Evaluation of set returning functions (SRFs_ in the targetlist (like SELECT
generate_series(1,5)) so far was done in the expression evaluation (i.e.
ExecEvalExpr()) and projection (i.e. ExecProject/ExecTargetList) code.
This meant that most executor nodes performing projection, and most
expression evaluation functions, had to deal with the possibility that an
evaluated expression could return a set of return values.
That's bad because it leads to repeated code in a lot of places. It also,
and that's my (Andres's) motivation, made it a lot harder to implement a
more efficient way of doing expression evaluation.
To fix this, introduce a new executor node (ProjectSet) that can evaluate
targetlists containing one or more SRFs. To avoid the complexity of the old
way of handling nested expressions returning sets (e.g. having to pass up
ExprDoneCond, and dealing with arguments to functions returning sets etc.),
those SRFs can only be at the top level of the node's targetlist. The
planner makes sure (via split_pathtarget_at_srfs()) that SRF evaluation is
only necessary in ProjectSet nodes and that SRFs are only present at the
top level of the node's targetlist. If there are nested SRFs the planner
creates multiple stacked ProjectSet nodes. The ProjectSet nodes always get
input from an underlying node.
We also discussed and prototyped evaluating targetlist SRFs using ROWS
FROM(), but that turned out to be more complicated than we'd hoped.
While moving SRF evaluation to ProjectSet would allow to retain the old
"least common multiple" behavior when multiple SRFs are present in one
targetlist (i.e. continue returning rows until all SRFs are at the end of
their input at the same time), we decided to instead only return rows till
all SRFs are exhausted, returning NULL for already exhausted ones. We
deemed the previous behavior to be too confusing, unexpected and actually
not particularly useful.
As a side effect, the previously prohibited case of multiple set returning
arguments to a function, is now allowed. Not because it's particularly
desirable, but because it ends up working and there seems to be no argument
for adding code to prohibit it.
Currently the behavior for COALESCE and CASE containing SRFs has changed,
returning multiple rows from the expression, even when the SRF containing
"arm" of the expression is not evaluated. That's because the SRFs are
evaluated in a separate ProjectSet node. As that's quite confusing, we're
likely to instead prohibit SRFs in those places. But that's still being
discussed, and the code would reside in places not touched here, so that's
a task for later.
There's a lot of, now superfluous, code dealing with set return expressions
around. But as the changes to get rid of those are verbose largely boring,
it seems better for readability to keep the cleanup as a separate commit.
Author: Tom Lane and Andres Freund
Discussion: https://postgr.es/m/20160822214023.aaxz5l4igypowyri@alap3.anarazel.de
create_merge_append_path should use the path rowcount it just computed,
not rel->tuples, for costing purposes. Those numbers should always be
the same at present, but if we ever support parameterized MergeAppend
paths (a case this function is otherwise prepared for), the former would
be right and the latter wrong.
No need for back-patch since the problem is only latent.
Ashutosh Bapat
Discussion: <CAFjFpRek+cLCnTo24youuGtsq4zRphEB8EUUPjDxZjnL4n4HYQ@mail.gmail.com>
We need to scan the whole parse tree for parallel-unsafe functions.
If there are none, we'll later need to determine whether particular
subtrees contain any parallel-restricted functions. The previous coding
retained no knowledge from the first scan, even though this is very
wasteful in the common case where the query contains only parallel-safe
functions. We can bypass all of the later scans by remembering that fact.
This provides a small but measurable speed improvement when the case
applies, and shouldn't cost anything when it doesn't.
Patch by me, reviewed by Robert Haas
Discussion: <3740.1471538387@sss.pgh.pa.us>
Commit 3fc6e2d7f5 introduced new "upper"
RelOptInfo structures but didn't set consider_parallel for them
correctly, a point I completely missed when reviewing it. Later,
commit e06a38965b made the situation
worse by doing it incorrectly for the grouping relation. Try to
straighten all of that out. Along the way, get rid of the annoying
wholePlanParallelSafe flag, which was only necessarily because of
the fact that upper planning stages didn't use paths at the time
that code was written.
The most important immediate impact of these changes is that
force_parallel_mode will provide useful test coverage in quite a few
more scenarios than it did previously, but it's also necessary
preparation for fixing some problems related to subqueries.
Patch by me, reviewed by Tom Lane.
The original coding had three separate booleans representing partial
aggregation behavior, which was confusing, unreadable, and error-prone,
not least because the booleans weren't always listed in the same order.
It was also inadequate for the allegedly-desirable future extension to
support intermediate partial aggregation, because we'd need separate
markers for serialization and deserialization in such a case.
Merge these bools into an enum "AggSplit" to provide symbolic names for
the supported operating modes (and document what those are). By assigning
the values of the enum constants carefully, we can treat AggSplit values
as options bitmasks so that tests of what to do aren't noticeably more
expensive than before.
While at it, get rid of Aggref.aggoutputtype. That's not needed since
commit 59a3795c2 got rid of setrefs.c's special-purpose Aggref comparison
code, and it likewise seemed more confusing than helpful.
Assorted comment cleanup as well (there's still more that I want to do
in that line).
catversion bump for change in Aggref node contents. Should be the last
one for partial-aggregation changes.
Discussion: <29309.1466699160@sss.pgh.pa.us>
The original upper-planner-pathification design (commit 3fc6e2d7f5)
assumed that we could always determine during Path formation whether or not
we would need a Result plan node to perform projection of a targetlist.
That turns out not to work very well, though, because createplan.c still
has some responsibilities for choosing the specific target list associated
with sorting/grouping nodes (in particular it might choose to add resjunk
columns for sorting). We might not ever refactor that --- doing so would
push more work into Path formation, which isn't attractive --- and we
certainly won't do so for 9.6. So, while create_projection_path and
apply_projection_to_path can tell for sure what will happen if the subpath
is projection-capable, they can't tell for sure when it isn't. This is at
least a latent bug in apply_projection_to_path, which might think it can
apply a target to a non-projecting node when the node will end up computing
something different.
Also, I'd tied the creation of a ProjectionPath node to whether or not a
Result is needed, but it turns out that we sometimes need a ProjectionPath
node anyway to avoid modifying a possibly-shared subpath node. Callers had
to use create_projection_path for such cases, and we added code to them
that knew about the potential omission of a Result node and attempted to
adjust the cost estimates for that. That was uncertainly correct and
definitely ugly/unmaintainable.
To fix, have create_projection_path explicitly check whether a Result
is needed and adjust its cost estimate accordingly, though it creates
a ProjectionPath in either case. apply_projection_to_path is now mostly
just an optimized version that can avoid creating an extra Path node when
the input is known to not be shared with any other live path. (There
is one case that create_projection_path doesn't handle, which is pushing
parallel-safe expressions below a Gather node. We could make it do that
by duplicating the GatherPath, but there seems no need as yet.)
create_projection_plan still has to recheck the tlist-match condition,
which means that if the matching situation does get changed by createplan.c
then we'll have made a slightly incorrect cost estimate. But there seems
no help for that in the near term, and I doubt it occurs often enough,
let alone would change planning decisions often enough, to be worth
stressing about.
I added a "dummypp" field to ProjectionPath to track whether
create_projection_path thinks a Result is needed. This is not really
necessary as-committed because create_projection_plan doesn't look at the
flag; but it seems like a good idea to remember what we thought when
forming the cost estimate, if only for debugging purposes.
In passing, get rid of the target_parallel parameter added to
apply_projection_to_path by commit 54f5c5150. I don't think that's a good
idea because it involves callers in what should be an internal decision,
and opens us up to missing optimization opportunities if callers think they
don't need to provide a valid flag, as most don't. For the moment, this
just costs us an extra has_parallel_hazard call when planning a Gather.
If that starts to look expensive, I think a better solution would be to
teach PathTarget to carry/cache knowledge of parallel-safety of its
contents.
