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
Formerly, we just punted when trying to estimate stats for variables coming
out of sub-queries using DISTINCT, on the grounds that whatever stats we
might have for underlying table columns would be inapplicable. But if the
sub-query has only one DISTINCT column, we can consider its output variable
as being unique, which is useful information all by itself. The scope of
this improvement is pretty narrow, but it costs nearly nothing, so we might
as well do it. Per discussion with Andres Freund.
This patch differs from the draft I submitted yesterday in updating various
comments about vardata.isunique (to reflect its extended meaning) and in
tweaking the interaction with security_barrier views. There does not seem
to be a reason why we can't use this sort of knowledge even when the
sub-query is such a view.
btcostestimate() makes an estimate of the number of index tuples that will
be visited based on knowledge of which index clauses can actually bound the
scan within nbtree. However, it forgot to account for partial indexes in
this calculation, with the result that the cost of the index scan could be
significantly overestimated for a partial index. Fix that by merging the
predicate with the abbreviated indexclause list, in the same way as we do
with the full list to estimate how many heap tuples will be visited.
Also, slightly increase the "fudge factor" that's meant to give preference
to smaller indexes over larger ones. While this is applied to all indexes,
it's most important for partial indexes since it can be the only factor
that makes a partial index look cheaper than a similar full index.
Experimentation shows that the existing value is so small as to easily get
swamped by noise such as page-boundary-roundoff behavior. I'm tempted to
kick it up more than this, but will refrain for now.
Per report from Ruben Blanco. These are long-standing issues, but given
the lack of prior complaints I'm not going to risk changing planner
behavior in back branches by back-patching.
This patch fixes the planner so that it can generate nestloop-with-
inner-indexscan plans even with one or more levels of joining between
the indexscan and the nestloop join that is supplying the parameter.
The executor was fixed to handle such cases some time ago, but the
planner was not ready. This should improve our plans in many situations
where join ordering restrictions formerly forced complete table scans.
There is probably a fair amount of tuning work yet to be done, because
of various heuristics that have been added to limit the number of
parameterized paths considered. However, we are not going to find out
what needs to be adjusted until the code gets some real-world use, so
it's time to get it in there where it can be tested easily.
Note API change for index AM amcostestimate functions. I'm not aware of
any non-core index AMs, but if there are any, they will need minor
adjustments.
In commit e2c2c2e8b1 I made use of nested
list structures to show which clauses went with which index columns, but
on reflection that's a data structure that only an old-line Lisp hacker
could love. Worse, it adds unnecessary complication to the many places
that don't much care which clauses go with which index columns. Revert
to the previous arrangement of flat lists of clauses, and instead add a
parallel integer list of column numbers. The places that care about the
pairing can chase both lists with forboth(), while the places that don't
care just examine one list the same as before.
The only real downside to this is that there are now two more lists that
need to be passed to amcostestimate functions in case they care about
column matching (which btcostestimate does, so not passing the info is not
an option). Rather than deal with 11-argument amcostestimate functions,
pass just the IndexPath and expect the functions to extract fields from it.
That gets us down to 7 arguments which is better than 11, and it seems
more future-proof against likely additions to the information we keep
about an index path.
It's potentially useful for an index to repeat the same indexable column
or expression in multiple index columns, if the columns have different
opclasses. (If they share opclasses too, the duplicate column is pretty
useless, but nonetheless we've allowed such cases since 9.0.) However,
the planner failed to cope with this, because createplan.c was relying on
simple equal() matching to figure out which index column each index qual
is intended for. We do have that information available upstream in
indxpath.c, though, so the fix is to not flatten the multi-level indexquals
list when putting it into an IndexPath. Then we can rely on the sublist
structure to identify target index columns in createplan.c. There's a
similar issue for index ORDER BYs (the KNNGIST feature), so introduce a
multi-level-list representation for that too. This adds a bit more
representational overhead, but we might more or less buy that back by not
having to search for matching index columns anymore in createplan.c;
likewise btcostestimate saves some cycles.
Per bug #6351 from Christian Rudolph. Likely symptoms include the "btree
index keys must be ordered by attribute" failure shown there, as well as
"operator MMMM is not a member of opfamily NNNN".
