The logic added in 9d9c02ccd to determine when a qual can be used as a
WindowClause run condition failed to correctly check for subqueries in the
qual. This was being done correctly for normal subquery qual pushdowns,
it's just that 9d9c02ccd failed to follow the lead on that.
This also fixes various other cases where transforming the qual into a
WindowClause run condition in the subquery should have been disallowed.
Bug: #17826
Reported-by: Anban Company
Discussion: https://postgr.es/m/17826-7d8750952f19a5f5@postgresql.org
Backpatch-through: 15, where 9d9c02ccd was introduced.
preprocess_targetlist thought PHVs couldn't appear here.
It was mistaken, as per report from Önder Kalacı.
Surveying other pull_var_clause calls, I noted no similar errors,
but I did notice that qual_is_pushdown_safe's assertion about
!contain_window_function was pointless, because the following
pull_var_clause call would complain about them anyway. In HEAD
only, remove the redundant Assert and improve the commentary.
Discussion: https://postgr.es/m/CACawEhUuum-gC_2S3sXLTcsk7bUSPSHOD+g1ZpfKaDK-KKPPWA@mail.gmail.com
adjust_appendrel_attrs can't transfer nullingrel labeling to a non-Var
translation expression (mainly because it's too late to wrap such an
expression in a PlaceHolderVar). I'd supposed in commit 2489d76c4
that that restriction was unreachable because we'd not attempt to push
problematic clauses down to an appendrel child relation. I forgot that
set_append_rel_size blindly converts all the parent rel's joininfo
clauses to child clauses, and that list could well contain clauses
from above a nulling outer join.
We might eventually have to devise a direct fix for this implementation
restriction, but for now it seems enough to filter out troublesome
clauses while constructing the child's joininfo list. Such clauses
are certainly not useful while constructing paths for the child rel;
they'll have to be applied later when we join the completed appendrel
to something else. So we don't need them here, and omitting them from
the list should save a few cycles while processing the child rel.
Per bug #17832 from Marko Tiikkaja.
Discussion: https://postgr.es/m/17832-d0a8106cdf1b722e@postgresql.org
In generate_orderedappend_paths(), when match_partition_order_desc was
true, we would lcons() items to various lists in a loop over each live
partition. When the number of live partitions was large, the lcons()
could show up in profiles due to it having to perform memmove() to make
way for the new list item.
Here we adjust things so that we just perform the loop over the live
partitions backwards when match_partition_order_desc is true. This allows
us to simplify the logic in the loop. Now, as far as the guts of the loop
knows, there's no difference between match_partition_order and
match_partition_order_desc. We can just set match_partition_order to true
so that we build the correct list of paths for the asc and desc case. Per
idea from Andres Freund.
Discussion: https://postgr.es/m/20230217002351.nyt4y5tdzg6hugdt@awork3.anarazel.de
Remove RestrictInfo.nullable_relids, along with a good deal of
infrastructure that calculated it. One use-case for it was in
join_clause_is_movable_to, but we can now replace that usage with
a check to see if the clause's relids include any outer join
that can null the target relation. The other use-case was in
join_clause_is_movable_into, but that test can just be dropped
entirely now that the clause's relids include outer joins.
Furthermore, join_clause_is_movable_into should now be
accurate enough that it will accept anything returned by
generate_join_implied_equalities, so we can restore the Assert
that was diked out in commit 95f4e59c3.
Remove the outerjoin_delayed mechanism. We needed this before to
prevent quals from getting evaluated below outer joins that should
null some of their vars. Now that we consider varnullingrels while
placing quals, that's taken care of automatically, so throw the
whole thing away.
Teach remove_useless_result_rtes to also remove useless FromExprs.
Having done that, the delay_upper_joins flag serves no purpose any
more and we can remove it, largely reverting 11086f2f2.
Use constant TRUE for "dummy" clauses when throwing back outer joins.
This improves on a hack I introduced in commit 6a6522529. If we
have a left-join clause l.x = r.y, and a WHERE clause l.x = constant,
we generate r.y = constant and then don't really have a need for the
join clause. But we must throw the join clause back anyway after
marking it redundant, so that the join search heuristics won't think
this is a clauseless join and avoid it. That was a kluge introduced
under time pressure, and after looking at it I thought of a better
way: let's just introduce constant-TRUE "join clauses" instead,
and get rid of them at the end. This improves the generated plans for
such cases by not having to test a redundant join clause. We can also
get rid of the ugly hack used to mark such clauses as redundant for
selectivity estimation.
Patch by me; thanks to Richard Guo for review.
Discussion: https://postgr.es/m/830269.1656693747@sss.pgh.pa.us
Traditionally we used the same Var struct to represent the value
of a table column everywhere in parse and plan trees. This choice
predates our support for SQL outer joins, and it's really a pretty
bad idea with outer joins, because the Var's value can depend on
where it is in the tree: it might go to NULL above an outer join.
