Tomas Vondra observed that the IO behavior for HashAgg tends to be
worse than for Sort. Penalize HashAgg IO costs accordingly.
Also, account for the CPU effort of spilling the tuples and reading
them back.
Discussion: https://postgr.es/m/20200906212112.nzoy5ytrzjjodpfh@development
Reviewed-by: Tomas Vondra
Backpatch-through: 13
Add a GUC that acts as a multiplier on work_mem. It gets applied when
sizing executor node hash tables that were previously size constrained
using work_mem alone.
The new GUC can be used to preferentially give hash-based nodes more
memory than the generic work_mem limit. It is intended to enable admin
tuning of the executor's memory usage. Overall system throughput and
system responsiveness can be improved by giving hash-based executor
nodes more memory (especially over sort-based alternatives, which are
often much less sensitive to being memory constrained).
The default value for hash_mem_multiplier is 1.0, which is also the
minimum valid value. This means that hash-based nodes continue to apply
work_mem in the traditional way by default.
hash_mem_multiplier is generally useful. However, it is being added now
due to concerns about hash aggregate performance stability for users
that upgrade to Postgres 13 (which added disk-based hash aggregation in
commit 1f39bce0). While the old hash aggregate behavior risked
out-of-memory errors, it is nevertheless likely that many users actually
benefited. Hash agg's previous indifference to work_mem during query
execution was not just faster; it also accidentally made aggregation
resilient to grouping estimate problems (at least in cases where this
didn't create destabilizing memory pressure).
hash_mem_multiplier can provide a certain kind of continuity with the
behavior of Postgres 12 hash aggregates in cases where the planner
incorrectly estimates that all groups (plus related allocations) will
fit in work_mem/hash_mem. This seems necessary because hash-based
aggregation is usually much slower when only a small fraction of all
groups can fit. Even when it isn't possible to totally avoid hash
aggregates that spill, giving hash aggregation more memory will reliably
improve performance (the same cannot be said for external sort
operations, which appear to be almost unaffected by memory availability
provided it's at least possible to get a single merge pass).
The PostgreSQL 13 release notes should advise users that increasing
hash_mem_multiplier can help with performance regressions associated
with hash aggregation. That can be taken care of by a later commit.
Author: Peter Geoghegan
Reviewed-By: Álvaro Herrera, Jeff Davis
Discussion: https://postgr.es/m/20200625203629.7m6yvut7eqblgmfo@alap3.anarazel.de
Discussion: https://postgr.es/m/CAH2-WzmD%2Bi1pG6rc1%2BCjc4V6EaFJ_qSuKCCHVnH%3DoruqD-zqow%40mail.gmail.com
Backpatch: 13-, where disk-based hash aggregation was introduced.
Note: This GUC was originally named enable_hashagg_disk when it appeared
in commit 1f39bce0, which added disk-based hash aggregation. It was
subsequently renamed in commit 92c58fd9.
Author: Peter Geoghegan
Reviewed-By: Jeff Davis, Álvaro Herrera
Discussion: https://postgr.es/m/9d9d1e1252a52ea1bad84ea40dbebfd54e672a0f.camel%40j-davis.com
Backpatch: 13-, where disk-based hash aggregation was introduced.
Eliminate enable_groupingsets_hash_disk, which was primarily useful
for testing grouping sets that use HashAgg and spill. Instead, hack
the table stats to convince the planner to choose hashed aggregation
for grouping sets that will spill to disk. Suggested by Melanie
Plageman.
Rename enable_hashagg_disk to hashagg_avoid_disk_plan, and invert the
meaning of on/off. The new name indicates more strongly that it only
affects the planner. Also, the word "avoid" is less definite, which
should avoid surprises when HashAgg still needs to use the
disk. Change suggested by Justin Pryzby, though I chose a different
GUC name.
Discussion: https://postgr.es/m/CAAKRu_aisiENMsPM2gC4oUY1hHG3yrCwY-fXUg22C6_MJUwQdA%40mail.gmail.com
Discussion: https://postgr.es/m/20200610021544.GA14879@telsasoft.com
Backpatch-through: 13
Thomas Munro fixed a longstanding annoyance in pg_bsd_indent, that
it would misformat lines containing IsA() macros on the assumption
that the IsA() call should be treated like a cast. This improves
some other cases involving field/variable names that match typedefs,
too. The only places that get worse are a couple of uses of the
OpenSSL macro STACK_OF(); we'll gladly take that trade-off.
Discussion: https://postgr.es/m/20200114221814.GA19630@alvherre.pgsql
Includes some manual cleanup of places that pgindent messed up,
most of which weren't per project style anyway.
Notably, it seems some people didn't absorb the style rules of
commit c9d297751, because there were a bunch of new occurrences
of function calls with a newline just after the left paren, all
with faulty expectations about how the rest of the call would get
indented.
When estimating the number of pre-sorted groups in cost_incremental_sort
we must not pass Vars with varno 0 to estimate_num_groups, which would
cause failues in find_base_rel. This may happen when sorting output of
set operations, thanks to generate_append_tlist.
