If we don't have any MCV statistics for the inner relation, and we don't
trust its numdistinct estimate either, eqjoinsel_semi falls back to a very
conservative estimate (that 50% of the outer rows have matches). This is
particularly problematic if the inner relation is completely empty, since
then even an explicit ANALYZE won't produce any pg_statistic entries,
so there's no way to budge the planner off the bad estimate.
We'd produce a better estimate in such cases if we used the nd2/nd1
selectivity heuristic, so an easy fix is to treat the nd2 estimate as
non-default if we derive it from clamping to the inner rel's rowcount
estimate. This won't fix every related case (mainly because the rowcount
estimate might be larger than DEFAULT_NUM_DISTINCT), but it seems like a
sane extension of the existing logic, so let's apply the change in HEAD
and see if anyone complains. Per bug #14438 from Nikolay Nikitin.
Report: https://postgr.es/m/20161128182113.6527.58926@wrigleys.postgresql.org
Discussion: https://postgr.es/m/31089.1480384713@sss.pgh.pa.us
If ANALYZE found no repeated non-null entries in its sample, it set the
column's stadistinct value to -1.0, intending to indicate that the entries
are all distinct. But what this value actually means is that the number
of distinct values is 100% of the table's rowcount, and thus it was
overestimating the number of distinct values by however many nulls there
are. This could lead to very poor selectivity estimates, as for example
in a recent report from Andreas Joseph Krogh. We should discount the
stadistinct value by whatever we've estimated the nulls fraction to be.
(That is what will happen if we choose to use a negative stadistinct for
a column that does have repeated entries, so this code path was just
inconsistent.)
In addition to fixing the stadistinct entries stored by several different
ANALYZE code paths, adjust the logic where get_variable_numdistinct()
forces an "all distinct" estimate on the basis of finding a relevant unique
index. Unique indexes don't reject nulls, so there's no reason to assume
that the null fraction doesn't apply.
Back-patch to all supported branches. Back-patching is a bit of a judgment
call, but this problem seems to affect only a few users (else we'd have
identified it long ago), and it's bad enough when it does happen that
destabilizing plan choices in a worse direction seems unlikely.
Patch by me, with documentation wording suggested by Dean Rasheed
Report: <VisenaEmail.26.df42f82acae38a58.156463942b8@tc7-visena>
Discussion: <16143.1470350371@sss.pgh.pa.us>
Now indexes (but only B-tree for now) can contain "extra" column(s) which
doesn't participate in index structure, they are just stored in leaf
tuples. It allows to use index only scan by using single index instead
of two or more indexes.
Author: Anastasia Lubennikova with minor editorializing by me
Reviewers: David Rowley, Peter Geoghegan, Jeff Janes
When adjusting the estimate for the number of distinct values from a
rel in a grouped query to take into account the selectivity of the
rel's restrictions, use a formula that is less likely to produce
under-estimates.
The old formula simply multiplied the number of distinct values in the
rel by the restriction selectivity, which would be correct if the
restrictions were fully correlated with the grouping expressions, but
can produce significant under-estimates in cases where they are not
well correlated.
The new formula is based on the random selection probability, and so
assumes that the restrictions are not correlated with the grouping
expressions. This is guaranteed to produce larger estimates, and of
course risks over-estimating in cases where the restrictions are
correlated, but that has less severe consequences than
under-estimating, which might lead to a HashAgg that consumes an
excessive amount of memory.
This could possibly be improved upon in the future by identifying
correlated restrictions and using a hybrid of the old and new
formulae.
Author: Tomas Vondra, with some hacking be me
Reviewed-by: Mark Dilger, Alexander Korotkov, Dean Rasheed and Tom Lane
Discussion: http://www.postgresql.org/message-id/flat/56CD0381.5060502@2ndquadrant.com
If the referenced rel was proven empty, we'd compute 0/0 here, which
results in the function returning NaN. That's a bit more serious
than the other zero-divide case. Still, it only seems to be possible
in HEAD, so no back-patch.