So expression nodes that are equal() per equalfuncs.c might not
represent the same value, which is a huge correctness hazard for
the planner.
To improve this, decorate Var nodes with a bitmapset showing
which outer joins (identified by RTE indexes) may have nulled
them at the point in the parse tree where the Var appears.
This allows us to trust that equal() Vars represent the same value.
A certain amount of klugery is still needed to cope with cases
where we re-order two outer joins, but it's possible to make it
work without sacrificing that core principle. PlaceHolderVars
receive similar decoration for the same reason.
In the planner, we include these outer join bitmapsets into the relids
that an expression is considered to depend on, and in consequence also
add outer-join relids to the relids of join RelOptInfos. This allows
us to correctly perceive whether an expression can be calculated above
or below a particular outer join.
This change affects FDWs that want to plan foreign joins. They *must*
follow suit when labeling foreign joins in order to match with the
core planner, but for many purposes (if postgres_fdw is any guide)
they'd prefer to consider only base relations within the join.
To support both requirements, redefine ForeignScan.fs_relids as
base+OJ relids, and add a new field fs_base_relids that's set up by
the core planner.
Large though it is, this commit just does the minimum necessary to
install the new mechanisms and get check-world passing again.
Follow-up patches will perform some cleanup. (The README additions
and comments mention some stuff that will appear in the follow-up.)
Patch by me; thanks to Richard Guo for review.
Discussion: https://postgr.es/m/830269.1656693747@sss.pgh.pa.us
When incremental sorts were added in v13 a 1.5x pessimism factor was added
to the cost modal. Seemingly this was done because the cost modal only
has an estimate of the total number of input rows and the number of
presorted groups. It assumes that the input rows will be evenly
distributed throughout the presorted groups. The 1.5x pessimism factor
was added to slightly reduce the likelihood of incremental sorts being
used in the hope to avoid performance regressions where an incremental
sort plan was picked and turned out slower due to a large skew in the
number of rows in the presorted groups.
An additional quirk with the path generation code meant that we could
consider both a sort and an incremental sort on paths with presorted keys.
This meant that with the pessimism factor, it was possible that we opted
to perform a sort rather than an incremental sort when the given path had
presorted keys.
Here we remove the 1.5x pessimism factor to allow incremental sorts to
have a fairer chance at being chosen against a full sort.
Previously we would generally create a sort path on the cheapest input
path (if that wasn't sorted already) and incremental sort paths on any
path which had presorted keys. This meant that if the cheapest input path
wasn't completely sorted but happened to have presorted keys, we would
create a full sort path *and* an incremental sort path on that input path.
Here we change this logic so that if there are presorted keys, we only
create an incremental sort path, and create sort paths only when a full
sort is required.
Both the removal of the cost pessimism factor and the changes made to the
path generation make it more likely that incremental sorts will now be
chosen. That, of course, as with teaching the planner any new tricks,
means an increased likelihood that the planner will perform an incremental
sort when it's not the best method. Our standard escape hatch for these
cases is an enable_* GUC. enable_incremental_sort already exists for
this.
This came out of a report by Pavel Luzanov where he mentioned that the
master branch was choosing to perform a Seq Scan -> Sort -> Group
Aggregate for his query with an ORDER BY aggregate function. The v15 plan
for his query performed an Index Scan -> Group Aggregate, of course, the
aggregate performed the final sort internally in nodeAgg.c for the
aggregate's ORDER BY. The ideal plan would have been to use the index,
which provided partially sorted input then use an incremental sort to
provide the aggregate with the sorted input. This was not being chosen
due to the pessimism in the incremental sort cost modal, so here we remove
that and rationalize the path generation so that sort and incremental sort
plans don't have to needlessly compete. We assume that it's senseless
to ever use a full sort on a given input path where an incremental sort
can be performed.
Reported-by: Pavel Luzanov
Reviewed-by: Richard Guo
Discussion: https://postgr.es/m/9f61ddbf-2989-1536-b31e-6459370a6baa%40postgrespro.ru
9d9c02ccd added window "run conditions", which allows the evaluation of
monotonic window functions to be skipped when the run condition is no
longer true. Prior to this commit, once the run condition was no longer
true and we stopped evaluating the window functions, we simply just left
the ecxt_aggvalues[] and ecxt_aggnulls[] arrays alone to store whatever
value was stored there the last time the window function was evaluated.
Leaving a stale value in there isn't really a problem on 64-bit builds as
all of the window functions which we recognize as monotonic all return
int8, which is passed by value on 64-bit builds. However, on 32-bit
builds, this was a problem as the value stored in the ecxt_values[]
element would be a by-ref value and it would be pointing to some memory
which would get reset once the tuple context is destroyed. Since the
WindowAgg node will output these values in the resulting tupleslot, this
could be problematic for the top-level WindowAgg node which must look at
these values to filter out the rows that don't meet its filter condition.