Commit 04ae11f62e removed some broken
code to apply the scan/join target to partial paths, but its theory
that this processing step is totally unnecessary turns out to be wrong.
Put similar code back again, but this time, check for parallel-safety
and avoid in-place modifications to paths that may already have been
used as part of some other path.
(This is not an entirely elegant solution to this problem; it might
be better, for example, to postpone generate_gather_paths for the
topmost scan/join rel until after the scan/join target has been
applied. But this is not the time for such redesign work.)
Amit Kapila and Robert Haas
This terminology provoked widespread complaints. So, instead, rename
the GUC max_parallel_degree to max_parallel_workers_per_gather
(leaving room for a possible future GUC max_parallel_workers that acts
as a system-wide limit), and rename the parallel_degree reloption to
parallel_workers. Rename structure members to match.
These changes create a dump/restore hazard for users of PostgreSQL
9.6beta1 who have set the reloption (or applied the GUC using ALTER
USER or ALTER DATABASE).
We mustn't run generate_gather_paths() during add_paths_to_joinrel(),
because that function can be invoked multiple times for the same target
joinrel. Not only is it wasteful to build GatherPaths repeatedly, but
a later add_partial_path() could delete the partial path that a previously
created GatherPath depends on. Instead establish the convention that we
do generate_gather_paths() for a rel only just before set_cheapest().
The code was accidentally not broken for baserels, because as of today there
never is more than one partial path for a baserel. But that assumption
obviously has a pretty short half-life, so move the generate_gather_paths()
calls for those cases as well.
Also add some generic comments explaining how and why this all works.
Per fuzz testing by Andreas Seltenreich.
Report: <871t5pgwdt.fsf@credativ.de>
This is necessary infrastructure for supporting parallel aggregation
for aggregates whose transition type is "internal". Such values
can't be passed between cooperating processes, because they are
just pointers.
David Rowley, reviewed by Tomas Vondra and by me.
Parallel workers can now partially aggregate the data and pass the
transition values back to the leader, which can combine the partial
results to produce the final answer.
David Rowley, based on earlier work by Haribabu Kommi. Reviewed by
Álvaro Herrera, Tomas Vondra, Amit Kapila, James Sewell, and me.
This means that, for example, "SELECT expensive_func(a) FROM bigtab
WHERE something" can compute expensive_func(a) in the workers rather
than the leader if it happens to be parallel-safe, which figures to be
a big win in some practical cases.
Currently, we can only do this if the entire target list is
parallel-safe. If we worked harder, we might be able to evaluate
parallel-safe targets in the worker and any parallel-restricted
targets in the leader, but that would be more complicated, and there
aren't that many parallel-restricted functions that people are likely
to use in queries anyway. I think. So just do the simple thing for
the moment.
Robert Haas, Amit Kapila, and Tom Lane
This is basically like the just-added create_upper_paths_hook, but
control is funneled only to the FDW responsible for all the baserels
of the current query; so providing such a callback is much less likely
to add useless overhead than using the hook function is.
The documentation is a bit sketchy. We'll likely want to improve it,
and/or adjust the call conventions, when we get some experience with
actually using this callback. Hopefully somebody will find time to
experiment with it before 9.6 feature freeze.
Although the default choice of rel->reltarget should typically be
sufficient for scan or join paths, it's not at all sufficient for the
purposes PathTargets were invented for; in particular not for
upper-relation Paths. So break API compatibility by adding a PathTarget
argument to create_foreignscan_path(). To ease updating of existing
code, accept a NULL value of the argument as selecting rel->reltarget.
In commit 19a541143a I did not make PathTarget a subtype of Node,
and embedded a RelOptInfo's reltarget directly into it rather than having
a separately-allocated Node. In hindsight that was misguided
micro-optimization, enabled by the fact that at that point we didn't have
any Paths with custom PathTargets. Now that PathTarget processing has
been fleshed out some more, it's easier to see that it's better to have
PathTarget as an indepedent Node type, even if it does cost us one more
palloc to create a RelOptInfo. So change it while we still can.
This commit just changes the representation, without doing anything more
interesting than that.
Instead of having planner.c compute a groupColIdx array and store it in
GroupingSetsPaths, make create_groupingsets_plan() find the grouping
columns by searching in the child plan node's tlist. Although that's
probably a bit slower for create_groupingsets_plan(), it's more like
the way every other plan node type does this, and it provides positive
confirmation that we know which child output columns we're supposed to be
grouping on. (Indeed, looking at this now, I'm not at all sure that it
wasn't broken before, because create_groupingsets_plan() isn't demanding
an exact tlist match from its child node.) Also, this allows substantial
simplification in planner.c, because it no longer needs to compute the
groupColIdx array at all; no other cases were using it.
I'd intended to put off this refactoring until later (like 9.7), but
in view of the likely bug fix and the need to rationalize planner.c's
tlist handling so we can do something sane with Konstantin Knizhnik's
function-evaluation-postponement patch, I think it can't wait.
This patch removes some redundant cost calculations that I left for later
cleanup in commit 3fc6e2d7f5. There's now a uniform policy that the
make_foo() convenience functions don't do any cost calculations. Most of
their callers copy costs from the source Path node, and for those that
don't, the calculation in the make_foo() function wasn't necessarily right
anyhow. (make_result() was particularly a mess, as it was serving multiple
callers using cost calcs designed for only the first one or two that had
ever existed.) Aside from saving a few cycles, this ensures that what
EXPLAIN prints matches the costs we used for planning purposes. It does
not change any planner decisions, since the decisions are already made.
I've been saying we needed to do this for more than five years, and here it
finally is. This patch removes the ever-growing tangle of spaghetti logic
that grouping_planner() used to use to try to identify the best plan for
post-scan/join query steps. Now, there is (nearly) independent
consideration of each execution step, and entirely separate construction of
Paths to represent each of the possible ways to do that step. We choose
the best Path or set of Paths using the same add_path() logic that's been
used inside query_planner() for years.
In addition, this patch removes the old restriction that subquery_planner()
could return only a single Plan. It now returns a RelOptInfo containing a
set of Paths, just as query_planner() does, and the parent query level can
use each of those Paths as the basis of a SubqueryScanPath at its level.
This allows finding some optimizations that we missed before, wherein a
subquery was capable of returning presorted data and thereby avoiding a
sort in the parent level, making the overall cost cheaper even though
delivering sorted output was not the cheapest plan for the subquery in
isolation. (A couple of regression test outputs change in consequence of
that. However, there is very little change in visible planner behavior
overall, because the point of this patch is not to get immediate planning
benefits but to create the infrastructure for future improvements.)
There is a great deal left to do here. This patch unblocks a lot of
planner work that was basically impractical in the old code structure,
such as allowing FDWs to implement remote aggregation, or rewriting
plan_set_operations() to allow consideration of multiple implementation
orders for set operations. (The latter will likely require a full
rewrite of plan_set_operations(); what I've done here is only to fix it
to return Paths not Plans.) I have also left unfinished some localized
refactoring in createplan.c and planner.c, because it was not necessary
to get this patch to a working state.
Thanks to Robert Haas, David Rowley, and Amit Kapila for review.
Up to now, there's been an assumption that all Paths for a given relation
compute the same output column set (targetlist). However, there are good
reasons to remove that assumption. For example, an indexscan on an
expression index might be able to return the value of an expensive function
"for free". While we have the ability to generate such a plan today in
simple cases, we don't have a way to model that it's cheaper than a plan
that computes the function from scratch, nor a way to create such a plan
in join cases (where the function computation would normally happen at
the topmost join node). Also, we need this so that we can have Paths
representing post-scan/join steps, where the targetlist may well change
from one step to the next. Therefore, invent a "struct PathTarget"
representing the columns we expect a plan step to emit. It's convenient
to include the output tuple width and tlist evaluation cost in this struct,
and there will likely be additional fields in future.