Per report from Piotr Stefaniak. I looked through the rest of selfuncs.c
and found no other likely trouble spots.
This avoids a possible divide-by-zero in the following calculation,
and rounding the number to an integer seems like saner behavior anyway.
Assuming IEEE math, the division would yield +Infinity which would get
replaced by 1.0 at the bottom of the function, so nothing really
interesting would ensue; but avoiding divide-by-zero seems like a
good idea on general principles.
Per report from Piotr Stefaniak. No back-patch since this seems
mostly cosmetic.
This enables external code to create access methods. This is useful so
that extensions can add their own access methods which can be formally
tracked for dependencies, so that DROP operates correctly. Also, having
explicit support makes pg_dump work correctly.
Currently only index AMs are supported, but we expect different types to
be added in the future.
Authors: Alexander Korotkov, Petr Jelínek
Reviewed-By: Teodor Sigaev, Petr Jelínek, Jim Nasby
Commitfest-URL: https://commitfest.postgresql.org/9/353/
Discussion: https://www.postgresql.org/message-id/CAPpHfdsXwZmojm6Dx+TJnpYk27kT4o7Ri6X_4OSWcByu1Rm+VA@mail.gmail.com
All along, this function should have treated WindowFuncs in a manner
similar to Aggrefs, ie with an option whether or not to recurse into them.
By not considering the case, it was always recursing, which is OK for most
callers (although I suspect that the case in prepare_sort_from_pathkeys
might represent a bug). But now we need return-without-recursing behavior
as well. There are also more than a few callers that should never see a
WindowFunc, and now we'll get some error checking on that.
In commit 1d97c19a0f and later c1d9579dd8, we extended
pull_var_clause's API by adding enum-type arguments. That's sort of a pain
to maintain, though, because it means every time we add a new behavior we
must touch every last one of the call sites, even if there's a reasonable
default behavior that most of them could use. Let's switch over to using a
bitmask of flags, instead; that seems more maintainable and might save a
nanosecond or two as well. This commit changes no behavior in itself,
though I'm going to follow it up with one that does add a new behavior.
In passing, remove flatten_tlist(), which has not been used since 9.1
and would otherwise need the same API changes.
Removing these enums means that optimizer/tlist.h no longer needs to
depend on optimizer/var.h. Changing that caused a number of C files to
need addition of #include "optimizer/var.h" (probably we can thank old
runs of pgrminclude for that); but on balance it seems like a good change
anyway.
This patch reduces pg_am to just two columns, a name and a handler
function. All the data formerly obtained from pg_am is now provided
in a C struct returned by the handler function. This is similar to
the designs we've adopted for FDWs and tablesample methods. There
are multiple advantages. For one, the index AM's support functions
are now simple C functions, making them faster to call and much less
error-prone, since the C compiler can now check function signatures.
For another, this will make it far more practical to define index access
methods in installable extensions.
A disadvantage is that SQL-level code can no longer see attributes
of index AMs; in particular, some of the crosschecks in the opr_sanity
regression test are no longer possible from SQL. We've addressed that
by adding a facility for the index AM to perform such checks instead.
(Much more could be done in that line, but for now we're content if the
amvalidate functions more or less replace what opr_sanity used to do.)
We might also want to expose some sort of reporting functionality, but
this patch doesn't do that.
Alexander Korotkov, reviewed by Petr Jelínek, and rather heavily
editorialized on by me.
A report from Andy Colson showed that gincostestimate() was not being
nearly paranoid enough about whether to believe the statistics it finds in
the index metapage. The problem is that the metapage stats (other than the
pending-pages count) are only updated by VACUUM, and in the worst case
could still reflect the index's original empty state even when it has grown
to many entries. We attempted to deal with that by scaling up the stats to
match the current index size, but if nEntries is zero then scaling it up
still gives zero. Moreover, the proportion of pages that are entry pages
vs. data pages vs. pending pages is unlikely to be estimated very well by
scaling if the index is now orders of magnitude larger than before.