Here we fix this by just zeroing the ecxt_aggvalues[] and setting the
ecxt_aggnulls[] array to true when the run condition first becomes false.
This results in the WindowAgg's output having NULLs for the WindowFunc's
columns rather than the stale or pointer pointing to possibly freed
memory. These tuples with the NULLs can only make it as far as the
top-level WindowAgg node before they're filtered out. To ensure that
these tuples *are* always filtered out, we now insist that OpExprs making
up the run condition are strict OpExprs. Currently, all the window
functions which the planner recognizes as monotonic return INT8 and the
operator which is used for the run condition must be a member of a btree
opclass. In reality, these restrictions exclude nothing that's built-in
to Postgres and are unlikely to exclude anyone's custom operators due to
the requirement that the operator is part of a btree opclass. It would be
unusual if those were not strict.
Reported-by: Sergey Shinderuk, using valgrind
Reviewed-by: Richard Guo, Sergey Shinderuk
Discussion: https://postgr.es/m/29184c50-429a-ebd7-f1fb-0589c6723a35@postgrespro.ru
Backpatch-through: 15, where 9d9c02ccd was added
We might fail to generate a partitionwise join, because
reparameterize_path_by_child() does not support all path types.
This should not be a hard failure condition: we should just fall back
to a non-partitioned join. However, generate_partitionwise_join_paths
did not consider this possibility and would emit the (misleading)
error "could not devise a query plan for the given query" if we'd
failed to make any paths for a child join. Fix it to give up on
partitionwise joining if so. (The accepted technique for giving up
appears to be to set rel->nparts = 0, which I find pretty bizarre,
but there you have it.)
I have not added a test case because there'd be little point:
any omissions of this sort that we identify would soon get fixed
by extending reparameterize_path_by_child(), so the test would stop
proving anything. However, right now there is a known test case based
on failure to cover MaterialPath, and with that I've found that this
is broken in all supported versions. Hence, patch all the way back.
Original report and patch by me; thanks to Richard Guo for
identifying a test case that works against committed versions.
Discussion: https://postgr.es/m/1854233.1669949723@sss.pgh.pa.us
Preprocessing of the HAVING clause will reduce havingQual to NIL
if the clause is constant-TRUE. This is one case where that
convention is rather unfortunate, because "HAVING TRUE" is not at all
the same as not having any HAVING clause at all. (Per the SQL spec,
it still forces the query to be grouped.) The planner deals with this
by having a boolean hasHavingQual that records whether havingQual was
originally nonempty; places that just want to check whether HAVING
was specified are supposed to consult that.
I found three places that got that wrong. Fortunately, these could
only affect cost estimates not correctness. It'd be hard even
to demonstrate the errors; for example, the one in allpaths.c would
only matter in a query that has HAVING TRUE but no GROUP BY and no
aggregates, which would require a completely variable-free SELECT
list, making the case probably of only academic interest. Hence,
while these are worth fixing before someone copies the incorrect
coding somewhere more critical, they don't seem worth back-patching.
I didn't bother trying to devise regression tests, either.
Discussion: https://postgr.es/m/2503888.1666042643@sss.pgh.pa.us
Two callers of generate_useful_gather_paths were testing the wrong
thing when deciding whether to call that function: they checked for
being at the top of the current join subproblem, rather than being at
the actual top join. This'd result in failing to construct parallel
paths for a sub-join for which they might be useful.
While set_rel_pathlist() isn't actively broken, it seems best to
make its identical-in-intention test for this be like the other two.
This has been wrong all along, but given the lack of field complaints
I'm hesitant to back-patch into stable branches; we usually prefer
to avoid non-bug-fix changes in plan choices in minor releases.
It seems not too late for v15 though.
Richard Guo, reviewed by Antonin Houska and Tom Lane
Discussion: https://postgr.es/m/CAMbWs4-mH8Zf87-w+3P2J=nJB+5OyicO28ia9q_9o=Lamf_VHg@mail.gmail.com
Commit 4f658dc8 provided the traditional BSD fls() function in
src/port/fls.c so it could be used in several places. Later we added a
bunch of similar facilities in pg_bitutils.h, based on compiler
builtins that map to hardware instructions. It's a bit confusing to
have both 1-based and 0-based variants of this operation in use in
different parts of the tree, and neither is blessed by a standard.
Let's drop fls.c and the configure probe, and reuse the newer code.
Reviewed-by: David Rowley <dgrowleyml@gmail.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/CA%2BhUKG%2B7dSX1XF8yFGmYk-%3D48dbjH2kmzZj16XvhbrWP-9BzRg%40mail.gmail.com
setrefs.c contains logic to discard no-op SubqueryScan nodes, that is,
ones that have no qual to check and copy the input targetlist unchanged.