Unlike recurse_set_operations we can't easily access the original target
list, so if we find any Vars with varno 0, we fall back to the default
estimate DEFAULT_NUM_DISTINCT.
Reported-by: Justin Pryzby
Discussion: https://postgr.es/m/20200411214639.GK2228%40telsasoft.com
Incremental Sort is an optimized variant of multikey sort for cases when
the input is already sorted by a prefix of the requested sort keys. For
example when the relation is already sorted by (key1, key2) and we need
to sort it by (key1, key2, key3) we can simply split the input rows into
groups having equal values in (key1, key2), and only sort/compare the
remaining column key3.
This has a number of benefits:
- Reduced memory consumption, because only a single group (determined by
values in the sorted prefix) needs to be kept in memory. This may also
eliminate the need to spill to disk.
- Lower startup cost, because Incremental Sort produce results after each
prefix group, which is beneficial for plans where startup cost matters
(like for example queries with LIMIT clause).
We consider both Sort and Incremental Sort, and decide based on costing.
The implemented algorithm operates in two different modes:
- Fetching a minimum number of tuples without check of equality on the
prefix keys, and sorting on all columns when safe.
- Fetching all tuples for a single prefix group and then sorting by
comparing only the remaining (non-prefix) keys.
We always start in the first mode, and employ a heuristic to switch into
the second mode if we believe it's beneficial - the goal is to minimize
the number of unnecessary comparions while keeping memory consumption
below work_mem.
This is a very old patch series. The idea was originally proposed by
Alexander Korotkov back in 2013, and then revived in 2017. In 2018 the
patch was taken over by James Coleman, who wrote and rewrote most of the
current code.
There were many reviewers/contributors since 2013 - I've done my best to
pick the most active ones, and listed them in this commit message.
Author: James Coleman, Alexander Korotkov
Reviewed-by: Tomas Vondra, Andreas Karlsson, Marti Raudsepp, Peter Geoghegan, Robert Haas, Thomas Munro, Antonin Houska, Andres Freund, Alexander Kuzmenkov
Discussion: https://postgr.es/m/CAPpHfdscOX5an71nHd8WSUH6GNOCf=V7wgDaTXdDd9=goN-gfA@mail.gmail.com
Discussion: https://postgr.es/m/CAPpHfds1waRZ=NOmueYq0sx1ZSCnt+5QJvizT8ndT2=etZEeAQ@mail.gmail.com
While performing hash aggregation, track memory usage when adding new
groups to a hash table. If the memory usage exceeds work_mem, enter
"spill mode".
In spill mode, new groups are not created in the hash table(s), but
existing groups continue to be advanced if input tuples match. Tuples
that would cause a new group to be created are instead spilled to a
logical tape to be processed later.
The tuples are spilled in a partitioned fashion. When all tuples from
the outer plan are processed (either by advancing the group or
spilling the tuple), finalize and emit the groups from the hash
table. Then, create new batches of work from the spilled partitions,
and select one of the saved batches and process it (possibly spilling
recursively).
Author: Jeff Davis
Reviewed-by: Tomas Vondra, Adam Lee, Justin Pryzby, Taylor Vesely, Melanie Plageman
Discussion: https://postgr.es/m/507ac540ec7c20136364b5272acbcd4574aa76ef.camel@j-davis.com
This follows multiple complains from Peter Geoghegan, Andres Freund and
Alvaro Herrera that this issue ought to be dug more before actually
happening, if it happens.
Discussion: https://postgr.es/m/20191226144606.GA5659@alvherre.pgsql
The following renaming is done so as source files related to index
access methods are more consistent with table access methods (the
original names used for index AMs ware too generic, and could be
confused as including features related to table AMs):
- amapi.h -> indexam.h.
- amapi.c -> indexamapi.c. Here we have an equivalent with
backend/access/table/tableamapi.c.
- amvalidate.c -> indexamvalidate.c.
- amvalidate.h -> indexamvalidate.h.
- genam.c -> indexgenam.c.
- genam.h -> indexgenam.h.
This has been discussed during the development of v12 when table AM was
worked on, but the renaming never happened.
Author: Michael Paquier
Reviewed-by: Fabien Coelho, Julien Rouhaud
Discussion: https://postgr.es/m/20191223053434.GF34339@paquier.xyz
In the wake of commit 1cff1b95a, the result of list_concat no longer
shares the ListCells of the second input. Therefore, we can replace
"list_concat(x, list_copy(y))" with just "list_concat(x, y)".
To improve call sites that were list_copy'ing the first argument,
or both arguments, invent "list_concat_copy()" which produces a new
list sharing no ListCells with either input. (This is a bit faster
than "list_concat(list_copy(x), y)" because it makes the result list
the right size to start with.)
In call sites that were not list_copy'ing the second argument, the new
semantics mean that we are usually leaking the second List's storage,
since typically there is no remaining pointer to it. We considered
inventing another list_copy variant that would list_free the second
input, but concluded that for most call sites it isn't worth worrying
about, given the relative compactness of the new List representation.