We can improve matters by expanding the use of the rule-of-thumb estimates
I introduced in commit 7fb008c5ee59b040: if the index has grown by more
than a cutoff amount (here set at 4X growth) since VACUUM, then use the
rule-of-thumb numbers instead of scaling. This might not be exactly right
but it seems much less likely to produce insane estimates.
I also improved both the scaling estimate and the rule-of-thumb estimate
to account for numPendingPages, since it's reasonable to expect that that
is accurate in any case, and certainly pages that are in the pending list
are not either entry or data pages.
As a somewhat separate issue, adjust the estimation equations that are
concerned with extra fetches for partial-match searches. These equations
suppose that a fraction partialEntries / numEntries of the entry and data
pages will be visited as a consequence of a partial-match search. Now,
it's physically impossible for that fraction to exceed one, but our
estimate of partialEntries is mostly bunk, and our estimate of numEntries
isn't exactly gospel either, so we could arrive at a silly value. In the
example presented by Andy we were coming out with a value of 100, leading
to insane cost estimates. Clamp the fraction to one to avoid that.
Like the previous patch, back-patch to all supported branches; this
problem can be demonstrated in one form or another in all of them.
We tried to fetch statistics data from the index metapage, which does not
work if the index isn't actually present. If the index is hypothetical,
instead extrapolate some plausible internal statistics based on the index
page count provided by the index-advisor plugin.
There was already some code in gincostestimate() to invent internal stats
in this way, but since it was only meant as a stopgap for pre-9.1 GIN
indexes that hadn't been vacuumed since upgrading, it was pretty crude.
If we want it to support index advisors, we should try a little harder.
A small amount of testing says that it's better to estimate the entry pages
as 90% of the index, not 100%. Also, estimating the number of entries
(keys) as equal to the heap tuple count could be wildly wrong in either
direction. Instead, let's estimate 100 entries per entry page.
Perhaps someday somebody will want the index advisor to be able to provide
these numbers more directly, but for the moment this should serve.
Problem report and initial patch by Julien Rouhaud; modified by me to
invent less-bogus internal statistics. Back-patch to all supported
branches, since we've supported index advisors since 9.0.
Previously, a function call appearing at the top level of WHERE had a
hard-wired selectivity estimate of 0.3333333, a kludge conveniently dated
in the source code itself to July 1992. The expectation at the time was
that somebody would soon implement estimator support functions analogous
to those for operators; but no such code has appeared, nor does it seem
likely to in the near future. We do have an alternative solution though,
at least for immutable functions on single relations: creating an
expression index on the function call will allow ANALYZE to gather stats
about the function's selectivity. But the code in clause_selectivity()
failed to make use of such data even if it exists.
Refactor so that that will happen. I chose to make it try this technique
for any clause type for which clause_selectivity() doesn't have a special
case, not just functions. To avoid adding unnecessary overhead in the
common case where we don't learn anything new, make selfuncs.c provide an
API that hooks directly to examine_variable() and then var_eq_const(),
rather than the previous coding which laboriously constructed an OpExpr
only so that it could be expensively deconstructed again.
I preserved the behavior that the default estimate for a function call
is 0.3333333. (For any other expression node type, it's 0.5, as before.)
I had originally thought to make the default be 0.5 across the board, but
changing a default estimate that's survived for twenty-three years seems
like something not to do without a lot more testing than I care to put
into it right now.
Per a complaint from Jehan-Guillaume de Rorthais. Back-patch into 9.5,
but not further, at least for the moment.
Previously, convert_one_string_to_scalar() would examine up to 20 bytes of
the input string, producing a scalar conversion with theoretical precision
far greater than is of any possible use considering the other limitations
on the accuracy of the resulting selectivity estimate. (I think this
choice might pre-date the caller-level logic that strips any common prefix
of the strings; before that, there could have been value in scanning the
strings far enough to use all the precision available in a double.)