(Formally it's not very nice to be applying such optimizations so late
in the planner, but there are practical reasons for it; mostly that we
can't unify relids between the subquery and the parent query until we
flatten the rangetable during setrefs.c.) This behavior falsifies our
previous cost estimates, since we would've charged cpu_tuple_cost per
row just to pass data through the node. Most of the time that's little
enough to not matter, but there are cases where this effect visibly
changes the plan compared to what you would've gotten with no
sub-select.
To improve the situation, make the callers of cost_subqueryscan tell
it whether they think the targetlist is trivial. cost_subqueryscan
already has the qual list, so it can check the other half of the
condition easily. It could make its own determination of tlist
triviality too, but doing so would be repetitive (for callers that
may call it several times) or unnecessarily expensive (for callers
that can determine this more cheaply than a general test would do).
This isn't a 100% solution, because createplan.c also does things
that can falsify any earlier estimate of whether the tlist is
trivial. However, it fixes nearly all cases in practice, if results
for the regression tests are anything to go by.
setrefs.c also contains logic to discard no-op Append and MergeAppend
nodes. We did have knowledge of that behavior at costing time, but
somebody failed to update it when a check on parallel-awareness was
added to the setrefs.c logic. Fix that while we're here.
These changes result in two minor changes in query plans shown in
our regression tests. Neither is relevant to the purposes of its
test case AFAICT.
Patch by me; thanks to Richard Guo for review.
Discussion: https://postgr.es/m/2581077.1651703520@sss.pgh.pa.us
Truncating off the end of a freshly copied List is not a very efficient
way of copying the first N elements of a List.
In many of the cases that are updated here, the pattern was only being
used to remove the final element of a List. That's about the best case
for it, but there were many instances where the truncate trimming the List
down much further.
4cc832f94 added list_copy_head(), so let's use it in cases where it's
useful.
Author: David Rowley
Discussion: https://postgr.es/m/1986787.1657666922%40sss.pgh.pa.us
9d9c02ccd added code to allow the executor to take shortcuts when quals
on monotonic window functions guaranteed that once the qual became false
it could never become true again. When possible, baserestrictinfo quals
are converted to become these quals, which we call run conditions.
Unfortunately, in 9d9c02ccd, I forgot to update
remove_unused_subquery_outputs to teach it about these run conditions.
This could cause a WindowFunc column which was unused in the target list
but referenced by an upper-level WHERE clause to be removed from the
subquery when the qual in the WHERE clause was converted into a window run
condition. Because of this, the entire WindowClause would be removed from
the query resulting in additional rows making it into the resultset when
they should have been filtered out by the WHERE clause.
Here we fix this by recording which target list items in the subquery have
run conditions. That gets passed along to remove_unused_subquery_outputs
to tell it not to remove these items from the target list.
Bug: #17495
Reported-by: Jeremy Evans
Reviewed-by: Richard Guo
Discussion: https://postgr.es/m/17495-7ffe2fa0b261b9fa@postgresql.org
inline_cte() expected to find exactly as many references to the
target CTE as its cterefcount indicates. While that should be
accurate for the tree as emitted by the parser, there are some
optimizations that occur upstream of here that could falsify it,
notably removal of unused subquery output expressions.
Trying to make the accounting 100% accurate seems expensive and
doomed to future breakage. It's not really worth it, because
all this code is protecting is downstream assumptions that every
referenced CTE has a plan. Let's convert those assertions to
regular test-and-elog just in case there's some actual problem,
and then drop the failing assertion.
Per report from Tomas Vondra (thanks also to Richard Guo for
analysis). Back-patch to v12 where the faulty code came in.
Discussion: https://postgr.es/m/29196a1e-ed47-c7ca-9be2-b1c636816183@enterprisedb.com
Window functions such as row_number() always return a value higher than
the previously returned value for tuples in any given window partition.
Traditionally queries such as;
SELECT * FROM (
SELECT *, row_number() over (order by c) rn
FROM t
) t WHERE rn <= 10;
were executed fairly inefficiently. Neither the query planner nor the
executor knew that once rn made it to 11 that nothing further would match
the outer query's WHERE clause. It would blindly continue until all
tuples were exhausted from the subquery.
Here we implement means to make the above execute more efficiently.
This is done by way of adding a pg_proc.prosupport function to various of
the built-in window functions and adding supporting code to allow the
support function to inform the planner if the window function is
monotonically increasing, monotonically decreasing, both or neither. The
planner is then able to make use of that information and possibly allow
the executor to short-circuit execution by way of adding a "run condition"
to the WindowAgg to allow it to determine if some of its execution work
can be skipped.
This "run condition" is not like a normal filter. These run conditions
are only built using quals comparing values to monotonic window functions.
For monotonic increasing functions, quals making use of the btree
operators for <, <= and = can be used (assuming the window function column
is on the left). You can see here that once such a condition becomes false
that a monotonic increasing function could never make it subsequently true
again. For monotonically decreasing functions the >, >= and = btree
operators for the given type can be used for run conditions.