(Note that in cases where such leakage would happen, the old code
already leaked the second List's header; so we're only discussing
the size of the leak not whether there is one. I did adjust two or
three places that had been troubling to free that header so that
they manually free the whole second List.)
Patch by me; thanks to David Rowley for review.
Discussion: https://postgr.es/m/11587.1550975080@sss.pgh.pa.us
Originally, Postgres Lists were a more or less exact reimplementation of
Lisp lists, which consist of chains of separately-allocated cons cells,
each having a value and a next-cell link. We'd hacked that once before
(commit d0b4399d8) to add a separate List header, but the data was still
in cons cells. That makes some operations -- notably list_nth() -- O(N),
and it's bulky because of the next-cell pointers and per-cell palloc
overhead, and it's very cache-unfriendly if the cons cells end up
scattered around rather than being adjacent.
In this rewrite, we still have List headers, but the data is in a
resizable array of values, with no next-cell links. Now we need at
most two palloc's per List, and often only one, since we can allocate
some values in the same palloc call as the List header. (Of course,
extending an existing List may require repalloc's to enlarge the array.
But this involves just O(log N) allocations not O(N).)
Of course this is not without downsides. The key difficulty is that
addition or deletion of a list entry may now cause other entries to
move, which it did not before.
For example, that breaks foreach() and sister macros, which historically
used a pointer to the current cons-cell as loop state. We can repair
those macros transparently by making their actual loop state be an
integer list index; the exposed "ListCell *" pointer is no longer state
carried across loop iterations, but is just a derived value. (In
practice, modern compilers can optimize things back to having just one
loop state value, at least for simple cases with inline loop bodies.)
In principle, this is a semantics change for cases where the loop body
inserts or deletes list entries ahead of the current loop index; but
I found no such cases in the Postgres code.
The change is not at all transparent for code that doesn't use foreach()
but chases lists "by hand" using lnext(). The largest share of such
code in the backend is in loops that were maintaining "prev" and "next"
variables in addition to the current-cell pointer, in order to delete
list cells efficiently using list_delete_cell(). However, we no longer
need a previous-cell pointer to delete a list cell efficiently. Keeping
a next-cell pointer doesn't work, as explained above, but we can improve
matters by changing such code to use a regular foreach() loop and then
using the new macro foreach_delete_current() to delete the current cell.
(This macro knows how to update the associated foreach loop's state so
that no cells will be missed in the traversal.)
There remains a nontrivial risk of code assuming that a ListCell *
pointer will remain good over an operation that could now move the list
contents. To help catch such errors, list.c can be compiled with a new
define symbol DEBUG_LIST_MEMORY_USAGE that forcibly moves list contents
whenever that could possibly happen. This makes list operations
significantly more expensive so it's not normally turned on (though it
is on by default if USE_VALGRIND is on).
There are two notable API differences from the previous code:
* lnext() now requires the List's header pointer in addition to the
current cell's address.
* list_delete_cell() no longer requires a previous-cell argument.
These changes are somewhat unfortunate, but on the other hand code using
either function needs inspection to see if it is assuming anything
it shouldn't, so it's not all bad.
Programmers should be aware of these significant performance changes:
* list_nth() and related functions are now O(1); so there's no
major access-speed difference between a list and an array.
* Inserting or deleting a list element now takes time proportional to
the distance to the end of the list, due to moving the array elements.
(However, it typically *doesn't* require palloc or pfree, so except in
long lists it's probably still faster than before.) Notably, lcons()
used to be about the same cost as lappend(), but that's no longer true
if the list is long. Code that uses lcons() and list_delete_first()
to maintain a stack might usefully be rewritten to push and pop at the
end of the list rather than the beginning.
* There are now list_insert_nth...() and list_delete_nth...() functions
that add or remove a list cell identified by index. These have the
data-movement penalty explained above, but there's no search penalty.
* list_concat() and variants now copy the second list's data into
storage belonging to the first list, so there is no longer any
sharing of cells between the input lists. The second argument is
now declared "const List *" to reflect that it isn't changed.
This patch just does the minimum needed to get the new implementation
in place and fix bugs exposed by the regression tests. As suggested
by the foregoing, there's a fair amount of followup work remaining to
do.
Also, the ENABLE_LIST_COMPAT macros are finally removed in this
commit. Code using those should have been gone a dozen years ago.
Patch by me; thanks to David Rowley, Jesper Pedersen, and others
for review.
Discussion: https://postgr.es/m/11587.1550975080@sss.pgh.pa.us
If we need ordered output from a scan of a partitioned table, but
the ordering matches the partition ordering, then we don't need to
use a MergeAppend to combine the pre-ordered per-partition scan
results: a plain Append will produce the same results. This
both saves useless comparison work inside the MergeAppend proper,
and allows us to start returning tuples after istarting up just
the first child node not all of them.
However, all is not peaches and cream, because if some of the
child nodes have high startup costs then there will be big
discontinuities in the tuples-returned-versus-elapsed-time curve.
The planner's cost model cannot handle that (yet, anyway).