Although this is a pre-existing problem that can be demonstrated in 9.0 and
9.1, I'm not going to back-patch it, because the API changes in the planner
seem likely to break things such as index plugins. The corner cases where
this matters seem too narrow to justify possibly breaking things in a minor
release.
When a view is marked as a security barrier, it will not be pulled up
into the containing query, and no quals will be pushed down into it,
so that no function or operator chosen by the user can be applied to
rows not exposed by the view. Views not configured with this
option cannot provide robust row-level security, but will perform far
better.
Patch by KaiGai Kohei; original problem report by Heikki Linnakangas
(in October 2009!). Review (in earlier versions) by Noah Misch and
others. Design advice by Tom Lane and myself. Further review and
cleanup by me.
The original coding of this function overlooked the possibility that
it could be passed anything except simple OpExpr indexquals. But
ScalarArrayOpExpr is possible too, and the code would probably crash
(and surely give ridiculous answers) in such a case. Add logic to try
to estimate sanely for such cases.
In passing, fix the treatment of inner-indexscan cost estimation: it was
failing to scale up properly for multiple iterations of a nestloop.
(I think somebody might've thought that index_pages_fetched() is linear,
but of course it's not.)
Report, diagnosis, and preliminary patch by Marti Raudsepp; I refactored
it a bit and fixed the cost estimation.
Back-patch into 9.1 where the bogus code was introduced.
SP-GiST is comparable to GiST in flexibility, but supports non-balanced
partitioned search structures rather than balanced trees. As described at
PGCon 2011, this new indexing structure can beat GiST in both index build
time and query speed for search problems that it is well matched to.
There are a number of areas that could still use improvement, but at this
point the code seems committable.
Teodor Sigaev and Oleg Bartunov, with considerable revisions by Tom Lane
Make sure that it considers all the possibilities that the old code did,
instead of trying only one possibility per character position. To keep the
runtime in bounds, instead tweak the character incrementers to not try
every possible multibyte character code. Remove unnecessary logic to
restore the old character value on failure. Additional comment and
formatting cleanup.
This infrastructure doesn't in any way guarantee that the character
we produce will sort before the one we incremented; but it does at least
make it much more likely that we'll end up with something that is a valid
character, which improves our chances.
Kyotaro Horiguchi, with various adjustments by me.
The uniqueness condition might fail to hold intra-transaction, and assuming
it does can give incorrect query results. Per report from Marti Raudsepp,
though this is not his proposed patch.
Back-patch to 9.0, where both these features were introduced. In the
released branches, add the new IndexOptInfo field to the end of the struct,
to try to minimize ABI breakage for third-party code that may be examining
that struct.
This code was looking at the sub-Query tree as seen in the parent query's
RangeTblEntry; but that's the pristine parser output, and what we need to
look at is the tree as it stands at the completion of planning. Otherwise
we might pick up a Var that references a subquery that got flattened and
hence has no RelOptInfo in the subroot. Per report from Peter Geoghegan.
As per my recent proposal, this refactors things so that these typedefs and
macros are available in a header that can be included in frontend-ish code.
I also changed various headers that were undesirably including
utils/timestamp.h to include datatype/timestamp.h instead. Unsurprisingly,
this showed that half the system was getting utils/timestamp.h by way of
xlog.h.
No actual code changes here, just header refactoring.
If a sub-select's output column is a simple Var, recursively look for
statistics applying to that Var, and use them if available. The need for
this was foreseen ages ago, but we didn't have enough infrastructure to do
it with reasonable speed until just now.
We punt and stick with default estimates if the subquery uses set
operations, GROUP BY, or DISTINCT, since those operations would change the
underlying column statistics (particularly, the relative frequencies of
different values) beyond recognition. This means that the types of
sub-selects for which this improvement applies are fairly limited, since
most subqueries satisfying those restrictions would have gotten flattened
into the parent query anyway. But it does help for some cases, such as
subqueries with ORDER BY or LIMIT.