The best-case situation for this is when there is a single WindowAgg node
without a PARTITION BY clause. Here when the run condition becomes false
the WindowAgg node can simply return NULL. No more tuples will ever match
the run condition. It's a little more complex when there is a PARTITION
BY clause. In this case, we cannot return NULL as we must still process
other partitions. To speed this case up we pull tuples from the outer
plan to check if they're from the same partition and simply discard them
if they are. When we find a tuple belonging to another partition we start
processing as normal again until the run condition becomes false or we run
out of tuples to process.
When there are multiple WindowAgg nodes to evaluate then this complicates
the situation. For intermediate WindowAggs we must ensure we always
return all tuples to the calling node. Any filtering done could lead to
incorrect results in WindowAgg nodes above. For all intermediate nodes,
we can still save some work when the run condition becomes false. We've
no need to evaluate the WindowFuncs anymore. Other WindowAgg nodes cannot
reference the value of these and these tuples will not appear in the final
result anyway. The savings here are small in comparison to what can be
saved in the top-level WingowAgg, but still worthwhile.
Intermediate WindowAgg nodes never filter out tuples, but here we change
WindowAgg so that the top-level WindowAgg filters out tuples that don't
match the intermediate WindowAgg node's run condition. Such filters
appear in the "Filter" clause in EXPLAIN for the top-level WindowAgg node.
Here we add prosupport functions to allow the above to work for;
row_number(), rank(), dense_rank(), count(*) and count(expr). It appears
technically possible to do the same for min() and max(), however, it seems
unlikely to be useful enough, so that's not done here.
Bump catversion
Author: David Rowley
Reviewed-by: Andy Fan, Zhihong Yu
Discussion: https://postgr.es/m/CAApHDvqvp3At8++yF8ij06sdcoo1S_b2YoaT9D4Nf+MObzsrLQ@mail.gmail.com
When building append paths, we've been looking only at startup and total
costs for the paths. When building fractional paths that may eliminate
the cheapest one, because it may be dominated by two separate paths (one
for startup, one for total cost).
This extends generate_orderedappend_paths() to also consider which paths
have lowest fractional cost. Currently we only consider paths matching
pathkeys - in the future this may be improved by also considering paths
that are only partially sorted, with an incremental sort on top.
Original report of an issue by Arne Roland, patch by me (based on a
suggestion by Tom Lane).
Reviewed-by: Arne Roland, Zhihong Yu
Discussion: https://postgr.es/m/e8f9ec90-546d-e948-acce-0525f3e92773%40enterprisedb.com
Discussion: https://postgr.es/m/1581042da8044e71ada2d6e3a51bf7bb%40index.de
959d00e9d added the ability to make use of an Append node instead of a
MergeAppend when we wanted to perform a scan of a partitioned table and
the required sort order was the same as the partitioned keys and the
partitioned table was defined in such a way that earlier partitions were
guaranteed to only contain lower-order values than later partitions.
However, previously we didn't allow these ordered partition scans for
LIST partitioned table when there were any partitions that allowed
multiple Datums. This was a very cheap check to make and we could likely
have done a little better by checking if there were interleaved
partitions, but at the time we didn't have visibility about which
partitions were pruned, so we still may have disallowed cases where all
interleaved partitions were pruned.
Since 475dbd0b7, we now have knowledge of pruned partitions, we can do a
much better job inside partitions_are_ordered().
Here we pass which partitions survived partition pruning into
partitions_are_ordered() and, for LIST partitioning, have it check to see
if any live partitions exist that are also in the new "interleaved_parts"
field defined in PartitionBoundInfo.
For RANGE partitioning we can relax the code which caused the partitions
to be unordered if a DEFAULT partition existed. Since we now know which
partitions were pruned, partitions_are_ordered() now returns true when the
DEFAULT partition was pruned.
Reviewed-by: Amit Langote, Zhihong Yu
Discussion: https://postgr.es/m/CAApHDvrdoN_sXU52i=QDXe2k3WAo=EVry29r2+Tq2WYcn2xhEA@mail.gmail.com
"Result Cache" was never a great name for this node, but nobody managed
to come up with another name that anyone liked enough. That was until
David Johnston mentioned "Node Memoization", which Tom Lane revised to
just "Memoize". People seem to like "Memoize", so let's do the rename.
Reviewed-by: Justin Pryzby
Discussion: https://postgr.es/m/20210708165145.GG1176@momjian.us
Backpatch-through: 14, where Result Cache was introduced
I didn't particularly like this function name, as it fails to
express what's going on. Also, returning the sort expression
alone isn't too helpful --- typically, a caller would also
need some other fields of the EquivalenceMember. But the
sole caller really only needs a bool result, so let's make
it "bool relation_can_be_sorted_early()".