While Path nodes that actually do have custom outputs will need their own
PathTargets, it will still be true that most Paths for a given relation
will compute the same tlist. To reduce the overhead added by this patch,
keep a "default PathTarget" in RelOptInfo, and allow Paths that compute
that column set to just point to their parent RelOptInfo's reltarget.
(In the patch as committed, actually every Path is like that, since we
do not yet have any cases of custom PathTargets.)
I took this opportunity to provide some more-honest costing of
PlaceHolderVar evaluation. Up to now, the assumption that "scan/join
reltargetlists have cost zero" was applied not only to Vars, where it's
reasonable, but also PlaceHolderVars where it isn't. Now, we add the eval
cost of a PlaceHolderVar's expression to the first plan level where it can
be computed, by including it in the PathTarget cost field and adding that
to the cost estimates for Paths. This isn't perfect yet but it's much
better than before, and there is a way forward to improve it more. This
costing change affects the join order chosen for a couple of the regression
tests, changing expected row ordering.
The core innovation of this patch is the introduction of the concept
of a partial path; that is, a path which if executed in parallel will
generate a subset of the output rows in each process. Gathering a
partial path produces an ordinary (complete) path. This allows us to
generate paths for parallel joins by joining a partial path for one
side (which at the baserel level is currently always a Partial Seq
Scan) to an ordinary path on the other side. This is subject to
various restrictions at present, especially that this strategy seems
unlikely to be sensible for merge joins, so only nested loops and
hash joins paths are generated.
This also allows an Append node to be pushed below a Gather node in
the case of a partitioned table.
Testing revealed that early versions of this patch made poor decisions
in some cases, which turned out to be caused by the fact that the
original cost model for Parallel Seq Scan wasn't very good. So this
patch tries to make some modest improvements in that area.
There is much more to be done in the area of generating good parallel
plans in all cases, but this seems like a useful step forward.
Patch by me, reviewed by Dilip Kumar and Amit Kapila.
Commit e7cb7ee145 provided basic
infrastructure for allowing a foreign data wrapper or custom scan
provider to replace a join of one or more tables with a scan.
However, this infrastructure failed to take into account the need
for possible EvalPlanQual rechecks, and ExecScanFetch would fail
an assertion (or just overwrite memory) if such a check was attempted
for a plan containing a pushed-down join. To fix, adjust the EPQ
machinery to skip some processing steps when scanrelid == 0, making
those the responsibility of scan's recheck method, which also has
the responsibility in this case of correctly populating the relevant
slot.
To allow foreign scans to gain control in the right place to make
use of this new facility, add a new, optional RecheckForeignScan
method. Also, allow a foreign scan to have a child plan, which can
be used to correctly populate the slot (or perhaps for something
else, but this is the only use currently envisioned).
KaiGai Kohei, reviewed by Robert Haas, Etsuro Fujita, and Kyotaro
Horiguchi.
In addition, this path fills in a number of missing bits and pieces in
the parallel infrastructure. Paths and plans now have a parallel_aware
flag indicating whether whatever parallel-aware logic they have should
be engaged. It is believed that we will need this flag for a number of
path/plan types, not just sequential scans, which is why the flag is
generic rather than part of the SeqScan structures specifically.
Also, execParallel.c now gives parallel nodes a chance to initialize
their PlanState nodes from the DSM during parallel worker startup.
Amit Kapila, with a fair amount of adjustment by me. Review of previous
patch versions by Haribabu Kommi and others.
A Gather executor node runs any number of copies of a plan in an equal
number of workers and merges all of the results into a single tuple
stream. It can also run the plan itself, if the workers are
unavailable or haven't started up yet. It is intended to work with
the Partial Seq Scan node which will be added in future commits.
It could also be used to implement parallel query of a different sort
by itself, without help from Partial Seq Scan, if the single_copy mode
is used. In that mode, a worker executes the plan, and the parallel
leader does not, merely collecting the worker's results. So, a Gather
node could be inserted into a plan to split the execution of that plan
across two processes. Nested Gather nodes aren't currently supported,
but we might want to add support for that in the future.
There's nothing in the planner to actually generate Gather nodes yet,
so it's not quite time to break out the champagne. But we're getting
close.
Amit Kapila. Some designs suggestions were provided by me, and I also
reviewed the patch. Single-copy mode, documentation, and other minor
changes also by me.
The original implementation of TABLESAMPLE modeled the tablesample method
API on index access methods, which wasn't a good choice because, without
specialized DDL commands, there's no way to build an extension that can
implement a TSM. (Raw inserts into system catalogs are not an acceptable
thing to do, because we can't undo them during DROP EXTENSION, nor will
pg_upgrade behave sanely.) Instead adopt an API more like procedural
language handlers or foreign data wrappers, wherein the only SQL-level
support object needed is a single handler function identified by having
a special return type. This lets us get rid of the supporting catalog
altogether, so that no custom DDL support is needed for the feature.
Adjust the API so that it can support non-constant tablesample arguments
(the original coding assumed we could evaluate the argument expressions at
ExecInitSampleScan time, which is undesirable even if it weren't outright
unsafe), and discourage sampling methods from looking at invisible tuples.
Make sure that the BERNOULLI and SYSTEM methods are genuinely repeatable
within and across queries, as required by the SQL standard, and deal more
honestly with methods that can't support that requirement.
Make a full code-review pass over the tablesample additions, and fix
assorted bugs, omissions, infelicities, and cosmetic issues (such as
failure to put the added code stanzas in a consistent ordering).
Improve EXPLAIN's output of tablesample plans, too.
Back-patch to 9.5 so that we don't have to support the original API
in production.
add_path_precheck was doing exact comparisons of path costs, but it really
needs to do them fuzzily to be sure it won't reject paths that could
survive add_path's comparisons. (This can only matter if the initial cost
estimate is very close to the final one, but that turns out to often be
true.)
Also, it should ignore startup cost for this purpose if and only if
compare_path_costs_fuzzily would do so. The previous coding always ignored
startup cost for parameterized paths, which is wrong as of commit
3f59be836c555fa6; it could result in improper early rejection of paths that
we care about for SEMI/ANTI joins. It also always considered startup cost
for unparameterized paths, which is just as wrong though the only effect is
to waste planner cycles on paths that can't survive. Instead, it should
consider startup cost only when directed to by the consider_startup/
consider_param_startup relation flags.
Likewise, compare_path_costs_fuzzily should have symmetrical behavior
for parameterized and unparameterized paths. In this case, the best
answer seems to be that after establishing that total costs are fuzzily
equal, we should compare startup costs whether or not the consider_xxx
flags are on. That is what it's always done for unparameterized paths,
so let's make the behavior for parameterized paths match.
These issues were noted while developing the SEMI/ANTI join costing fix
of commit 3f59be836c, but we chose not to back-patch these fixes,
because they can cause changes in the planner's choices among
nearly-same-cost plans. (There is in fact one minor change in plan choice
within the core regression tests.) Destabilizing plan choices in back
branches without very clear improvements is frowned on, so we'll just fix
this in HEAD.
When the inner side of a nestloop SEMI or ANTI join is an indexscan that
uses all the join clauses as indexquals, it can be presumed that both
matched and unmatched outer rows will be processed very quickly: for
matched rows, we'll stop after fetching one row from the indexscan, while
for unmatched rows we'll have an indexscan that finds no matching index
entries, which should also be quick. The planner already knew about this,
but it was nonetheless charging for at least one full run of the inner
indexscan, as a consequence of concerns about the behavior of materialized
inner scans --- but those concerns don't apply in the fast case. If the
inner side has low cardinality (many matching rows) this could make an
indexscan plan look far more expensive than it actually is. To fix,
rearrange the work in initial_cost_nestloop/final_cost_nestloop so that we
don't add the inner scan cost until we've inspected the indexquals, and
then we can add either the full-run cost or just the first tuple's cost as
appropriate.