Aside from wasting cycles to little purpose, this choice meant that the
"denom" variable could grow to as much as 256^21 = 3.74e50, which could
overflow in some non-IEEE float arithmetics. While we don't really support
any machines with non-IEEE arithmetic anymore, this still seems like quite
an unnecessary platform dependency. Limit the scan to 12 bytes instead,
thus limiting "denom" to 256^13 = 2.03e31, a value more likely to be
computable everywhere.
Per testing by Greg Stark, which showed overflow failures in our standard
regression tests on VAX.
Although I think on all modern machines floating division by zero
results in Infinity not SIGFPE, we still don't want infinities
running around in the planner's costing estimates; too much risk
of that leading to insane behavior.
grouping_planner() failed to consider the possibility that final_rel
might be known dummy and hence have zero rowcount. (I wonder if it
would be better to set a rows estimate of 1 for dummy relations?
But at least in the back branches, changing this convention seems
like a bad idea, so I'll leave that for another day.)
Make certain that get_variable_numdistinct() produces a nonzero result.
The case that can be shown to be broken is with stadistinct < 0.0 and
small ntuples; we did not prevent the result from rounding to zero.
For good luck I applied clamp_row_est() to all the nonconstant return
values.
In ExecChooseHashTableSize(), Assert that we compute positive nbuckets
and nbatch. I know of no reason to think this isn't the case, but it
seems like a good safety check.
Per reports from Piotr Stefaniak. Back-patch to all active branches.
This formalizes a decision implicit in commit
4ea51cdfe8 and adds clean detection of
affected systems. Vendor updates are available for each such known bug.
Back-patch to 9.5, where the aforementioned commit first appeared.
This SQL standard functionality allows to aggregate data by different
GROUP BY clauses at once. Each grouping set returns rows with columns
grouped by in other sets set to NULL.
This could previously be achieved by doing each grouping as a separate
query, conjoined by UNION ALLs. Besides being considerably more concise,
grouping sets will in many cases be faster, requiring only one scan over
the underlying data.
The current implementation of grouping sets only supports using sorting
for input. Individual sets that share a sort order are computed in one
pass. If there are sets that don't share a sort order, additional sort &
aggregation steps are performed. These additional passes are sourced by
the previous sort step; thus avoiding repeated scans of the source data.
The code is structured in a way that adding support for purely using
hash aggregation or a mix of hashing and sorting is possible. Sorting
was chosen to be supported first, as it is the most generic method of
implementation.
Instead of, as in an earlier versions of the patch, representing the
chain of sort and aggregation steps as full blown planner and executor
nodes, all but the first sort are performed inside the aggregation node
itself. This avoids the need to do some unusual gymnastics to handle
having to return aggregated and non-aggregated tuples from underlying
nodes, as well as having to shut down underlying nodes early to limit
memory usage. The optimizer still builds Sort/Agg node to describe each
phase, but they're not part of the plan tree, but instead additional
data for the aggregation node. They're a convenient and preexisting way
to describe aggregation and sorting. The first (and possibly only) sort
step is still performed as a separate execution step. That retains
similarity with existing group by plans, makes rescans fairly simple,
avoids very deep plans (leading to slow explains) and easily allows to
avoid the sorting step if the underlying data is sorted by other means.
A somewhat ugly side of this patch is having to deal with a grammar
ambiguity between the new CUBE keyword and the cube extension/functions
named cube (and rollup). To avoid breaking existing deployments of the
cube extension it has not been renamed, neither has cube been made a
reserved keyword. Instead precedence hacking is used to make GROUP BY
cube(..) refer to the CUBE grouping sets feature, and not the function
cube(). To actually group by a function cube(), unlikely as that might
be, the function name has to be quoted.
Needs a catversion bump because stored rules may change.
Author: Andrew Gierth and Atri Sharma, with contributions from Andres Freund
Reviewed-By: Andres Freund, Noah Misch, Tom Lane, Svenne Krap, Tomas
Vondra, Erik Rijkers, Marti Raudsepp, Pavel Stehule
Discussion: CAOeZVidmVRe2jU6aMk_5qkxnB7dfmPROzM7Ur8JPW5j8Y5X-Lw@mail.gmail.com
The new type has the scope of whole the database cluster so it doesn't
behave the same as the existing OID alias types which have database
scope,
concerning object dependency. To avoid confusion constants of the new
type are prohibited from appearing where dependencies are made involving
it.