Discussion: https://postgr.es/m/91f3ec99-85a4-fa55-ea74-33f85a5c651f@swarm64.com
Here we add a new executor node type named "Result Cache". The planner
can include this node type in the plan to have the executor cache the
results from the inner side of parameterized nested loop joins. This
allows caching of tuples for sets of parameters so that in the event that
the node sees the same parameter values again, it can just return the
cached tuples instead of rescanning the inner side of the join all over
again. Internally, result cache uses a hash table in order to quickly
find tuples that have been previously cached.
For certain data sets, this can significantly improve the performance of
joins. The best cases for using this new node type are for join problems
where a large portion of the tuples from the inner side of the join have
no join partner on the outer side of the join. In such cases, hash join
would have to hash values that are never looked up, thus bloating the hash
table and possibly causing it to multi-batch. Merge joins would have to
skip over all of the unmatched rows. If we use a nested loop join with a
result cache, then we only cache tuples that have at least one join
partner on the outer side of the join. The benefits of using a
parameterized nested loop with a result cache increase when there are
fewer distinct values being looked up and the number of lookups of each
value is large. Also, hash probes to lookup the cache can be much faster
than the hash probe in a hash join as it's common that the result cache's
hash table is much smaller than the hash join's due to result cache only
caching useful tuples rather than all tuples from the inner side of the
join. This variation in hash probe performance is more significant when
the hash join's hash table no longer fits into the CPU's L3 cache, but the
result cache's hash table does. The apparent "random" access of hash
buckets with each hash probe can cause a poor L3 cache hit ratio for large
hash tables. Smaller hash tables generally perform better.
The hash table used for the cache limits itself to not exceeding work_mem
* hash_mem_multiplier in size. We maintain a dlist of keys for this cache
and when we're adding new tuples and realize we've exceeded the memory
budget, we evict cache entries starting with the least recently used ones
until we have enough memory to add the new tuples to the cache.
For parameterized nested loop joins, we now consider using one of these
result cache nodes in between the nested loop node and its inner node. We
determine when this might be useful based on cost, which is primarily
driven off of what the expected cache hit ratio will be. Estimating the
cache hit ratio relies on having good distinct estimates on the nested
loop's parameters.
For now, the planner will only consider using a result cache for
parameterized nested loop joins. This works for both normal joins and
also for LATERAL type joins to subqueries. It is possible to use this new
node for other uses in the future. For example, to cache results from
correlated subqueries. However, that's not done here due to some
difficulties obtaining a distinct estimation on the outer plan to
calculate the estimated cache hit ratio. Currently we plan the inner plan
before planning the outer plan so there is no good way to know if a result
cache would be useful or not since we can't estimate the number of times
the subplan will be called until the outer plan is generated.
The functionality being added here is newly introducing a dependency on
the return value of estimate_num_groups() during the join search.
Previously, during the join search, we only ever needed to perform
selectivity estimations. With this commit, we need to use
estimate_num_groups() in order to estimate what the hit ratio on the
result cache will be. In simple terms, if we expect 10 distinct values
and we expect 1000 outer rows, then we'll estimate the hit ratio to be
99%. Since cache hits are very cheap compared to scanning the underlying
nodes on the inner side of the nested loop join, then this will
significantly reduce the planner's cost for the join. However, it's
fairly easy to see here that things will go bad when estimate_num_groups()
incorrectly returns a value that's significantly lower than the actual
number of distinct values. If this happens then that may cause us to make
use of a nested loop join with a result cache instead of some other join
type, such as a merge or hash join. Our distinct estimations have been
known to be a source of trouble in the past, so the extra reliance on them
here could cause the planner to choose slower plans than it did previous
to having this feature. Distinct estimations are also fairly hard to
estimate accurately when several tables have been joined already or when a
WHERE clause filters out a set of values that are correlated to the
expressions we're estimating the number of distinct value for.
For now, the costing we perform during query planning for result caches
does put quite a bit of faith in the distinct estimations being accurate.
When these are accurate then we should generally see faster execution
times for plans containing a result cache. However, in the real world, we
may find that we need to either change the costings to put less trust in
the distinct estimations being accurate or perhaps even disable this
feature by default. There's always an element of risk when we teach the
query planner to do new tricks that it decides to use that new trick at
the wrong time and causes a regression. Users may opt to get the old
behavior by turning the feature off using the enable_resultcache GUC.
Currently, this is enabled by default. It remains to be seen if we'll
maintain that setting for the release.
Additionally, the name "Result Cache" is the best name I could think of
for this new node at the time I started writing the patch. Nobody seems
to strongly dislike the name. A few people did suggest other names but no
other name seemed to dominate in the brief discussion that there was about
names. Let's allow the beta period to see if the current name pleases
enough people. If there's some consensus on a better name, then we can
change it before the release. Please see the 2nd discussion link below
for the discussion on the "Result Cache" name.