Examination of examples provided by Mark Kirkwood and others has convinced
me that actually commit 7f3eba30c9 was quite
a few bricks shy of a load. The useful part of that patch was clamping
ndistinct for the inner side of a semi or anti join, and the reason why
that's needed is that it's the only way that restriction clauses
eliminating rows from the inner relation can affect the estimated size of
the join result. I had not clearly understood why the clamping was
appropriate, and so mis-extrapolated to conclude that we should clamp
ndistinct for the outer side too, as well as for both sides of regular
joins. These latter actions were all wrong, and are reverted with this
patch. In addition, the clamping logic is now made to affect the behavior
of both paths in eqjoinsel_semi, with or without MCV lists to compare.
When we have MCVs, we suppose that the most common values are the ones
that are most likely to survive the decimation resulting from a lower
restriction clause, so we think of the clamping as eliminating non-MCV
values, or potentially even the least-common MCVs for the inner relation.
Back-patch to 8.4, same as previous fixes in this area.
This patch fixes an oversight in my commit
7f3eba30c9 of 2008-10-23. That patch
accounted for baserel restriction clauses that reduced the number of rows
coming out of a table (and hence the number of possibly-distinct values of
a join variable), but not for join restriction clauses that might have been
applied at a lower level of join. To account for the latter, look up the
sizes of the min_lefthand and min_righthand inputs of the current join,
and clamp with those in the same way as for the base relations.
Noted while investigating a complaint from Ben Chobot, although this in
itself doesn't seem to explain his report.
Back-patch to 8.4; previous versions used different estimation methods
for which this heuristic isn't relevant.
Regular aggregate functions in combination with, or within the arguments
of, window functions are OK per spec; they have the semantics that the
aggregate output rows are computed and then we run the window functions
over that row set. (Thus, this combination is not really useful unless
there's a GROUP BY so that more than one aggregate output row is possible.)
The case without GROUP BY could fail, as recently reported by Jeff Davis,
because sloppy construction of the Agg node's targetlist resulted in extra
references to possibly-ungrouped Vars appearing outside the aggregate
function calls themselves. See the added regression test case for an
example.
Fixing this requires modifying the API of flatten_tlist and its underlying
function pull_var_clause. I chose to make pull_var_clause's API for
aggregates identical to what it was already doing for placeholders, since
the useful behaviors turn out to be the same (error, report node as-is, or
recurse into it). I also tightened the error checking in this area a bit:
if it was ever valid to see an uplevel Var, Aggref, or PlaceHolderVar here,
that was a long time ago, so complain instead of ignoring them.
Backpatch into 9.1. The failure exists in 8.4 and 9.0 as well, but seeing
that it only occurs in a basically-useless corner case, it doesn't seem
worth the risks of changing a function API in a minor release. There might
be third-party code using pull_var_clause.
Since collation is effectively an argument, not a property of the function,
FmgrInfo is really the wrong place for it; and this becomes critical in
cases where a cached FmgrInfo is used for varying purposes that might need
different collation settings. Fix by passing it in FunctionCallInfoData
instead. In particular this allows a clean fix for bug #5970 (record_cmp
not working). This requires touching a bit more code than the original
method, but nobody ever thought that collations would not be an invasive
patch...
In particular, if we don't have real ndistinct estimates for both sides,
fall back to assuming that half of the left-hand rows have join partners.
This is what was done in 8.2 and 8.3 (cf nulltestsel() in those versions).
It's pretty stupid but it won't lead us to think that an antijoin produces
no rows out, as seen in recent example from Uwe Schroeder.
This warning is new in gcc 4.6 and part of -Wall. This patch cleans
up most of the noise, but there are some still warnings that are
trickier to remove.
This is necessary, not optional, now that ILIKE and regexes are collation
aware --- else we might derive a wrong comparison constant for index
optimized pattern matches.
The previous coding worked only if ltproc->fn_collation was always either
DEFAULT_COLLATION_OID or a C-compatible locale. While that's true at the
moment, it wasn't documented (and in fact wasn't true when this code was
committed...). But it only takes a couple more lines to make its internal
caching behavior locale-aware, so let's do that.
In nearly all cases, the caller already knows the correct collation, and
in a number of places, the value the caller has handy is more correct than
the default for the type would be. (In particular, this patch makes it
significantly less likely that eval_const_expressions will result in
changing the exposed collation of an expression.) So an internal lookup
is both expensive and wrong.
All expression nodes now have an explicit output-collation field, unless
they are known to only return a noncollatable data type (such as boolean
or record). Also, nodes that can invoke collation-aware functions store
a separate field that is the collation value to pass to the function.