Experimentation with this fix uncovered another problem: add_path and
friends were coded to disregard cheap startup cost when considering
parameterized paths. That's usually okay (and desirable, because it thins
the path herd faster); but in this fast case for SEMI/ANTI joins, it could
result in throwing away the desired plain indexscan path in favor of a
bitmap scan path before we ever get to the join costing logic. In the
many-matching-rows cases of interest here, a bitmap scan will do a lot more
work than required, so this is a problem. To fix, add a per-relation flag
consider_param_startup that works like the existing consider_startup flag,
but applies to parameterized paths, and set it for relations that are the
inside of a SEMI or ANTI join.
To make this patch reasonably safe to back-patch, care has been taken to
avoid changing the planner's behavior except in the very narrow case of
SEMI/ANTI joins with inner indexscans. There are places in
compare_path_costs_fuzzily and add_path_precheck that are not terribly
consistent with the new approach, but changing them will affect planner
decisions at the margins in other cases, so we'll leave that for a
HEAD-only fix.
Back-patch to 9.3; before that, the consider_startup flag didn't exist,
meaning that the second aspect of the patch would be too invasive.
Per a complaint from Peter Holzer and analysis by Tomas Vondra.
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
Add a TABLESAMPLE clause to SELECT statements that allows
user to specify random BERNOULLI sampling or block level
SYSTEM sampling. Implementation allows for extensible
sampling functions to be written, using a standard API.
Basic version follows SQLStandard exactly. Usable
concrete use cases for the sampling API follow in later
commits.
Petr Jelinek
Reviewed by Michael Paquier and Simon Riggs
If we have a semijoin, say
SELECT * FROM x WHERE x1 IN (SELECT y1 FROM y)
and we're estimating the cost of a parameterized indexscan on x, the number
of repetitions of the indexscan should not be taken as the size of y; it'll
really only be the number of distinct values of y1, because the only valid
plan with y on the outside of a nestloop would require y to be unique-ified
before joining it to x. Most of the time this doesn't make that much
difference, but sometimes it can lead to drastically underestimating the
cost of the indexscan and hence choosing a bad plan, as pointed out by
David Kubečka.
Fixing this is a bit difficult because parameterized indexscans are costed
out quite early in the planning process, before we have the information
that would be needed to call estimate_num_groups() and thereby estimate the
number of distinct values of the join column(s). However we can move the
code that extracts a semijoin RHS's unique-ification columns, so that it's
done in initsplan.c rather than on-the-fly in create_unique_path(). That
shouldn't make any difference speed-wise and it's really a bit cleaner too.
The other bit of information we need is the size of the semijoin RHS,
which is easy if it's a single relation (we make those estimates before
considering indexscan costs) but problematic if it's a join relation.
The solution adopted here is just to use the product of the sizes of the
join component rels. That will generally be an overestimate, but since
estimate_num_groups() only uses this input as a clamp, an overestimate
shouldn't hurt us too badly. In any case we don't allow this new logic
to produce a value larger than we would have chosen before, so that at
worst an overestimate leaves us no wiser than we were before.
Instead of register_custom_path_provider and a CreateCustomScanPath
callback, let's just provide a standard function hook in set_rel_pathlist.
This is more flexible than what was previously committed, is more like the
usual conventions for planner hooks, and requires less support code in the
core. We had discussed this design (including centralizing the
set_cheapest() calls) back in March or so, so I'm not sure why it wasn't
done like this already.
Get rid of the pernicious entanglement between planner and executor headers
introduced by commit 0b03e5951b.
Also, rearrange the CustomFoo struct/typedef definitions so that all the
typedef names are seen as used by the compiler. Without this pgindent
will mess things up a bit, which is not so important perhaps, but it also
removes a bizarre discrepancy between the declaration arrangement used for
CustomExecMethods and that used for CustomScanMethods and
CustomPathMethods.
Clean up the commentary around ExecSupportsMarkRestore to reflect the
rather large change in its API.
Const-ify register_custom_path_provider's argument. This necessitates
casting away const in the function, but that seems better than forcing
callers of the function to do so (or else not const-ify their method
pointer structs, which was sort of the whole point).
De-export fix_expr_common. I don't like the exporting of fix_scan_expr
or replace_nestloop_params either, but this one surely has got little
excuse.
This allows extension modules to define their own methods for
scanning a relation, and get the core code to use them. It's
unclear as yet how much use this capability will find, but we
won't find out if we never commit it.
KaiGai Kohei, reviewed at various times and in various levels
of detail by Shigeru Hanada, Tom Lane, Andres Freund, Álvaro
Herrera, and myself.
We can remove a left join to a relation if the relation's output is
provably distinct for the columns involved in the join clause (considering
only equijoin clauses) and the relation supplies no variables needed above
the join. Previously, the join removal logic could only prove distinctness
by reference to unique indexes of a table. This patch extends the logic
to consider subquery relations, wherein distinctness might be proven by
reference to GROUP BY, DISTINCT, etc.
We actually already had some code to check that a subquery's output was
provably distinct, but it was hidden inside pathnode.c; which was a pretty
bad place for it really, since that file is mostly boilerplate Path
construction and comparison. Move that code to analyzejoins.c, which is
arguably a more appropriate location, and is certainly the site of the
new usage for it.
David Rowley, reviewed by Simon Riggs
While the x output of "select x from t group by x" can be presumed unique,
this does not hold for "select x, generate_series(1,10) from t group by x",
because we may expand the set-returning function after the grouping step.
(Perhaps that should be re-thought; but considering all the other oddities
involved with SRFs in targetlists, it seems unlikely we'll change it.)
Put a check in query_is_distinct_for() so it's not fooled by such cases.
Back-patch to all supported branches.
David Rowley
This patch adds the ability to write TABLE( function1(), function2(), ...)
as a single FROM-clause entry. The result is the concatenation of the
first row from each function, followed by the second row from each
function, etc; with NULLs inserted if any function produces fewer rows than
others. This is believed to be a much more useful behavior than what
Postgres currently does with multiple SRFs in a SELECT list.
This syntax also provides a reasonable way to combine use of column
definition lists with WITH ORDINALITY: put the column definition list
inside TABLE(), where it's clear that it doesn't control the ordinality
column as well.
Also implement SQL-compliant multiple-argument UNNEST(), by turning
UNNEST(a,b,c) into TABLE(unnest(a), unnest(b), unnest(c)).
The SQL standard specifies TABLE() with only a single function, not
multiple functions, and it seems to require an implicit UNNEST() which is
not what this patch does. There may be something wrong with that reading
of the spec, though, because if it's right then the spec's TABLE() is just
a pointless alternative spelling of UNNEST(). After further review of
that, we might choose to adopt a different syntax for what this patch does,
but in any case this functionality seems clearly worthwhile.
Andrew Gierth, reviewed by Zoltán Böszörményi and Heikki Linnakangas, and
significantly revised by me
We can detect whether the planner top level is going to care at all about
cheap startup cost (it will only do so if query_planner's tuple_fraction
argument is greater than zero). If it isn't, we might as well discard
paths immediately whose only advantage over others is cheap startup cost.
This turns out to get rid of quite a lot of paths in complex queries ---
I saw planner runtime reduction of more than a third on one large query.
Since add_path isn't currently passed the PlannerInfo "root", the easiest
way to tell it whether to do this was to add a bool flag to RelOptInfo.
That's a bit redundant, since all relations in a given query level will
have the same setting. But in the future it's possible that we'd refine
the control decision to work on a per-relation basis, so this seems like
a good arrangement anyway.
Per my suggestion of a few months ago.
In the initial cut at LATERAL, I kept the rule that cheapest_total_path
was always unparameterized, which meant it had to be NULL if the relation
has no unparameterized paths. It turns out to work much more nicely if
we always have *some* path nominated as cheapest-total for each relation.
In particular, let's still say it's the cheapest unparameterized path if
there is one; if not, take the cheapest-total-cost path among those of
the minimum available parameterization. (The first rule is actually
a special case of the second.)
This allows reversion of some temporary lobotomizations I'd put in place.