Author: David Rowley
Reviewed-by: Andy Fan, Justin Pryzby, Zhihong Yu, Hou Zhijie
Tested-By: Konstantin Knizhnik
Discussion: https://postgr.es/m/CAApHDvrPcQyQdWERGYWx8J%2B2DLUNgXu%2BfOSbQ1UscxrunyXyrQ%40mail.gmail.com
Discussion: https://postgr.es/m/CAApHDvq=yQXr5kqhRviT2RhNKwToaWr9JAN5t+5_PzhuRJ3wvg@mail.gmail.com
This removes "Add Result Cache executor node". It seems that something
weird is going on with the tracking of cache hits and misses as
highlighted by many buildfarm animals. It's not yet clear what the
problem is as other parts of the plan indicate that the cache did work
correctly, it's just the hits and misses that were being reported as 0.
This is especially a bad time to have the buildfarm so broken, so
reverting before too many more animals go red.
Discussion: https://postgr.es/m/CAApHDvq_hydhfovm4=izgWs+C5HqEeRScjMbOgbpC-jRAeK3Yw@mail.gmail.com
Here we add a new executor node type named "Result Cache". The planner
can include this node type in the plan to have the executor cache the
results from the inner side of parameterized nested loop joins. This
allows caching of tuples for sets of parameters so that in the event that
the node sees the same parameter values again, it can just return the
cached tuples instead of rescanning the inner side of the join all over
again. Internally, result cache uses a hash table in order to quickly
find tuples that have been previously cached.
For certain data sets, this can significantly improve the performance of
joins. The best cases for using this new node type are for join problems
where a large portion of the tuples from the inner side of the join have
no join partner on the outer side of the join. In such cases, hash join
would have to hash values that are never looked up, thus bloating the hash
table and possibly causing it to multi-batch. Merge joins would have to
skip over all of the unmatched rows. If we use a nested loop join with a
result cache, then we only cache tuples that have at least one join
partner on the outer side of the join. The benefits of using a
parameterized nested loop with a result cache increase when there are
fewer distinct values being looked up and the number of lookups of each
value is large. Also, hash probes to lookup the cache can be much faster
than the hash probe in a hash join as it's common that the result cache's
hash table is much smaller than the hash join's due to result cache only
caching useful tuples rather than all tuples from the inner side of the
join. This variation in hash probe performance is more significant when
the hash join's hash table no longer fits into the CPU's L3 cache, but the
result cache's hash table does. The apparent "random" access of hash
buckets with each hash probe can cause a poor L3 cache hit ratio for large
hash tables. Smaller hash tables generally perform better.
The hash table used for the cache limits itself to not exceeding work_mem
* hash_mem_multiplier in size. We maintain a dlist of keys for this cache
and when we're adding new tuples and realize we've exceeded the memory
budget, we evict cache entries starting with the least recently used ones
until we have enough memory to add the new tuples to the cache.
For parameterized nested loop joins, we now consider using one of these
result cache nodes in between the nested loop node and its inner node. We
determine when this might be useful based on cost, which is primarily
driven off of what the expected cache hit ratio will be. Estimating the
cache hit ratio relies on having good distinct estimates on the nested
loop's parameters.
For now, the planner will only consider using a result cache for
parameterized nested loop joins. This works for both normal joins and
also for LATERAL type joins to subqueries. It is possible to use this new
node for other uses in the future. For example, to cache results from
correlated subqueries. However, that's not done here due to some
difficulties obtaining a distinct estimation on the outer plan to
calculate the estimated cache hit ratio. Currently we plan the inner plan
before planning the outer plan so there is no good way to know if a result
cache would be useful or not since we can't estimate the number of times
the subplan will be called until the outer plan is generated.
The functionality being added here is newly introducing a dependency on
the return value of estimate_num_groups() during the join search.
Previously, during the join search, we only ever needed to perform
selectivity estimations. With this commit, we need to use
estimate_num_groups() in order to estimate what the hit ratio on the
result cache will be. In simple terms, if we expect 10 distinct values
and we expect 1000 outer rows, then we'll estimate the hit ratio to be
99%. Since cache hits are very cheap compared to scanning the underlying
nodes on the inner side of the nested loop join, then this will
significantly reduce the planner's cost for the join. However, it's
fairly easy to see here that things will go bad when estimate_num_groups()
incorrectly returns a value that's significantly lower than the actual
number of distinct values. If this happens then that may cause us to make
use of a nested loop join with a result cache instead of some other join
type, such as a merge or hash join. Our distinct estimations have been
known to be a source of trouble in the past, so the extra reliance on them
here could cause the planner to choose slower plans than it did previous
to having this feature. Distinct estimations are also fairly hard to
estimate accurately when several tables have been joined already or when a
WHERE clause filters out a set of values that are correlated to the
expressions we're estimating the number of distinct value for.
For now, the costing we perform during query planning for result caches
does put quite a bit of faith in the distinct estimations being accurate.
