. With this commit, expressions are supported
by all existing extended statistics kinds, improving the same types of
estimates. A simple example may look like this:
CREATE TABLE t (a int);
CREATE STATISTICS s ON mod(a,10), mod(a,20) FROM t;
ANALYZE t;
The collected statistics are useful e.g. to estimate queries with those
expressions in WHERE or GROUP BY clauses:
SELECT * FROM t WHERE mod(a,10) = 0 AND mod(a,20) = 0;
SELECT 1 FROM t GROUP BY mod(a,10), mod(a,20);
This introduces new internal statistics kind 'e' (expressions) which is
built automatically when the statistics object definition includes any
expressions. This represents single-expression statistics, as if there
was an expression index (but without the index maintenance overhead).
The statistics is stored in pg_statistics_ext_data as an array of
composite types, which is possible thanks to 79f6a942bd.
CREATE STATISTICS allows building statistics on a single expression, in
which case in which case it's not possible to specify statistics kinds.
A new system view pg_stats_ext_exprs can be used to display expression
statistics, similarly to pg_stats and pg_stats_ext views.
ALTER TABLE ... ALTER COLUMN ... TYPE now treats indexes the same way it
treats indexes, i.e. it drops and recreates the statistics. This means
all statistics are reset, and we no longer try to preserve at least the
functional dependencies. This should not be a major issue in practice,
as the functional dependencies actually rely on per-column statistics,
which were always reset anyway.
Author: Tomas Vondra
Reviewed-by: Justin Pryzby, Dean Rasheed, Zhihong Yu
Discussion: https://postgr.es/m/ad7891d2-e90c-b446-9fe2-7419143847d7%40enterprisedb.com
When estimating the number of groups using extended statistics, the code
was discarding information about system attributes. This led to strange
situation that
SELECT 1 FROM t GROUP BY ctid;
could have produced higher estimate (equal to pg_class.reltuples) than
SELECT 1 FROM t GROUP BY a, b, ctid;
with extended statistics on (a,b). Fixed by retaining information about
the system attribute.
Backpatch all the way to 10, where extended statistics were introduced.
Author: Tomas Vondra
Backpatch-through: 10
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
We have a dozen or so places that need to iterate over all but the
first cell of a List. Prior to v13 this was typically written as
for_each_cell(lc, lnext(list_head(list)))
Commit 1cff1b95a changed these to
for_each_cell(lc, list, list_second_cell(list))
This patch introduces a new macro for_each_from() which expresses
the start point as a list index, allowing these to be written as
for_each_from(lc, list, 1)
This is marginally more efficient, since ForEachState.i can be
initialized directly instead of backing into it from a ListCell
address. It also seems clearer and less typo-prone.
Some of the remaining uses of for_each_cell() look like they could
profitably be changed to for_each_from(), but here I confined myself
to changing uses of list_second_cell().
Also, fix for_each_cell_setup() and for_both_cell_setup() to
const-ify their arguments; that's a simple oversight in 1cff1b95a.
Back-patch into v13, on the grounds that (1) the const-ification
is a minor bug fix, and (2) it's better for back-patching purposes
if we only have two ways to write these loops rather than three.
In HEAD, also remove list_third_cell() and list_fourth_cell(),
which were also introduced in 1cff1b95a, and are unused as of
cc99baa43. It seems unlikely that any third-party code would
have started to use them already; anyone who has can be directed
to list_nth_cell instead.
Discussion: https://postgr.es/m/CAApHDvpo1zj9KhEpU2cCRZfSM3Q6XGdhzuAS2v79PH7WJBkYVA@mail.gmail.com
To make GetSnapshotData() more scalable, it cannot not look at at each proc's
xmin: While snapshot contents do not need to change whenever a read-only
transaction commits or a snapshot is released, a proc's xmin is modified in
those cases. The frequency of xmin modifications leads to, particularly on
higher core count systems, many cache misses inside GetSnapshotData(), despite
the data underlying a snapshot not changing. That is the most
significant source of GetSnapshotData() scaling poorly on larger systems.
Without accessing xmins, GetSnapshotData() cannot calculate accurate horizons /
thresholds as it has so far. But we don't really have to: The horizons don't
actually change that much between GetSnapshotData() calls. Nor are the horizons
actually used every time a snapshot is built.
The trick this commit introduces is to delay computation of accurate horizons
until there use and using horizon boundaries to determine whether accurate
horizons need to be computed.