Also, add a note to the docs about possible MVCC violation and
optimization issues, which are general over the all reg* types.
Kyotaro Horiguchi
genericcostestimate() and friends used the cost of the entire indexqual
expressions as the charge for initial evaluation of indexscan arguments.
But of course the index column is not evaluated, only the other side
of the qual expression, so this was a bad overestimate if the index
column was an expensive expression.
To fix, refactor the logic in this area so that there's a single routine
charged with deconstructing index quals and figuring out what is the index
column and what is the comparison expression. This is more or less free in
the case of btree indexes, since btcostestimate() was doing equivalent
deconstruction already. It probably adds a bit of new overhead in the cases
of other index types, but not a lot. (In the case of GIN I think I saved
something by getting rid of code that wasn't aware that the index column
associations were already available "for free".)
Per recent gripe from Jeff Janes.
Arguably this is a bug fix, but I'm hesitant to back-patch because of the
possibility of destabilizing plan choices that people may be happy with.
BRIN is a new index access method intended to accelerate scans of very
large tables, without the maintenance overhead of btrees or other
traditional indexes. They work by maintaining "summary" data about
block ranges. Bitmap index scans work by reading each summary tuple and
comparing them with the query quals; all pages in the range are returned
in a lossy TID bitmap if the quals are consistent with the values in the
summary tuple, otherwise not. Normal index scans are not supported
because these indexes do not store TIDs.
As new tuples are added into the index, the summary information is
updated (if the block range in which the tuple is added is already
summarized) or not; in the latter case, a subsequent pass of VACUUM or
the brin_summarize_new_values() function will create the summary
information.
For data types with natural 1-D sort orders, the summary info consists
of the maximum and the minimum values of each indexed column within each
page range. This type of operator class we call "Minmax", and we
supply a bunch of them for most data types with B-tree opclasses.
Since the BRIN code is generalized, other approaches are possible for
things such as arrays, geometric types, ranges, etc; even for things
such as enum types we could do something different than minmax with
better results. In this commit I only include minmax.
Catalog version bumped due to new builtin catalog entries.
There's more that could be done here, but this is a good step forwards.
Loosely based on ideas from Simon Riggs; code mostly by Álvaro Herrera,
with contribution by Heikki Linnakangas.
Patch reviewed by: Amit Kapila, Heikki Linnakangas, Robert Haas.
Testing help from Jeff Janes, Erik Rijkers, Emanuel Calvo.
PS:
The research leading to these results has received funding from the
European Union's Seventh Framework Programme (FP7/2007-2013) under
grant agreement n° 318633.
If there are lots of uncommitted tuples at the end of the index range,
get_actual_variable_range() ends up fetching each one and doing an MVCC
visibility check on it, until it finally hits a visible tuple. This is
bad enough in isolation, considering that we don't need an exact answer
only an approximate one. But because the tuples are not yet committed,
each visibility check does a TransactionIdIsInProgress() test, which
involves scanning the ProcArray. When multiple sessions do this
concurrently, the ensuing contention results in horrid performance loss.
20X overall throughput loss on not-too-complicated queries is easy to
demonstrate in the back branches (though someone's made it noticeably
less bad in HEAD).
We can dodge the problem fairly effectively by using SnapshotDirty rather
than a normal MVCC snapshot. This will cause the index probe to take
uncommitted tuples as good, so that we incur only one tuple fetch and test
even if there are many such tuples. The extent to which this degrades the
estimate is debatable: it's possible the result is actually a more accurate
prediction than before, if the endmost tuple has become committed by the
time we actually execute the query being planned. In any case, it's not
very likely that it makes the estimate a lot worse.