If we model the Append's startup cost as being just the first
child's startup cost, we may drastically underestimate the cost
of fetching slightly more tuples than are available from the first
child. Since we've had bad experiences with over-optimistic choices
of "fast start" plans for ORDER BY LIMIT queries, that seems scary.
As a klugy workaround, set the startup cost estimate for an ordered
Append to be the sum of its children's startup costs (as MergeAppend
would). This doesn't really describe reality, but it's less likely
to cause a bad plan choice than an underestimated startup cost would.
In practice, the cases where we really care about this optimization
will have child plans that are IndexScans with zero startup cost,
so that the overly conservative estimate is still just zero.
David Rowley, reviewed by Julien Rouhaud and Antonin Houska
Discussion: https://postgr.es/m/CAKJS1f-hAqhPLRk_RaSFTgYxd=Tz5hA7kQ2h4-DhJufQk8TGuw@mail.gmail.com
Add support function requests for estimating the selectivity, cost,
and number of result rows (if a SRF) of the target function.
The lack of a way to estimate selectivity of a boolean-returning
function in WHERE has been a recognized deficiency of the planner
since Berkeley days. This commit finally fixes it.
In addition, non-constant estimates of cost and number of output
rows are now possible. We still fall back to looking at procost
and prorows if the support function doesn't service the request,
of course.
To make concrete use of the possibility of estimating output rowcount
for SRFs, this commit adds support functions for array_unnest(anyarray)
and the integer variants of generate_series; the lack of plausible
rowcount estimates for those, even when it's obvious to a human,
has been a repeated subject of complaints. Obviously, much more
could now be done in this line, but I'm mostly just trying to get
the infrastructure in place.
Discussion: https://postgr.es/m/15193.1548028093@sss.pgh.pa.us
In place of three separate but interrelated lists (indexclauses,
indexquals, and indexqualcols), an IndexPath now has one list
"indexclauses" of IndexClause nodes. This holds basically the same
information as before, but in a more useful format: in particular, there
is now a clear connection between an indexclause (an original restriction
clause from WHERE or JOIN/ON) and the indexquals (directly usable index
conditions) derived from it.
We also change the ground rules a bit by mandating that clause commutation,
if needed, be done up-front so that what is stored in the indexquals list
is always directly usable as an index condition. This gets rid of repeated
re-determination of which side of the clause is the indexkey during costing
and plan generation, as well as repeated lookups of the commutator
operator. To minimize the added up-front cost, the typical case of
commuting a plain OpExpr is handled by a new special-purpose function
commute_restrictinfo(). For RowCompareExprs, generating the new clause
properly commuted to begin with is not really any more complex than before,
it's just different --- and we can save doing that work twice, as the
pretty-klugy original implementation did.
Tracking the connection between original and derived clauses lets us
also track explicitly whether the derived clauses are an exact or lossy
translation of the original. This provides a cheap solution to getting
rid of unnecessary rechecks of boolean index clauses, which previously
seemed like it'd be more expensive than it was worth.
Another pleasant (IMO) side-effect is that EXPLAIN now always shows
index clauses with the indexkey on the left; this seems less confusing.
This commit leaves expand_indexqual_conditions() and some related
functions in a slightly messy state. I didn't bother to change them
any more than minimally necessary to work with the new data structure,
because all that code is going to be refactored out of existence in
a follow-on patch.
Discussion: https://postgr.es/m/22182.1549124950@sss.pgh.pa.us
The old name of this file was never a very good indication of what it
was for. Now that there's also access/relation.h, we have a potential
confusion hazard as well, so let's rename it to something more apropos.
Per discussion, "pathnodes.h" is reasonable, since a good fraction of
the file is Path node definitions.
While at it, tweak a couple of other headers that were gratuitously
importing relation.h into modules that don't need it.
Discussion: https://postgr.es/m/7719.1548688728@sss.pgh.pa.us
Create a new header optimizer/optimizer.h, which exposes just the
planner functions that can be used "at arm's length", without need
to access Paths or the other planner-internal data structures defined
in nodes/relation.h. This is intended to provide the whole planner
API seen by most of the rest of the system; although FDWs still need
to use additional stuff, and more thought is also needed about just
what selfuncs.c should rely on.
The main point of doing this now is to limit the amount of new
#include baggage that will be needed by "planner support functions",
which I expect to introduce later, and which will be in relevant
datatype modules rather than anywhere near the planner.
This commit just moves relevant declarations into optimizer.h from
other header files (a couple of which go away because everything
got moved), and adjusts #include lists to match. There's further
cleanup that could be done if we want to decide that some stuff
being exposed by optimizer.h doesn't belong in the planner at all,
but I'll leave that for another day.
Discussion: https://postgr.es/m/11460.1548706639@sss.pgh.pa.us
Move a few very simple node-creation and node-type-testing functions
from the planner's clauses.c to nodes/makefuncs and nodes/nodeFuncs.
There's nothing planner-specific about them, as evidenced by the
number of other places that were using them.
While at it, rename and_clause() etc to is_andclause() etc, to clarify
that they are node-type-testing functions not node-creation functions.
And use "static inline" implementations for the shortest ones.