In particular, the planner can now consider hash and merge joins for
joins below a parameter-supplying nestloop, even if there aren't any
unparameterized paths available. This should bring planning of
LATERAL-containing queries to the same level as queries not using that
feature.
Along the way, simplify management of parameterized paths in add_path()
and friends. In the original coding for parameterized paths in 9.2,
I tried to minimize the logic changes in add_path(), so it just treated
parameterization as yet another dimension of comparison for paths.
We later made it ignore pathkeys (sort ordering) of parameterized paths,
on the grounds that ordering isn't a useful property for the path on the
inside of a nestloop, so we might as well get rid of useless parameterized
paths as quickly as possible. But we didn't take that reasoning as far as
we should have. Startup cost isn't a useful property inside a nestloop
either, so add_path() ought to discount startup cost of parameterized paths
as well. Having done that, the secondary sorting I'd implemented (in
add_parameterized_path) is no longer needed --- any parameterized path that
survives add_path() at all is worth considering at higher levels. So this
should be a bit faster as well as simpler.
This patch takes care of a number of problems having to do with failure
to choose valid join orders and incorrect handling of lateral references
pulled up from subqueries. Notable changes:
* Add a LateralJoinInfo data structure similar to SpecialJoinInfo, to
represent join ordering constraints created by lateral references.
(I first considered extending the SpecialJoinInfo structure, but the
semantics are different enough that a separate data structure seems
better.) Extend join_is_legal() and related functions to prevent trying
to form unworkable joins, and to ensure that we will consider joins that
satisfy lateral references even if the joins would be clauseless.
* Fill in the infrastructure needed for the last few types of relation scan
paths to support parameterization. We'd have wanted this eventually
anyway, but it is necessary now because a relation that gets pulled up out
of a UNION ALL subquery may acquire a reltargetlist containing lateral
references, meaning that its paths *have* to be parameterized whether or
not we have any code that can push join quals down into the scan.
* Compute data about lateral references early in query_planner(), and save
in RelOptInfo nodes, to avoid repetitive calculations later.
* Assorted corner-case bug fixes.
There's probably still some bugs left, but this is a lot closer to being
real than it was before.
Re-allow subquery pullup for LATERAL subqueries, except when the subquery
is below an outer join and contains lateral references to relations outside
that outer join. If we pull up in such a case, we risk introducing lateral
cross-references into outer joins' ON quals, which is something the code is
entirely unprepared to cope with right now; and I'm not sure it'll ever be
worth coping with.
Support lateral refs in VALUES (this seems to be the only additional path
type that needs such support as a consequence of re-allowing subquery
pullup).
Put in a slightly hacky fix for joinpath.c's refusal to consider
parameterized join paths even when there cannot be any unparameterized
ones. This was causing "could not devise a query plan for the given query"
failures in queries involving more than two FROM items.
Put in an even more hacky fix for distribute_qual_to_rels() being unhappy
with join quals that contain references to rels outside their syntactic
scope; which is to say, disable that test altogether. Need to think about
how to preserve some sort of debugging cross-check here, while not
expending more cycles than befits a debugging cross-check.
This patch implements the standard syntax of LATERAL attached to a
sub-SELECT in FROM, and also allows LATERAL attached to a function in FROM,
since set-returning function calls are expected to be one of the principal
use-cases.
The main change here is a rewrite of the mechanism for keeping track of
which relations are visible for column references while the FROM clause is
being scanned. The parser "namespace" lists are no longer lists of bare
RTEs, but are lists of ParseNamespaceItem structs, which carry an RTE
pointer as well as some visibility-controlling flags. Aside from
supporting LATERAL correctly, this lets us get rid of the ancient hacks
that required rechecking subqueries and JOIN/ON and function-in-FROM
expressions for invalid references after they were initially parsed.
Invalid column references are now always correctly detected on sight.
In passing, remove assorted parser error checks that are now dead code by
virtue of our having gotten rid of add_missing_from, as well as some
comments that are obsolete for the same reason. (It was mainly
add_missing_from that caused so much fudging here in the first place.)
The planner support for this feature is very minimal, and will be improved
in future patches. It works well enough for testing purposes, though.
catversion bump forced due to new field in RangeTblEntry.
Instead of an exact cost comparison, use a fuzzy comparison with 1e-10
delta after all other path metrics have proved equal. This is to avoid
having platform-specific roundoff behaviors determine the choice when
two paths are really the same to our cost estimators. Adjust the
recently-added test case that made it obvious we had a problem here.
This patch adjusts the treatment of parameterized paths so that all paths
with the same parameterization (same set of required outer rels) for the
same relation will have the same rowcount estimate. We cache the rowcount
estimates to ensure that property, and hopefully save a few cycles too.
Doing this makes it practical for add_path_precheck to operate without
a rowcount estimate: it need only assume that paths with different
parameterizations never dominate each other, which is close enough to
true anyway for coarse filtering, because normally a more-parameterized
path should yield fewer rows thanks to having more join clauses to apply.
In add_path, we do the full nine yards of comparing rowcount estimates
along with everything else, so that we can discard parameterized paths that
don't actually have an advantage. This fixes some issues I'd found with
add_path rejecting parameterized paths on the grounds that they were more
expensive than not-parameterized ones, even though they yielded many fewer
rows and hence would be cheaper once subsequent joining was considered.
To make the same-rowcounts assumption valid, we have to require that any
parameterized path enforce *all* join clauses that could be obtained from
the particular set of outer rels, even if not all of them are useful for
indexing. This is required at both base scans and joins. It's a good
thing anyway since the net impact is that join quals are checked at the
lowest practical level in the join tree. Hence, discard the original
rather ad-hoc mechanism for choosing parameterization joinquals, and build
a better one that has a more principled rule for when clauses can be moved.
The original rule was actually buggy anyway for lack of knowledge about
which relations are part of an outer join's outer side; getting this right
requires adding an outer_relids field to RestrictInfo.
Further reflection shows that a single callback isn't very workable if we
desire to let FDWs generate multiple Paths, because that forces the FDW to
do all work necessary to generate a valid Plan node for each Path. Instead
split the former PlanForeignScan API into three steps: GetForeignRelSize,
GetForeignPaths, GetForeignPlan. We had already bit the bullet of breaking
the 9.1 FDW API for 9.2, so this shouldn't cause very much additional pain,
and it's substantially more flexible for complex FDWs.
Add an fdw_private field to RelOptInfo so that the new functions can save
state there rather than possibly having to recalculate information two or
three times.
In addition, we'd not thought through what would be needed to allow an FDW
to set up subexpressions of its choice for runtime execution. We could
treat ForeignScan.fdw_private as an executable expression but that seems
likely to break existing FDWs unnecessarily (in particular, it would
restrict the set of node types allowable in fdw_private to those supported
by expression_tree_walker). Instead, invent a separate field fdw_exprs
which will receive the postprocessing appropriate for expression trees.
(One field is enough since it can be a list of expressions; also, we assume
the corresponding expression state tree(s) will be held within fdw_state,
so we don't need to add anything to ForeignScanState.)
Per review of Hanada Shigeru's pgsql_fdw patch. We may need to tweak this
further as we continue to work on that patch, but to me it feels a lot
closer to being right now.
The original API specification only allowed an FDW to create a single
access path, which doesn't seem like a terribly good idea in hindsight.
Instead, move the responsibility for building the Path node and calling
add_path() into the FDW's PlanForeignScan function. Now, it can do that
more than once if appropriate. There is no longer any need for the
transient FdwPlan struct, so get rid of that.
Etsuro Fujita, Shigeru Hanada, Tom Lane
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.
If the right-hand side of a semijoin is unique, then we can treat it like a
normal join (or another way to say that is: we don't need to explicitly
unique-ify the data before doing it as a normal join). We were recognizing
such cases when the RHS was a sub-query with appropriate DISTINCT or GROUP
BY decoration, but there's another way: if the RHS is a plain relation with
unique indexes, we can check if any of the indexes prove the output is
unique. Most of the infrastructure for that was there already in the join
removal code, though I had to rearrange it a bit. Per reflection about a
recent example in pgsql-performance.