SnapshotDirty will still reject tuples that are known committed dead, so
we won't give bogus answers if an invalid outlier has been deleted but not
yet vacuumed from the index. (Because btrees know how to mark such tuples
dead in the index, we shouldn't have a big performance problem in the case
that there are many of them at the end of the range.) This consideration
motivates not using SnapshotAny, which was also considered as a fix.
Note: the back branches were using SnapshotNow instead of an MVCC snapshot,
but the problem and solution are the same.
Per performance complaints from Bartlomiej Romanski, Josh Berkus, and
others. Back-patch to 9.0, where the issue was introduced (by commit
40608e7f94).
Most estimation functions apply estimate_expression_value to see if they
can reduce an expression to a constant; the key difference is that it
allows evaluation of stable as well as immutable functions in hopes of
ending up with a simple Const node. scalararraysel didn't get the memo
though, and neither did gincost_opexpr/gincost_scalararrayopexpr. Fix
that, and remove a now-unnecessary estimate_expression_value step in the
subsidiary function scalararraysel_containment.
Per complaint from Alexey Klyukin. Back-patch to 9.3. The problem
goes back further, but I'm hesitant to change estimation behavior in
long-stable release branches.
Simple oversight in commit 1cb108efb0 ---
recursively examining a subquery output column is only sane if the
original Var refers to a single output column. Found by Kevin Grittner.
Instead, use the active snapshot. Per Tom Lane, this function is
most interested in knowing the range of tuples our scan will actually
see.
This is another step towards full removal of SnapshotNow.
In a boolean column that contains mostly nulls, ANALYZE might not find
enough non-null values to populate the most-common-values stats,
but it would still create a pg_statistic entry with stanullfrac set.
The logic in booltestsel() for this situation did the wrong thing for
"col IS NOT TRUE" and "col IS NOT FALSE" tests, forgetting that null
values would satisfy these tests (so that the true selectivity would
be close to one, not close to zero). Per bug #8274.
Fix by Andrew Gierth, some comment-smithing by me.
This case doesn't normally happen, because the planner usually clamps
all row estimates to at least one row; but I found that it can arise
when dealing with relations excluded by constraints. Without a defense,
estimate_num_groups() can return zero, which leads to divisions by zero
inside the planner as well as assertion failures in the executor.
An alternative fix would be to change set_dummy_rel_pathlist() to make
the size estimate for a dummy relation 1 row instead of 0, but that seemed
pretty ugly; and probably someday we'll want to drop the convention that
the minimum rowcount estimate is 1 row.
Back-patch to 8.4, as the problem can be demonstrated that far back.
This works by extracting trigrams from the given regular expression,
in generally the same spirit as the previously-existing support for
LIKE searches, though of course the details are far more complicated.
Currently, only GIN indexes are supported. We might be able to make
it work with GiST indexes later.
The implementation includes adding API functions to backend/regex/
to provide a view of the search NFA created from a regular expression.
These functions are meant to be generic enough to be supportable in
a standalone version of the regex library, should that ever happen.
Alexander Korotkov, reviewed by Heikki Linnakangas and Tom Lane
Historically we've used a couple of very ad-hoc fudge factors to try to
get the right results when indexes of different sizes would satisfy a
query with the same number of index leaf tuples being visited. In
commit 21a39de580 I tweaked one of these
fudge factors, with results that proved disastrous for larger indexes.
Commit bf01e34b55 fudged it some more,
but still with not a lot of principle behind it.
What seems like a better way to address these issues is to explicitly model
index-descent costs, since that's what's really at stake when considering
diferent indexes with similar leaf-page-level costs. We tried that once
long ago, and found that charging random_page_cost per page descended
through was way too much, because upper btree levels tend to stay in cache
in real-world workloads. However, there's still CPU costs to think about,
and the previous fudge factors can be seen as a crude attempt to account
for those costs. So this patch replaces those fudge factors with explicit
charges for the number of tuple comparisons needed to descend the index
tree, plus a small charge per page touched in the descent. The cost
multipliers are chosen so that the resulting charges are in the vicinity of
the historical (pre-9.2) fudge factors for indexes of up to about a million
tuples, while not ballooning unreasonably beyond that, as the old fudge
factor did (even more so in 9.2).