Also, modify flatten_join_alias_vars() and some subsidiary functions
to take a Query not a PlannerInfo to define the join structure that
Vars should be translated according to. They were only using the
"parse" field of the PlannerInfo anyway, so this just requires removing
one level of indirection. The advantage is that now parse_agg.c can
use flatten_join_alias_vars() without the horrid kluge of creating an
incomplete PlannerInfo, which will allow that file to be decoupled from
relation.h in a subsequent patch.
Discussion: https://postgr.es/m/11460.1548706639@sss.pgh.pa.us
The fact that "SELECT expression" has no base relations has long been a
thorn in the side of the planner. It makes it hard to flatten a sub-query
that looks like that, or is a trivial VALUES() item, because the planner
generally uses relid sets to identify sub-relations, and such a sub-query
would have an empty relid set if we flattened it. prepjointree.c contains
some baroque logic that works around this in certain special cases --- but
there is a much better answer. We can replace an empty FROM clause with a
dummy RTE that acts like a table of one row and no columns, and then there
are no such corner cases to worry about. Instead we need some logic to
get rid of useless dummy RTEs, but that's simpler and covers more cases
than what was there before.
For really trivial cases, where the query is just "SELECT expression" and
nothing else, there's a hazard that adding the extra RTE makes for a
noticeable slowdown; even though it's not much processing, there's not
that much for the planner to do overall. However testing says that the
penalty is very small, close to the noise level. In more complex queries,
this is able to find optimizations that we could not find before.
The new RTE type is called RTE_RESULT, since the "scan" plan type it
gives rise to is a Result node (the same plan we produced for a "SELECT
expression" query before). To avoid confusion, rename the old ResultPath
path type to GroupResultPath, reflecting that it's only used in degenerate
grouping cases where we know the query produces just one grouped row.
(It wouldn't work to unify the two cases, because there are different
rules about where the associated quals live during query_planner.)
Note: although this touches readfuncs.c, I don't think a catversion
bump is required, because the added case can't occur in stored rules,
only plans.
Patch by me, reviewed by David Rowley and Mark Dilger
Discussion: https://postgr.es/m/15944.1521127664@sss.pgh.pa.us
Up to now we've not worried much about joins where the join key is a
relation's CTID column, reasoning that storing a table's CTIDs in some
other table would be pretty useless. However, there are use-cases for
this sort of query involving self-joins, so that argument doesn't really
hold water.
This patch allows generating plans for joins on CTID that use a nestloop
with inner TidScan, similar to what we might do with an index on the join
column. This is the most efficient way to join when the outer side of
the nestloop is expected to yield relatively few rows.
This change requires upgrading tidpath.c and the generated TidPaths
to work with RestrictInfos instead of bare qual clauses, but that's
long-postponed technical debt anyway.
Discussion: https://postgr.es/m/17443.1545435266@sss.pgh.pa.us
"rescanratio" was computed as 1 + rescanned-tuples / total-inner-tuples,
which is sensible if it's to be multiplied by total-inner-tuples or a cost
value corresponding to scanning all the inner tuples. But in reality it
was (mostly) multiplied by inner_rows or a related cost, numbers that take
into account the possibility of stopping short of scanning the whole inner
relation thanks to a limited key range in the outer relation. This'd
still make sense if we could expect that stopping short would result in a
proportional decrease in the number of tuples that have to be rescanned.
It does not, however. The argument that establishes the validity of our
estimate for that number is independent of whether we scan all of the inner
relation or stop short, and experimentation also shows that stopping short
doesn't reduce the number of rescanned tuples. So the correct calculation
is 1 + rescanned-tuples / inner_rows, and we should be sure to multiply
that by inner_rows or a corresponding cost value.
Most of the time this doesn't make much difference, but if we have
both a high rescan rate (due to lots of duplicate values) and an outer
key range much smaller than the inner key range, then the error can
be significant, leading to a large underestimate of the cost associated
with rescanning.
Per report from Vijaykumar Jain. This thinko appears to go all the way
back to the introduction of the rescan estimation logic in commit
70fba7043, so back-patch to all supported branches.
Discussion: https://postgr.es/m/CAE7uO5hMb_TZYJcZmLAgO6iD68AkEK6qCe7i=vZUkCpoKns+EQ@mail.gmail.com
In final_cost_hashjoin(), commit 9c7f5229a allowed inner_unique cases
to follow a code path previously used only for SEMI/ANTI joins; but it
neglected to fix an if-test within that path that assumed SEMI and ANTI
were the only possible cases. This resulted in a wrong value for
hashjointuples, and an ensuing bad cost estimate, for inner_unique normal
joins. Fortunately, for inner_unique normal joins we can assume the number
of joined tuples is the same as for a SEMI join; so there's no need for
more code, we just have to invert the test to check for ANTI not SEMI.
It turns out that in two contrib tests in which commit 9c7f5229a
changed the plan expected for a query, the change was actually wrong
and induced by this estimation error, not by any real improvement.
Hence this patch also reverts those changes.
Per report from RK Korlapati. Backpatch to v10 where the error was
introduced.