The use of RecentGlobal[Data]Xmin to decide whether a row version could be
removed has been replaces with new GlobalVisTest* functions. These use two
thresholds to determine whether a row can be pruned:
1) definitely_needed, indicating that rows deleted by XIDs >= definitely_needed
are definitely still visible.
2) maybe_needed, indicating that rows deleted by XIDs < maybe_needed can
definitely be removed
GetSnapshotData() updates definitely_needed to be the xmin of the computed
snapshot.
When testing whether a row can be removed (with GlobalVisTestIsRemovableXid())
and the tested XID falls in between the two (i.e. XID >= maybe_needed && XID <
definitely_needed) the boundaries can be recomputed to be more accurate. As it
is not cheap to compute accurate boundaries, we limit the number of times that
happens in short succession. As the boundaries used by
GlobalVisTestIsRemovableXid() are never reset (with maybe_needed updated by
GetSnapshotData()), it is likely that further test can benefit from an earlier
computation of accurate horizons.
To avoid regressing performance when old_snapshot_threshold is set (as that
requires an accurate horizon to be computed), heap_page_prune_opt() doesn't
unconditionally call TransactionIdLimitedForOldSnapshots() anymore. Both the
computation of the limited horizon, and the triggering of errors (with
SetOldSnapshotThresholdTimestamp()) is now only done when necessary to remove
tuples.
This commit just removes the accesses to PGXACT->xmin from
GetSnapshotData(), but other members of PGXACT residing in the same
cache line are accessed. Therefore this in itself does not result in a
significant improvement. Subsequent commits will take advantage of the
fact that GetSnapshotData() now does not need to access xmins anymore.
Note: This contains a workaround in heap_page_prune_opt() to keep the
snapshot_too_old tests working. While that workaround is ugly, the tests
currently are not meaningful, and it seems best to address them separately.
Author: Andres Freund <andres@anarazel.de>
Reviewed-By: Robert Haas <robertmhaas@gmail.com>
Reviewed-By: Thomas Munro <thomas.munro@gmail.com>
Reviewed-By: David Rowley <dgrowleyml@gmail.com>
Discussion: https://postgr.es/m/20200301083601.ews6hz5dduc3w2se@alap3.anarazel.de
Since commit 044c99bc5, eqjoinsel passes the passed-in collation
to any operators it invokes. However, neqjoinsel failed to pass
on whatever collation it got, so that if we invoked a
collation-dependent operator via that code path, we'd get "could not
determine which collation to use for string comparison" or the like.
Per report from Justin Pryzby. Back-patch to v12, like the previous
commit.
Discussion: https://postgr.es/m/20200721191606.GL5748@telsasoft.com
ineq_histogram_selectivity() can be invoked in situations where the
ordering we care about is not that of the column's histogram. We could
be considering some other collation, or even more drastically, the
query operator might not agree at all with what was used to construct
the histogram. (We'll get here for anything using scalarineqsel-based
estimators, so that's quite likely to happen for extension operators.)
Up to now we just ignored this issue and assumed we were dealing with
an operator/collation whose sort order exactly matches the histogram,
possibly resulting in junk estimates if the binary search gets confused.
It's past time to improve that, since the use of nondefault collations
is increasing. What we can do is verify that the given operator and
collation match what's recorded in pg_statistic, and use the existing
code only if so. When they don't match, instead execute the operator
against each histogram entry, and take the fraction of successes as our
selectivity estimate. This gives an estimate that is probably good to
about 1/histogram_size, with no assumptions about ordering. (The quality
of the estimate is likely to degrade near the ends of the value range,
since the two orderings probably don't agree on what is an extremal value;
but this is surely going to be more reliable than what we did before.)
At some point we might further improve matters by storing more than one
histogram calculated according to different orderings. But this code
would still be good fallback logic when no matches exist, so that is
not an argument for not doing this.
While here, also improve get_variable_range() to deal more honestly
with non-default collations.
This isn't back-patchable, because it requires adding another argument
to ineq_histogram_selectivity, and because it might have significant
impact on the estimation results for extension operators relying on
scalarineqsel --- mostly for the better, one hopes, but in any case
destabilizing plan choices in back branches is best avoided.
Per investigation of a report from James Lucas.
Discussion: https://postgr.es/m/CAAFmbbOvfi=wMM=3qRsPunBSLb8BFREno2oOzSBS=mzfLPKABw@mail.gmail.com
Peter Geoghegan
Tom Lane
Bruce Momjian
David Rowley
Dean Rasheed
Michael Paquier
Alvaro Herrera
Alexander Korotkov
Peter Eisentraut
Tomas Vondra
Heikki Linnakangas
Andres Freund