To make this work accurately for btree indexes, add some code that allows
extraction of the known root-page height from a btree. There's no
equivalent number readily available for other index types, but we can use
the log of the number of index pages as an approximate substitute.
This seems like too much of a behavioral change to risk back-patching,
but it should improve matters going forward. In 9.2 I'll just revert
the fudge-factor change.
To provide some bias against using a large index when a small one would do
as well, genericcostestimate adds a "fudge factor", which for a long time
was random_page_cost * index_pages/10000. However, this can grow to be the
dominant term in indexscan cost estimates when the index involved is large
enough, a behavior that was never intended. Change to a ln(1 + n/10000)
formulation, which has nearly the same behavior up to a few hundred pages
but tails off significantly thereafter. (A log curve seems correct on
first principles, since what we're trying to account for here is index
descent costs, which are typically logarithmic.) Per bug #7619 from Niko
Kiirala.
Possibly this change should get back-patched, but I'm hesitant to mess with
cost estimates in stable branches.
examine_simple_variable supposed that any RTE_SUBQUERY rel it gets pointed
at must have been planned already. However, this isn't a safe assumption
because we must do selectivity estimation while generating indexscan paths,
and that code might look at join clauses involving a rel that the loop in
set_base_rel_sizes() hasn't reached yet. The simplest fix is to play dumb
in such a situation, that is give up trying to extract any stats for the
Var. This could possibly be improved by making a separate pass over the
RTE list to plan each unflattened subquery before we start the main
planning work --- but that would be pretty invasive and it doesn't seem
worth it, for now at least. (We couldn't just break set_base_rel_sizes()
into two loops: the prescan would need to handle all subquery rels in the
query, not only those in the current join subproblem.)
This bug was introduced in commit 1cb108efb0,
although I think that subsequent changes may have exposed it more than it
was originally. Per bug #7580 from Maxim Boguk.
This reduces unnecessary exposure of other headers through htup.h, which
is very widely included by many files.
I have chosen to move the function prototypes to the new file as well,
because that means htup.h no longer needs to include tupdesc.h. In
itself this doesn't have much effect in indirect inclusion of tupdesc.h
throughout the tree, because it's also required by execnodes.h; but it's
something to explore in the future, and it seemed best to do the htup.h
change now while I'm busy with it.
To generate btree-indexable conditions from regex WHERE conditions (such as
WHERE indexed_col ~ '^foo'), we need to be able to identify any fixed
prefix that a regex might have; that is, find any string that must be a
prefix of all strings satisfying the regex. We used to do that with
entirely ad-hoc code that looked at the source text of the regex. It
didn't know very much about regex syntax, which mostly meant that it would
fail to identify some optimizable cases; but Viktor Rosenfeld reported that
it would produce actively wrong answers for quantified parenthesized
subexpressions, such as '^(foo)?bar'. Rather than trying to extend the
ad-hoc code to cover this, let's get rid of it altogether in favor of
identifying prefixes by examining the compiled form of a regex.
To do this, I've added a new entry point "pg_regprefix" to the regex library;
hopefully it is defined in a sufficiently general fashion that it can remain
in the library when/if that code gets split out as a standalone project.
Since this bug has been there for a very long time, this fix needs to get
back-patched. However it depends on some other recent commits (particularly
the addition of wchar-to-database-encoding conversion), so I'll commit this
separately and then go to work on back-porting the necessary fixes.
Previously, pattern_fixed_prefix() was defined to return whatever fixed
prefix it could extract from the pattern, plus the "rest" of the pattern.
That definition was sensible for LIKE patterns, but not so much for
regexes, where reconstituting a valid pattern minus the prefix could be
quite tricky (certainly the existing code wasn't doing that correctly).
Since the only thing that callers ever did with the "rest" of the pattern
was to pass it to like_selectivity() or regex_selectivity(), let's cut out
the middle-man and just have pattern_fixed_prefix's subroutines do this
directly. Then pattern_fixed_prefix can return a simple selectivity
number, and the question of how to cope with partial patterns is removed
from its API specification.