David Rowley
Discussion: https://postgr.es/m/CA+SNy03bhq0fodsfOkeWDCreNjJVjsdHwUsb7AG=jpe0PtZc_g@mail.gmail.com
We include <float.h> in every place that needs isnan(), because MSVC
used to require it. However, since MSVC 2013 that's no longer necessary
(cf. commit cec8394b5c), so we can retire the inclusion to a
version-specific stanza in win32_port.h, where it doesn't need to
pollute random .c files. The header is of course still needed in a few
places for other reasons.
I (Álvaro) removed float.h from a few more files than in Emre's original
patch. This doesn't break the build in my system, but we'll see what
the buildfarm has to say about it all.
Author: Emre Hasegeli
Discussion: https://postgr.es/m/CAE2gYzyc0+5uG+Cd9-BSL7NKC8LSHLNg1Aq2=8ubjnUwut4_iw@mail.gmail.com
This controls both plan-time and execution-time new-style partition
pruning. While finer-grain control is possible (maybe using an enum GUC
instead of boolean), there doesn't seem to be much need for that.
This new parameter controls partition pruning for all queries:
trivially, SELECT queries that affect partitioned tables are naturally
under its control since they are using the new technology. However,
while UPDATE/DELETE queries do not use the new code, we make the new GUC
control their behavior also (stealing control from
constraint_exclusion), because it is more natural, and it leads to a
more natural transition to the future in which those queries will also
use the new pruning code.
Constraint exclusion still controls pruning for regular inheritance
situations (those not involving partitioned tables).
Author: David Rowley
Review: Amit Langote, Ashutosh Bapat, Justin Pryzby, David G. Johnston
Discussion: https://postgr.es/m/CAKJS1f_0HwsxJG9m+nzU+CizxSdGtfe6iF_ykPYBiYft302DCw@mail.gmail.com
On further reflection, commit e5d83995e didn't go far enough: pretty much
everywhere in the planner that examines a clause's is_pushed_down flag
ought to be changed to use the more complicated behavior where we also
check the clause's required_relids. Otherwise we could make incorrect
decisions about whether, say, a clause is safe to use as a hash clause.
Some (many?) of these places are safe as-is, either because they are
never reached while considering a parameterized path, or because there
are additional checks that would reject a pushed-down clause anyway.
However, it seems smarter to just code them all the same way rather
than rely on easily-broken reasoning of that sort.
In support of that, invent a new macro RINFO_IS_PUSHED_DOWN that should
be used in place of direct tests on the is_pushed_down flag.
Like the previous patch, back-patch to all supported branches.
Discussion: https://postgr.es/m/f8128b11-c5bf-3539-48cd-234178b2314d@proxel.se
If the partition keys of input relation are part of the GROUP BY
clause, all the rows belonging to a given group come from a single
partition. This allows aggregation/grouping over a partitioned
relation to be broken down * into aggregation/grouping on each
partition. This should be no worse, and often better, than the normal
approach.
If the GROUP BY clause does not contain all the partition keys, we can
still perform partial aggregation for each partition and then finalize
aggregation after appending the partial results. This is less certain
to be a win, but it's still useful.
Jeevan Chalke, Ashutosh Bapat, Robert Haas. The larger patch series
of which this patch is a part was also reviewed and tested by Antonin
Houska, Rajkumar Raghuwanshi, David Rowley, Dilip Kumar, Konstantin
Knizhnik, Pascal Legrand, and Rafia Sabih.
Discussion: http://postgr.es/m/CAM2+6=V64_xhstVHie0Rz=KPEQnLJMZt_e314P0jaT_oJ9MR8A@mail.gmail.com
Commit 1804284042 established that single-batch
parallel-aware hash joins could create one large shared hash table using the
combined work_mem budget of all participants. The costing accidentally
assumed that parallel-oblivious hash joins could also do that. The
documentation for initial_cost_hashjoin() also failed to mention the new
argument. Repair.
Author: Thomas Munro
Reported-By: Antonin Houska
Reviewed-By: Antonin Houska
Discussion: https://postgr.es/m/12441.1513935950%40localhost
Previously, Append didn't charge anything at all, and MergeAppend
charged only cpu_operator_cost, about half the value used here. This
change might make MergeAppend plans slightly more likely to be chosen
than before, since this commit increases the assumed cost for Append
-- with default values -- by 0.005 per tuple but MergeAppend by only
0.0025 per tuple. Since the comparisons required by MergeAppend are
costed separately, it's not clear why MergeAppend needs to be
otherwise more expensive than Append, so hopefully this is OK.
Prior to partition-wise join, it didn't really matter whether or not
an Append node had any cost of its own, because every plan had to use
the same number of Append or MergeAppend nodes and in the same places.
Only the relative cost of Append vs. MergeAppend made a difference.
Now, however, it is possible to avoid some of the Append nodes using a
partition-wise join, so it's worth making an effort. Pending patches
for partition-wise aggregate care too, because an Append of Aggregate
nodes will incur the Append overhead fewer times than an Aggregate
over an Append. Although in most cases this change will favor the use
of partition-wise techniques, it does the opposite when the join
cardinality is greater than the sum of the input cardinalities. Since
this situation arises in an existing regression test, I [rhaas]
adjusted it to keep the overall plan shape approximately the same.