This commit changes index-only scans so that data is read directly from the
index tuple without first generating a faux heap tuple. The only immediate
benefit is that indexes on system columns (such as OID) can be used in
index-only scans, but this is necessary infrastructure if we are ever to
support index-only scans on expression indexes. The executor is now ready
for that, though the planner still needs substantial work to recognize
the possibility.
To do this, Vars in index-only plan nodes have to refer to index columns
not heap columns. I introduced a new special varno, INDEX_VAR, to mark
such Vars to avoid confusion. (In passing, this commit renames the two
existing special varnos to OUTER_VAR and INNER_VAR.) This allows
ruleutils.c to handle them with logic similar to what we use for subplan
reference Vars.
Since index-only scans are now fundamentally different from regular
indexscans so far as their expression subtrees are concerned, I also chose
to change them to have their own plan node type (and hence, their own
executor source file).
When a btree index contains all columns required by the query, and the
visibility map shows that all tuples on a target heap page are
visible-to-all, we don't need to fetch that heap page. This patch depends
on the previous patches that made the visibility map reliable.
There's a fair amount left to do here, notably trying to figure out a less
chintzy way of estimating the cost of an index-only scan, but the core
functionality seems ready to commit.
Robert Haas and Ibrar Ahmed, with some previous work by Heikki Linnakangas.
The previous coding failed to account properly for the costs of evaluating
the input expressions of aggregates and window functions, as seen in a
recent gripe from Claudio Freire. (I said at the time that it wasn't
counting these costs at all; but on closer inspection, it was effectively
charging these costs once per output tuple. That is completely wrong for
aggregates, and not exactly right for window functions either.)
There was also a hard-wired assumption that aggregates and window functions
had procost 1.0, which is now fixed to respect the actual cataloged costs.
The costing of WindowAgg is still pretty bogus, since it doesn't try to
estimate the effects of spilling data to disk, but that seems like a
separate issue.
This commit provides the core code and documentation needed. A contrib
module test case will follow shortly.
Shigeru Hanada, Jan Urbanski, Heikki Linnakangas
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
Fix things so that top-N sorting can be used in child Sort nodes of a
MergeAppend node, when there is a LIMIT and no intervening joins or
grouping. Actually doing this on the executor side isn't too bad,
but it's a bit messier to get the planner to cost it properly.
Per gripe from Robert Haas.
In passing, fix an oversight in the original top-N-sorting patch:
query_planner should not assume that a LIMIT can be used to make an
explicit sort cheaper when there will be grouping or aggregation in
between. Possibly this should be back-patched, but I'm not sure the
mistake is serious enough to be a real problem in practice.
The core of this patch is hash_array() and associated typcache
infrastructure, which works just about exactly like the existing support
for array comparison.
In addition I did some work to ensure that the planner won't think that an
array type is hashable unless its element type is hashable, and similarly
for sorting. This includes adding a datatype parameter to op_hashjoinable
and op_mergejoinable, and adding an explicit "hashable" flag to
SortGroupClause. The lack of a cross-check on the element type was a
pre-existing bug in mergejoin support --- but it didn't matter so much
before, because if you couldn't sort the element type there wasn't any good
alternative to failing anyhow. Now that we have the alternative of hashing
the array type, there are cases where we can avoid a failure by being picky
at the planner stage, so it's time to be picky.
The issue of exactly how to combine the per-element hash values to produce
an array hash is still open for discussion, but the rest of this is pretty
solid, so I'll commit it as-is.
This patch eliminates the former need to sort the output of an Append scan
when an ordered scan of an inheritance tree is wanted. This should be
particularly useful for fast-start cases such as queries with LIMIT.
Original patch by Greg Stark, with further hacking by Hans-Jurgen Schonig,
Robert Haas, and Tom Lane.
mergejoin to shield it from doing mark/restore and refetches. Put an explicit
flag in MergePath so we can centralize the logic that knows about this,
and add costing logic that considers using Materialize even when it's not
forced by the previously-existing considerations. This is in response to
a discussion back in August that suggested that materializing an inner
indexscan can be helpful when the refetch percentage is high enough.
is unique and is not referenced above the join. In this case the inner
side doesn't affect the query result and can be thrown away entirely.
Although perhaps nobody would ever write such a thing by hand, it's
a reasonably common case in machine-generated SQL.
The current implementation only recognizes the case where the inner side
is a simple relation with a unique index matching the query conditions.
This is enough for the use-cases that have been shown so far, but we
might want to try to handle other cases later.
Robert Haas, somewhat rewritten by Tom
an explicit model of rescan costs being different from first-time costs.
The costing of Material nodes in particular now has some visible relationship
to the actual runtime behavior, where before it was essentially fantasy.
This also fixes up a couple of places where different materialized plan types
were treated differently for no very good reason (probably just oversights).
A couple of the regression tests are affected, because the planner now chooses
to put the other relation on the inside of a nestloop-with-materialize.
So far as I can see both changes are sane, and the planner is now more
consistently following the expectation that it should prefer to materialize
the smaller of two relations.
Per a recent discussion with Robert Haas.
"physical tlist" optimization on the outer relation (ie, force a projection
step to occur in its scan). This avoids storing useless column values when
the outer relation's tuples are written to temporary batch files.
Modified version of a patch by Michael Henderson and Ramon Lawrence.
to be syntactically part of a semijoin clause. For example given
WHERE EXISTS(SELECT ... WHERE upper.var = lower.var AND some-condition)
where some-condition is just a restriction on the lower relation, we can
use unique-ification on lower.var after having applied some-condition within
the scan on lower.
There are some unimplemented aspects: recursive queries must use UNION ALL
(should allow UNION too), and we don't have SEARCH or CYCLE clauses.
These might or might not get done for 8.4, but even without them it's a
pretty useful feature.
There are also a couple of small loose ends and definitional quibbles,
which I'll send a memo about to pgsql-hackers shortly. But let's land
the patch now so we can get on with other development.
Yoshiyuki Asaba, with lots of help from Tatsuo Ishii and Tom Lane
inserting a materialize node above an inner-side sort node, when the sort is
expected to spill to disk. (The materialize protects the sort from having
to support mark/restore, allowing it to do its final merge pass on-the-fly.)
We neglected to teach cost_mergejoin about that hack, so it was failing to
include the materialize's costs in the estimated cost of the mergejoin.
The materialize's costs are generally going to be pretty negligible in
comparison to the sort's, so this is only a small error and probably not
worth back-patching; but it's still wrong.
In the similar case where a materialize is inserted to protect an inner-side
node that can't do mark/restore at all, it's still true that the materialize
should not spill to disk, and so we should cost it cheaply rather than
expensively.
Noted while thinking about a question from Tom Raney.
the old JOIN_IN code, but antijoins are new functionality.) Teach the planner
to convert appropriate EXISTS and NOT EXISTS subqueries into semi and anti
joins respectively. Also, LEFT JOINs with suitable upper-level IS NULL
filters are recognized as being anti joins. Unify the InClauseInfo and
OuterJoinInfo infrastructure into "SpecialJoinInfo". With that change,
it becomes possible to associate a SpecialJoinInfo with every join attempt,
which permits some cleanup of join selectivity estimation. That needs to be
taken much further than this patch does, but the next step is to change the
API for oprjoin selectivity functions, which seems like material for a
separate patch. So for the moment the output size estimates for semi and
especially anti joins are quite bogus.
but seem like a separate patch since most of the remaining work is on the
executor side.) I took the opportunity to push selection of the grouping
operators for set operations into the parser where it belongs. Otherwise this
is just a small exercise in making prepunion.c consider both alternatives.