While at it, adjust the API spec so that callers who don't actually care
about the pattern's selectivity (which is a lot of them) can pass NULL for
the selectivity pointer to skip doing the work of computing a selectivity
estimate.
This patch is only an API refactoring that doesn't actually change any
processing, other than allowing a little bit of useless work to be skipped.
However, it's necessary infrastructure for my upcoming fix to regex prefix
extraction, because after that change there won't be any simple way to
identify the "rest" of the regex, not even to the low level of fidelity
needed by regex_selectivity. We can cope with that if regex_fixed_prefix
and regex_selectivity communicate directly, but not if we have to work
within the old API. Hence, back-patch to all active branches.
We can do this without creating an API break for estimation functions
by passing the collation using the existing fmgr functionality for
passing an input collation as a hidden parameter.
The need for this was foreseen at the outset, but we didn't get around to
making it happen in 9.1 because of the decision to sort all pg_statistic
histograms according to the database's default collation. That meant that
selectivity estimators generally need to use the default collation too,
even if they're estimating for an operator that will do something
different. The reason it's suddenly become more interesting is that
regexp interpretation also uses a collation (for its LC_TYPE not LC_COLLATE
property), and we no longer want to use the wrong collation when examining
regexps during planning. It's not that the selectivity estimate is likely
to change much from this; rather that we are thinking of caching compiled
regexps during planner estimation, and we won't get the intended benefit
if we cache them with a different collation than the executor will use.
Back-patch to 9.1, both because the regexp change is likely to get
back-patched and because we might as well get this right in all
collation-supporting branches, in case any third-party code wants to
rely on getting the collation. The patch turns out to be minuscule
now that I've done it ...
estimate_num_groups() gets unhappy with
create table empty();
select * from empty except select * from empty e2;
I can't see any actual use-case for such a query (and the table is illegal
per SQL spec), but it seems like a good idea that it not cause an assert
failure.
For those variables only used when asserts are enabled, use a new
macro PG_USED_FOR_ASSERTS_ONLY, which expands to
__attribute__((unused)) when asserts are not enabled.
In constructs such as "x IN (1,2,3,4)" and "x <> ALL(ARRAY[1,2,3,4])",
we formerly always used a general-purpose assumption that the probability
of success is independent for each comparison of "x" to an array element.
But in real-world usage of these constructs, that's a pretty poor
assumption; it's much saner to assume that the array elements are distinct
and so the match probabilities are disjoint. Apply that assumption if the
operator appears to behave as equality (for ANY) or inequality (for ALL).
But fall back to the normal independent-probabilities calculation if this
yields an impossible result, ie probability > 1 or < 0. We could protect
ourselves against bad estimates even more by explicitly checking for equal
array elements, but that is expensive and doesn't seem worthwhile: doing
it would amount to optimizing for poorly-written queries at the expense
of well-written ones.
Daniele Varrazzo and Tom Lane, after a suggestion by Ants Aasma
This patch improves selectivity estimation for the array <@, &&, and @>
(containment and overlaps) operators. It enables collection of statistics
about individual array element values by ANALYZE, and introduces
operator-specific estimators that use these stats. In addition,
ScalarArrayOpExpr constructs of the forms "const = ANY/ALL (array_column)"
and "const <> ANY/ALL (array_column)" are estimated by treating them as
variants of the containment operators.
Since we still collect scalar-style stats about the array values as a
whole, the pg_stats view is expanded to show both these stats and the
array-style stats in separate columns. This creates an incompatible change
in how stats for tsvector columns are displayed in pg_stats: the stats
about lexemes are now displayed in the array-related columns instead of the
original scalar-related columns.
There are a few loose ends here, notably that it'd be nice to be able to
suppress either the scalar-style stats or the array-element stats for
columns for which they're not useful. But the patch is in good enough
shape to commit for wider testing.
Alexander Korotkov, reviewed by Noah Misch and Nathan Boley
Tom Lane