Jeevan Chalke, per a suggestion from David Rowley. Reviewed by
Ashutosh Bapat. Some changes by me. The larger patch series of which
this patch is a part was also reviewed and tested by Antonin Houska,
Rajkumar Raghuwanshi, David Rowley, Dilip Kumar, Konstantin Knizhnik,
Pascal Legrand, Rafia Sabih, and me.
Discussion: http://postgr.es/m/CAKJS1f9UXdk6ZYyqbJnjFO9a9hyHKGW7B=ZRh-rxy9qxfPA5Gw@mail.gmail.com
To make this work, tuplesort.c and logtape.c must also support
parallelism, so this patch adds that infrastructure and then applies
it to the particular case of parallel btree index builds. Testing
to date shows that this can often be 2-3x faster than a serial
index build.
The model for deciding how many workers to use is fairly primitive
at present, but it's better than not having the feature. We can
refine it as we get more experience.
Peter Geoghegan with some help from Rushabh Lathia. While Heikki
Linnakangas is not an author of this patch, he wrote other patches
without which this feature would not have been possible, and
therefore the release notes should possibly credit him as an author
of this feature. Reviewed by Claudio Freire, Heikki Linnakangas,
Thomas Munro, Tels, Amit Kapila, me.
Discussion: http://postgr.es/m/CAM3SWZQKM=Pzc=CAHzRixKjp2eO5Q0Jg1SoFQqeXFQ647JiwqQ@mail.gmail.com
Discussion: http://postgr.es/m/CAH2-Wz=AxWqDoVvGU7dq856S4r6sJAj6DBn7VMtigkB33N5eyg@mail.gmail.com
- Remove unnecessary #include mistakenly added in execnodes.h.
- Fix mistake in comment in choose_next_subplan_for_leader.
- Adjust row estimates in cost_append for a possibly-different
parallel divisor.
- Clamp row estimates in cost_append after operations that may
not produce integers.
Amit Kapila, with cosmetic adjustments by me.
Discussion: http://postgr.es/m/CAA4eK1+qcbeai3coPpRW=GFCzFeLUsuY4T-AKHqMjxpEGZBPQg@mail.gmail.com
Introduce parallel-aware hash joins that appear in EXPLAIN plans as Parallel
Hash Join with Parallel Hash. While hash joins could already appear in
parallel queries, they were previously always parallel-oblivious and had a
partial subplan only on the outer side, meaning that the work of the inner
subplan was duplicated in every worker.
After this commit, the planner will consider using a partial subplan on the
inner side too, using the Parallel Hash node to divide the work over the
available CPU cores and combine its results in shared memory. If the join
needs to be split into multiple batches in order to respect work_mem, then
workers process different batches as much as possible and then work together
on the remaining batches.
The advantages of a parallel-aware hash join over a parallel-oblivious hash
join used in a parallel query are that it:
* avoids wasting memory on duplicated hash tables
* avoids wasting disk space on duplicated batch files
* divides the work of building the hash table over the CPUs
One disadvantage is that there is some communication between the participating
CPUs which might outweigh the benefits of parallelism in the case of small
hash tables. This is avoided by the planner's existing reluctance to supply
partial plans for small scans, but it may be necessary to estimate
synchronization costs in future if that situation changes. Another is that
outer batch 0 must be written to disk if multiple batches are required.
A potential future advantage of parallel-aware hash joins is that right and
full outer joins could be supported, since there is a single set of matched
bits for each hashtable, but that is not yet implemented.
A new GUC enable_parallel_hash is defined to control the feature, defaulting
to on.
Author: Thomas Munro
Reviewed-By: Andres Freund, Robert Haas
Tested-By: Rafia Sabih, Prabhat Sahu
Discussion:
https://postgr.es/m/CAEepm=2W=cOkiZxcg6qiFQP-dHUe09aqTrEMM7yJDrHMhDv_RA@mail.gmail.comhttps://postgr.es/m/CAEepm=37HKyJ4U6XOLi=JgfSHM3o6B-GaeO-6hkOmneTDkH+Uw@mail.gmail.com
When we create an Append node, we can spread out the workers over the
subplans instead of piling on to each subplan one at a time, which
should typically be a bit more efficient, both because the startup
cost of any plan executed entirely by one worker is paid only once and
also because of reduced contention. We can also construct Append
plans using a mix of partial and non-partial subplans, which may allow
for parallelism in places that otherwise couldn't support it.
Unfortunately, this patch doesn't handle the important case of
parallelizing UNION ALL by running each branch in a separate worker;
the executor infrastructure is added here, but more planner work is
needed.
Amit Khandekar, Robert Haas, Amul Sul, reviewed and tested by
Ashutosh Bapat, Amit Langote, Rafia Sabih, Amit Kapila, and
Rajkumar Raghuwanshi.