This avoids confusion that arises when a function has collatable inputs
and noncollatable output type, or vice versa.
Also, replace the parser's on-the-fly collation assignment method with
a post-pass over the completed expression tree. This allows us to use
a more complex (and hopefully more nearly spec-compliant) assignment
rule without paying for it in extra storage in every expression node.
Fix assorted bugs in the planner's handling of collations by making
collation one of the defining properties of an EquivalenceClass and
by converting CollateExprs into discardable RelabelType nodes during
expression preprocessing.
While this will give wrong answers when estimating selectivity for a
comparison operator that's using a non-default collation, the estimation
error probably won't be large; and anyway the former approach created
estimation errors of its own by trying to use a histogram that might have
been computed with some other collation. So we'll adopt this simplified
approach for now and perhaps improve it sometime in the future.
This patch incorporates changes from Andres Freund to make sure that
selfuncs.c passes a valid collation OID to any datatype-specific function
it calls, in case that function wants collation information. Said OID will
now always be DEFAULT_COLLATION_OID, but at least we won't get errors.
That function was supposing that indexoid == 0 for a hypothetical index,
but that is not likely to be true in any non-toy implementation of an index
adviser, since assigning a fake OID is the only way to know at EXPLAIN time
which hypothetical index got selected. Fix by adding a flag to
IndexOptInfo to mark hypothetical indexes. Back-patch to 9.0 where
get_actual_variable_range() was added.
Gurjeet Singh
This adds collation support for columns and domains, a COLLATE clause
to override it per expression, and B-tree index support.
Peter Eisentraut
reviewed by Pavel Stehule, Itagaki Takahiro, Robert Haas, Noah Misch
The only reason this wasn't crashing while testing the core anyarray
operators was that it was disabled for those cases because of passing the
wrong type information to get_opfamily_proc :-(. So fix that too, and make
it insist on finding the support proc --- in hindsight, silently doing
nothing is not as sane a coping mechanism as all that.
The "date" type supports a wider range of dates than int64 timestamps do.
However, there is pre-int64-timestamp code in the planner that assumes that
all date values can be converted to timestamp with impunity. Fortunately,
what we really need out of the conversion is always a double (float8)
value; so even when the date is out of timestamp's range it's possible to
produce a sane answer. All we need is a code path that doesn't try to
force the result into int64. Per trouble report from David Rericha.
Back-patch to all supported versions. Although this is surely a corner
case, there's not much point in advertising a date range wider than
timestamp's if we will choke on such values in unexpected places.
This is a heavily revised version of builtin_knngist_core-0.9. The
ordering operators are no longer mixed in with actual quals, which would
have confused not only humans but significant parts of the planner.
Instead, ordering operators are carried separately throughout planning and
execution.
Since the API for ambeginscan and amrescan functions had to be changed
anyway, this commit takes the opportunity to rationalize that a bit.
RelationGetIndexScan no longer forces a premature index_rescan call;
instead, callers of index_beginscan must call index_rescan too. Aside from
making the AM-side initialization logic a bit less peculiar, this has the
advantage that we do not make a useless extra am_rescan call when there are
runtime key values. AMs formerly could not assume that the key values
passed to amrescan were actually valid; now they can.
Teodor Sigaev and Tom Lane
Formerly we looked up the operators associated with each index (caching
them in relcache) and then the planner looked up the btree opfamily
containing such operators in order to build the btree-centric pathkey
representation that describes the index's sort order. This is quite
pointless for btree indexes: we might as well just use the index's opfamily
information directly. That saves syscache lookup cycles during planning,
and furthermore allows us to eliminate the relcache's caching of operators
altogether, which may help in reducing backend startup time.
I added code to plancat.c to perform the same type of double lookup
on-the-fly if it's ever faced with a non-btree amcanorder index AM.
If such a thing actually becomes interesting for production, we should
replace that logic with some more-direct method for identifying the
corresponding btree opfamily; but it's not worth spending effort on now.
There is considerably more to do pursuant to my recent proposal to get rid
of sort-operator-based representations of sort orderings, but this patch
grabs some of the low-hanging fruit. I'll look at the remainder of that
work after the current commitfest.