As with the recent DISTINCT patch, this means we can UNION on datatypes that
can hash but not sort, and it means that UNION without ORDER BY is no longer
certain to produce sorted output.
as per my recent proposal:
1. Fold SortClause and GroupClause into a single node type SortGroupClause.
We were already relying on them to be struct-equivalent, so using two node
tags wasn't accomplishing much except to get in the way of comparing items
with equal().
2. Add an "eqop" field to SortGroupClause to carry the associated equality
operator. This is cheap for the parser to get at the same time it's looking
up the sort operator, and storing it eliminates the need for repeated
not-so-cheap lookups during planning. In future this will also let us
represent GROUP/DISTINCT operations on datatypes that have hash opclasses
but no btree opclasses (ie, they have equality but no natural sort order).
The previous representation simply didn't work for that, since its only
indicator of comparison semantics was a sort operator.
3. Add a hasDistinctOn boolean to struct Query to explicitly record whether
the distinctClause came from DISTINCT or DISTINCT ON. This allows removing
some complicated and not 100% bulletproof code that attempted to figure
that out from the distinctClause alone.
This patch doesn't in itself create any new capability, but it's necessary
infrastructure for future attempts to use hash-based grouping for DISTINCT
and UNION/INTERSECT/EXCEPT.
output is not of the same type that's needed for the IN comparison (ie,
where the parser inserted an implicit coercion above the subselect result).
We should record the coerced expression, not just a raw Var referencing
the subselect output, as the quantity that needs to be unique-ified if
we choose to implement the IN as Unique followed by a plain join.
As of 8.3 this error was causing crashes, as seen in bug #4113 from Javier
Hernandez, because the executor was being told to hash or sort the raw
subselect output column using operators appropriate to the coerced type.
In prior versions there was no crash because the executor chose the
hash or sort operators for itself based on the column type it saw.
However, that's still not really right, because what's unique for one data
type might not be unique for another. In corner cases we could get multiple
outputs of a row that should appear only once, as demonstrated by the
regression test case included in this commit.
However, this patch doesn't apply cleanly to 8.2 or before, and the code
involved has shifted enough over time that I'm hesitant to try to back-patch.
Given the lack of complaints from the field about such corner cases, I think
the bug may not be important enough to risk breaking other things with a
back-patch.
need be returned. We keep a heap of the current best N tuples and sift-up
new tuples into it as we scan the input. For M input tuples this means
only about M*log(N) comparisons instead of M*log(M), not to mention a lot
less workspace when N is small --- avoiding spill-to-disk for large M
is actually the most attractive thing about it. Patch includes planner
and executor support for invoking this facility in ORDER BY ... LIMIT
queries. Greg Stark, with some editorialization by moi.
are mostly excluded by constraints: do the CE test a bit earlier to save
some adjust_appendrel_attrs() work on excluded children, and arrange to
use array indexing rather than rt_fetch() to fetch RTEs in the main body
of the planner. The latter is something I'd wanted to do for awhile anyway,
but seeing list_nth_cell() as 35% of the runtime gets one's attention.
that aren't turned into true joins). Since this is the last missing bit of
infrastructure, go ahead and fill out the hash integer_ops and float_ops
opfamilies with cross-type operators. The operator family project is now
DONE ... er, except for documentation ...
representation of equivalence classes of variables. This is an extensive
rewrite, but it brings a number of benefits:
* planner no longer fails in the presence of "incomplete" operator families
that don't offer operators for every possible combination of datatypes.
* avoid generating and then discarding redundant equality clauses.
* remove bogus assumption that derived equalities always use operators
named "=".
* mergejoins can work with a variety of sort orders (e.g., descending) now,
instead of tying each mergejoinable operator to exactly one sort order.
* better recognition of redundant sort columns.
* can make use of equalities appearing underneath an outer join.
which comparison operators to use for plan nodes involving tuple comparison
(Agg, Group, Unique, SetOp). Formerly the executor looked up the default
equality operator for the datatype, which was really pretty shaky, since it's
possible that the data being fed to the node is sorted according to some
nondefault operator class that could have an incompatible idea of equality.
The planner knows what it has sorted by and therefore can provide the right
equality operator to use. Also, this change moves a couple of catalog lookups
out of the executor and into the planner, which should help startup time for
pre-planned queries by some small amount. Modify the planner to remove some
other cavalier assumptions about always being able to use the default
operators. Also add "nulls first/last" info to the Plan node for a mergejoin
--- neither the executor nor the planner can cope yet, but at least the API is
in place.
cases. Operator classes now exist within "operator families". While most
families are equivalent to a single class, related classes can be grouped
into one family to represent the fact that they are semantically compatible.
Cross-type operators are now naturally adjunct parts of a family, without
having to wedge them into a particular opclass as we had done originally.
This commit restructures the catalogs and cleans up enough of the fallout so
that everything still works at least as well as before, but most of the work
needed to actually improve the planner's behavior will come later. Also,
there are not yet CREATE/DROP/ALTER OPERATOR FAMILY commands; the only way
to create a new family right now is to allow CREATE OPERATOR CLASS to make
one by default. I owe some more documentation work, too. But that can all
be done in smaller pieces once this infrastructure is in place.
(e.g. "INSERT ... VALUES (...), (...), ...") and elsewhere as allowed
by the spec. (e.g. similar to a FROM clause subselect). initdb required.
Joe Conway and Tom Lane.
effects in a nestloop inner indexscan, I had only dealt with plain index
scans and the index portion of bitmap scans. But there will be cache
benefits for the heap accesses of bitmap scans too, so fix
cost_bitmap_heap_scan() to account for that.
clauses containing no variables and no volatile functions. Such a clause
can be used as a one-time qual in a gating Result plan node, to suppress
plan execution entirely when it is false. Even when the clause is true,
putting it in a gating node wins by avoiding repeated evaluation of the
clause. In previous PG releases, query_planner() would do this for
pseudoconstant clauses appearing at the top level of the jointree, but
there was no ability to generate a gating Result deeper in the plan tree.
To fix it, get rid of the special case in query_planner(), and instead
process pseudoconstant clauses through the normal RestrictInfo qual
distribution mechanism. When a pseudoconstant clause is found attached to
a path node in create_plan(), pull it out and generate a gating Result at
that point. This requires special-casing pseudoconstants in selectivity
estimation and cost_qual_eval, but on the whole it's pretty clean.
It probably even makes the planner a bit faster than before for the normal
case of no pseudoconstants, since removing pull_constant_clauses saves one
useless traversal of the qual tree. Per gripe from Phil Frost.
that the Mackert-Lohmann formula applies across all the repetitions of the
nestloop, not just each scan independently. We use the M-L formula to
estimate the number of pages fetched from the index as well as from the table;
that isn't what it was designed for, but it seems reasonably applicable
anyway. This makes large numbers of repetitions look much cheaper than
before, which accords with many reports we've received of overestimation
of the cost of a nestloop. Also, change the index access cost model to
charge random_page_cost per index leaf page touched, while explicitly
not counting anything for access to metapage or upper tree pages. This
may all need tweaking after we get some field experience, but in simple
tests it seems to be giving saner results than before. The main thing
is to get the infrastructure in place to let cost_index() and amcostestimate
functions take repeated scans into account at all. Per my recent proposal.
Note: this patch changes pg_proc.h, but I did not force initdb because
the changes are basically cosmetic --- the system does not look into
pg_proc to decide how to call an index amcostestimate function, and
there's no way to call such a function from SQL at all.
"ctid IN (list)" will still work after we convert IN to ScalarArrayOpExpr.