Discussion: http://postgr.es/m/CAJ3gD9dy0K_E8r727heqXoBmWZ83HwLFwdcaSSmBQ1+S+vRuUQ@mail.gmail.com
For some reason, we have never accounted for either the evaluation cost
or the selectivity of filter conditions attached to Agg and Group nodes
(which, in practice, are always conditions from a HAVING clause).
Applying our regular selectivity logic to post-grouping conditions is a
bit bogus, but it's surely better than taking the selectivity as 1.0.
Perhaps someday the extended-statistics mechanism can be taught to provide
statistics that would help us in getting non-default estimates here.
Per a gripe from Benjamin Coutu. This is surely a bug fix, but I'm
hesitant to back-patch because of the prospect of destabilizing existing
plan choices. Given that it took us this long to notice the bug, it's
probably not hurting too many people in the field.
Discussion: https://postgr.es/m/20968.1509486337@sss.pgh.pa.us
Instead of joining two partitioned tables in their entirety we can, if
it is an equi-join on the partition keys, join the matching partitions
individually. This involves teaching the planner about "other join"
rels, which are related to regular join rels in the same way that
other member rels are related to baserels. This can use significantly
more CPU time and memory than regular join planning, because there may
now be a set of "other" rels not only for every base relation but also
for every join relation. In most practical cases, this probably
shouldn't be a problem, because (1) it's probably unusual to join many
tables each with many partitions using the partition keys for all
joins and (2) if you do that scenario then you probably have a big
enough machine to handle the increased memory cost of planning and (3)
the resulting plan is highly likely to be better, so what you spend in
planning you'll make up on the execution side. All the same, for now,
turn this feature off by default.
Currently, we can only perform joins between two tables whose
partitioning schemes are absolutely identical. It would be nice to
cope with other scenarios, such as extra partitions on one side or the
other with no match on the other side, but that will have to wait for
a future patch.
Ashutosh Bapat, reviewed and tested by Rajkumar Raghuwanshi, Amit
Langote, Rafia Sabih, Thomas Munro, Dilip Kumar, Antonin Houska, Amit
Khandekar, and by me. A few final adjustments by me.
Discussion: http://postgr.es/m/CAFjFpRfQ8GrQvzp3jA2wnLqrHmaXna-urjm_UY9BqXj=EaDTSA@mail.gmail.com
Discussion: http://postgr.es/m/CAFjFpRcitjfrULr5jfuKWRPsGUX0LQ0k8-yG0Qw2+1LBGNpMdw@mail.gmail.com
Allowing arrays with a domain type as their element type was left un-done
in the original domain patch, but not for any very good reason. This
omission leads to such surprising results as array_agg() not working on
a domain column, because the parser can't identify a suitable output type
for the polymorphic aggregate.
In order to fix this, first clean up the APIs of coerce_to_domain() and
some internal functions in parse_coerce.c so that we consistently pass
around a CoercionContext along with CoercionForm. Previously, we sometimes
passed an "isExplicit" boolean flag instead, which is strictly less
information; and coerce_to_domain() didn't even get that, but instead had
to reverse-engineer isExplicit from CoercionForm. That's contrary to the
documentation in primnodes.h that says that CoercionForm only affects
display and not semantics. I don't think this change fixes any live bugs,
but it makes things more consistent. The main reason for doing it though
is that now build_coercion_expression() receives ccontext, which it needs
in order to be able to recursively invoke coerce_to_target_type().
Next, reimplement ArrayCoerceExpr so that the node does not directly know
any details of what has to be done to the individual array elements while
performing the array coercion. Instead, the per-element processing is
represented by a sub-expression whose input is a source array element and
whose output is a target array element. This simplifies life in
parse_coerce.c, because it can build that sub-expression by a recursive
invocation of coerce_to_target_type(). The executor now handles the
per-element processing as a compiled expression instead of hard-wired code.
The main advantage of this is that we can use a single ArrayCoerceExpr to
handle as many as three successive steps per element: base type conversion,
typmod coercion, and domain constraint checking. The old code used two
stacked ArrayCoerceExprs to handle type + typmod coercion, which was pretty
inefficient, and adding yet another array deconstruction to do domain
constraint checking seemed very unappetizing.
In the case where we just need a single, very simple coercion function,
doing this straightforwardly leads to a noticeable increase in the
per-array-element runtime cost. Hence, add an additional shortcut evalfunc
in execExprInterp.c that skips unnecessary overhead for that specific form
of expression. The runtime speed of simple cases is within 1% or so of
where it was before, while cases that previously required two levels of
array processing are significantly faster.
Finally, create an implicit array type for every domain type, as we do for
base types, enums, etc. Everything except the array-coercion case seems
to just work without further effort.
Tom Lane, reviewed by Andrew Dunstan
Discussion: https://postgr.es/m/9852.1499791473@sss.pgh.pa.us
Jeff Davis
Peter Geoghegan
Andres Freund
Peter Eisentraut
Tom Lane
Tomas Vondra
Bruce Momjian
Michael Paquier
Alvaro Herrera
Robert Haas
Peter Eisentraut