This patch eliminates various bizarre behaviors caused by sloppy thinking
about the difference between a domain type and its underlying array type.
In particular, the operation of updating one element of such an array
has to be considered as yielding a value of the underlying array type,
*not* a value of the domain, because there's no assurance that the
domain's CHECK constraints are still satisfied. If we're intending to
store the result back into a domain column, we have to re-cast to the
domain type so that constraints are re-checked.
For similar reasons, such a domain can't be blindly matched to an ANYARRAY
polymorphic parameter, because the polymorphic function is likely to apply
array-ish operations that could invalidate the domain constraints. For the
moment, we just forbid such matching. We might later wish to insert an
automatic downcast to the underlying array type, but such a change should
also change matching of domains to ANYELEMENT for consistency.
To ensure that all such logic is rechecked, this patch removes the original
hack of setting a domain's pg_type.typelem field to match its base type;
the typelem will always be zero instead. In those places where it's really
okay to look through the domain type with no other logic changes, use the
newly added get_base_element_type function in place of get_element_type.
catversion bumped due to change in pg_type contents.
Per bug #5717 from Richard Huxton and subsequent discussion.
The better estimate requires more statistics than we previously stored:
in particular, counts of "entry" versus "data" pages within the index,
as well as knowledge of the number of distinct key values. We collect
this information during initial index build and update it during VACUUM,
storing the info in new fields on the index metapage. No initdb is
required because these fields will read as zeroes in a pre-existing
index, and the new gincostestimate code is coded to behave (reasonably)
sanely if they are zeroes.
Teodor Sigaev, reviewed by Jan Urbanski, Tom Lane, and Itagaki Takahiro.
The purpose of this change is to eliminate the need for every caller
of SearchSysCache, SearchSysCacheCopy, SearchSysCacheExists,
GetSysCacheOid, and SearchSysCacheList to know the maximum number
of allowable keys for a syscache entry (currently 4). This will
make it far easier to increase the maximum number of keys in a
future release should we choose to do so, and it makes the code
shorter, too.
Design and review by Tom Lane.
This patch only supports seq_page_cost and random_page_cost as parameters,
but it provides the infrastructure to scalably support many more.
In particular, we may want to add support for effective_io_concurrency,
but I'm leaving that as future work for now.
Thanks to Tom Lane for design help and Alvaro Herrera for the review.
"column < constant", and the comparison value is in the first or last
histogram bin or outside the histogram entirely, try to fetch the actual
column min or max value using an index scan (if there is an index on the
column). If successful, replace the lower or upper histogram bound with
that value before carrying on with the estimate. This limits the
estimation error caused by moving min/max values when the comparison
value is close to the min or max. Per a complaint from Josh Berkus.
It is tempting to consider using this mechanism for mergejoinscansel as well,
but that would inject index fetches into main-line join estimation not just
endpoint cases. I'm refraining from that until we can get a better handle
on the costs of doing this type of lookup.
to be just a minor extension of the previous patch that made "x IS NULL"
indexable, because we can treat the IS NOT NULL condition as if it were
"x < NULL" or "x > NULL" (depending on the index's NULLS FIRST/LAST option),
just like IS NULL is treated like "x = NULL". Aside from any possible
usefulness in its own right, this is an important improvement for
index-optimized MAX/MIN aggregates: it is now reliably possible to get
a column's min or max value cheaply, even when there are a lot of nulls
cluttering the interesting end of the index.
and teach ANALYZE to compute such stats for tables that have subclasses.
Per my proposal of yesterday.
autovacuum still needs to be taught about running ANALYZE on parent tables
when their subclasses change, but the feature is useful even without that.
style by default. Per discussion, there seems to be hardly anything that
really relies on being able to change the regex flavor, so the ability to
select it via embedded options ought to be enough for any stragglers.
Also, if we didn't remove the GUC, we'd really be morally obligated to
mark the regex functions non-immutable, which'd possibly create performance
issues.
Both hex format and the traditional "escape" format are automatically
handled on input. The output format is selected by the new GUC variable
bytea_output.
As committed, bytea_output defaults to HEX, which is an *incompatible
change*. We will keep it this way for awhile for testing purposes, but
should consider whether to switch to the more backwards-compatible
default of ESCAPE before 8.5 is released.
Peter Eisentraut