Make some minor efficiency improvements while at it, such as ensuring that
multiple TIDs are fetched in physical heap order. And fix EXPLAIN so that
it shows what's really going on for a TID scan.
coding would ignore startup cost differences of less than 1% of the
estimated total cost; which was OK for normal planning but highly not OK
if a very small LIMIT was applied afterwards, so that startup cost becomes
the name of the game. Instead, compare startup and total costs fuzzily
but independently. This changes the plan selected for two queries in the
regression tests; adjust expected-output files for resulting changes in
row order. Per reports from Dawid Kuroczko and Sam Mason.
cases: we can't just consider whether the subquery's output is unique on its
own terms, we have to check whether the set of output columns we are going to
use will be unique. Per complaint from Luca Pireddu and test case from
Michael Fuhr.
a new PlannerInfo struct, which is passed around instead of the bare
Query in all the planning code. This commit is essentially just a
code-beautification exercise, but it does open the door to making
larger changes to the planner data structures without having to muck
with the widely-known Query struct.
node, as this behavior is now better done as a bitmap OR indexscan.
This allows considerable simplification in nodeIndexscan.c itself as
well as several planner modules concerned with indexscan plan generation.
Also we can improve the sharing of code between regular and bitmap
indexscans, since they are now working with nigh-identical Plan nodes.
but the code is basically working. Along the way, rewrite the entire
approach to processing OR index conditions, and make it work in join
cases for the first time ever. orindxpath.c is now basically obsolete,
but I left it in for the time being to allow easy comparison testing
against the old implementation.
logic operations during planning. Seems cleaner to create two new Path
node types, instead --- this avoids duplication of cost-estimation code.
Also, create an enable_bitmapscan GUC parameter to control use of bitmap
plans.
scans, using in-memory tuple ID bitmaps as the intermediary. The planner
frontend (path creation and cost estimation) is not there yet, so none
of this code can be executed. I have tested it using some hacked planner
code that is far too ugly to see the light of day, however. Committing
now so that the bulk of the infrastructure changes go in before the tree
drifts under me.
few palloc's. I also chose to eliminate the restype and restypmod fields
entirely, since they are redundant with information stored in the node's
contained expression; re-examining the expression at need seems simpler
and more reliable than trying to keep restype/restypmod up to date.
initdb forced due to change in contents of stored rules.
structs. There are many places in the planner where we were passing
both a rel and an index to subroutines, and now need only pass the
index struct. Notationally simpler, and perhaps a tad faster.
for boolean indexes. Previously we would only use such an index with
WHERE clauses like 'indexkey = true' or 'indexkey = false'. The new
code transforms the cases 'indexkey', 'NOT indexkey', 'indexkey IS TRUE',
and 'indexkey IS FALSE' into one of these. While this is only marginally
useful in itself, I intend soon to change constant-expression simplification
so that 'foo = true' and 'foo = false' are reduced to just 'foo' and
'NOT foo' ... which would lose the ability to use boolean indexes for
such queries at all, if the indexscan machinery couldn't make the
reverse transformation.
Formerly, if such a clause contained no aggregate functions we mistakenly
treated it as equivalent to WHERE. Per spec it must cause the query to
be treated as a grouped query of a single group, the same as appearance
of aggregate functions would do. Also, the HAVING filter must execute
after aggregate function computation even if it itself contains no
aggregate functions.
Also performed an initial run through of upgrading our Copyright date to
extend to 2005 ... first run here was very simple ... change everything
where: grep 1996-2004 && the word 'Copyright' ... scanned through the
generated list with 'less' first, and after, to make sure that I only
picked up the right entries ...
In the past, we used a 'Lispy' linked list implementation: a "list" was
merely a pointer to the head node of the list. The problem with that
design is that it makes lappend() and length() linear time. This patch
fixes that problem (and others) by maintaining a count of the list
length and a pointer to the tail node along with each head node pointer.
A "list" is now a pointer to a structure containing some meta-data
about the list; the head and tail pointers in that structure refer
to ListCell structures that maintain the actual linked list of nodes.
The function names of the list API have also been changed to, I hope,
be more logically consistent. By default, the old function names are
still available; they will be disabled-by-default once the rest of
the tree has been updated to use the new API names.
same path keys and nearly equivalent costs will be considered redundant.
The exact nature of the fuzziness may get adjusted later based on current
discussions, but no one has shot a hole in the basic idea yet ...
Make btree index creation and initial validation of foreign-key constraints
use maintenance_work_mem rather than work_mem as their memory limit.
Add some code to guc.c to allow these variables to be referenced by their
old names in SHOW and SET commands, for backwards compatibility.
with index qual clauses in the Path representation. This saves a little
work during createplan and (probably more importantly) allows reuse of
cached selectivity estimates during indexscan planning. Also fix latent
bug: wrong plan would have been generated for a 'special operator' used
in a nestloop-inner-indexscan join qual, because the special operator
would not have gotten into the list of quals to recheck. This bug is
only latent because at present the special-operator code could never
trigger on a join qual, but sooner or later someone will want to do it.
join conditions in which each OR subclause includes a constraint on
the same relation. This implements the other useful side-effect of
conversion to CNF format, without its unpleasant side-effects. As
per pghackers discussion of a few weeks ago.
node emits only those vars that are actually needed above it in the
plan tree. (There were comments in the code suggesting that this was
done at some point in the dim past, but for a long time we have just
made join nodes emit everything that either input emitted.) Aside from
being marginally more efficient, this fixes the problem noted by Peter
Eisentraut where a join above an IN-implemented-as-join might fail,
because the subplan targetlist constructed in the latter case didn't
meet the expectation of including everything.
Along the way, fix some places that were O(N^2) in the targetlist
length. This is not all the trouble spots for wide queries by any
means, but it's a step forward.
some cases of redundant clauses that were formerly not caught. We have
to special-case this because the clauses involved never get attached to
the same join restrictlist and so the existing logic does not notice
that they are redundant.
blanks, in hopes of reducing the surprise factor for newbies. Remove
redundant operators for VARCHAR (it depends wholly on TEXT operations now).
Clean up resolution of ambiguous operators/functions to avoid surprising
choices for domains: domains are treated as equivalent to their base types
and binary-coercibility is no longer considered a preference item when
choosing among multiple operators/functions. IsBinaryCoercible now correctly
reflects the notion that you need *only* relabel the type to get from type
A to type B: that is, a domain is binary-coercible to its base type, but
not vice versa. Various marginal cleanup, including merging the essentially
duplicate resolution code in parse_func.c and parse_oper.c. Improve opr_sanity
regression test to understand about binary compatibility (using pg_cast),
and fix a couple of small errors in the catalogs revealed thereby.
Restructure "special operator" handling to fetch operators via index opclasses
rather than hardwiring assumptions about names (cleans up the pattern_ops
stuff a little).
the outer query. (The implementation is a bit klugy, but it would take
nontrivial restructuring to make it nicer, which this is probably not
worth.) This avoids unnecessary sort steps in examples like
SELECT foo,count(*) FROM (SELECT ... ORDER BY foo,bar) sub GROUP BY foo
which means there is now a reasonable technique for controlling the
order of inputs to custom aggregates, even in the grouping case.
Try to model the effect of rescanning input tuples in mergejoins;
account for JOIN_IN short-circuiting where appropriate. Also, recognize
that mergejoin and hashjoin clauses may now be more than single operator
calls, so we have to charge appropriate execution costs.
There are two implementation techniques: the executor understands a new
JOIN_IN jointype, which emits at most one matching row per left-hand row,
or the result of the IN's sub-select can be fed through a DISTINCT filter
and then joined as an ordinary relation.
Along the way, some minor code cleanup in the optimizer; notably, break
out most of the jointree-rearrangement preprocessing in planner.c and
put it in a new file prep/prepjointree.c.
to plan nodes, not vice-versa. All executor state nodes now inherit from
struct PlanState. Copying of plan trees has been simplified by not
storing a list of SubPlans in Plan nodes (eliminating duplicate links).
The executor still needs such a list, but it can build it during
ExecutorStart since it has to scan the plan tree anyway.
No initdb forced since no stored-on-disk structures changed, but you
will need a full recompile because of node-numbering changes.
Andres Freund
Bruce Momjian
Tom Lane
Robert Haas
Simon Riggs
Magnus Hagander
Joe Conway
PostgreSQL Daemon
Neil Conway