When these are accurate then we should generally see faster execution
times for plans containing a result cache. However, in the real world, we
may find that we need to either change the costings to put less trust in
the distinct estimations being accurate or perhaps even disable this
feature by default. There's always an element of risk when we teach the
query planner to do new tricks that it decides to use that new trick at
the wrong time and causes a regression. Users may opt to get the old
behavior by turning the feature off using the enable_resultcache GUC.
Currently, this is enabled by default. It remains to be seen if we'll
maintain that setting for the release.
Additionally, the name "Result Cache" is the best name I could think of
for this new node at the time I started writing the patch. Nobody seems
to strongly dislike the name. A few people did suggest other names but no
other name seemed to dominate in the brief discussion that there was about
names. Let's allow the beta period to see if the current name pleases
enough people. If there's some consensus on a better name, then we can
change it before the release. Please see the 2nd discussion link below
for the discussion on the "Result Cache" name.
Author: David Rowley
Reviewed-by: Andy Fan, Justin Pryzby, Zhihong Yu
Tested-By: Konstantin Knizhnik
Discussion: https://postgr.es/m/CAApHDvrPcQyQdWERGYWx8J%2B2DLUNgXu%2BfOSbQ1UscxrunyXyrQ%40mail.gmail.com
Discussion: https://postgr.es/m/CAApHDvq=yQXr5kqhRviT2RhNKwToaWr9JAN5t+5_PzhuRJ3wvg@mail.gmail.com
This patch makes two closely related sets of changes:
1. For UPDATE, the subplan of the ModifyTable node now only delivers
the new values of the changed columns (i.e., the expressions computed
in the query's SET clause) plus row identity information such as CTID.
ModifyTable must re-fetch the original tuple to merge in the old
values of any unchanged columns. The core advantage of this is that
the changed columns are uniform across all tables of an inherited or
partitioned target relation, whereas the other columns might not be.
A secondary advantage, when the UPDATE involves joins, is that less
data needs to pass through the plan tree. The disadvantage of course
is an extra fetch of each tuple to be updated. However, that seems to
be very nearly free in context; even worst-case tests don't show it to
add more than a couple percent to the total query cost. At some point
it might be interesting to combine the re-fetch with the tuple access
that ModifyTable must do anyway to mark the old tuple dead; but that
would require a good deal of refactoring and it seems it wouldn't buy
all that much, so this patch doesn't attempt it.
2. For inherited UPDATE/DELETE, instead of generating a separate
subplan for each target relation, we now generate a single subplan
that is just exactly like a SELECT's plan, then stick ModifyTable
on top of that. To let ModifyTable know which target relation a
given incoming row refers to, a tableoid junk column is added to
the row identity information. This gets rid of the horrid hack
that was inheritance_planner(), eliminating O(N^2) planning cost
and memory consumption in cases where there were many unprunable
target relations.
Point 2 of course requires point 1, so that there is a uniform
definition of the non-junk columns to be returned by the subplan.
We can't insist on uniform definition of the row identity junk
columns however, if we want to keep the ability to have both
plain and foreign tables in a partitioning hierarchy. Since
it wouldn't scale very far to have every child table have its
own row identity column, this patch includes provisions to merge
similar row identity columns into one column of the subplan result.
In particular, we can merge the whole-row Vars typically used as
row identity by FDWs into one column by pretending they are type
RECORD. (It's still okay for the actual composite Datums to be
labeled with the table's rowtype OID, though.)
There is more that can be done to file down residual inefficiencies
in this patch, but it seems to be committable now.
FDW authors should note several API changes:
* The argument list for AddForeignUpdateTargets() has changed, and so
has the method it must use for adding junk columns to the query. Call
add_row_identity_var() instead of manipulating the parse tree directly.
You might want to reconsider exactly what you're adding, too.
* PlanDirectModify() must now work a little harder to find the
ForeignScan plan node; if the foreign table is part of a partitioning
hierarchy then the ForeignScan might not be the direct child of
ModifyTable. See postgres_fdw for sample code.
* To check whether a relation is a target relation, it's no
longer sufficient to compare its relid to root->parse->resultRelation.
Instead, check it against all_result_relids or leaf_result_relids,
as appropriate.
Amit Langote and Tom Lane
Discussion: https://postgr.es/m/CA+HiwqHpHdqdDn48yCEhynnniahH78rwcrv1rEX65-fsZGBOLQ@mail.gmail.com
Here we aim to reduce duplicate work done by contain_volatile_functions()
by caching whether PathTargets and RestrictInfos contain any volatile
functions the first time contain_volatile_functions() is called for them.
Any future calls for these nodes just use the cached value rather than
going to the trouble of recursively checking the sub-node all over again.
Thanks to Tom Lane for the idea.
Any locations in the code which make changes to a PathTarget or
RestrictInfo which could change the outcome of the volatility check must
change the cached value back to VOLATILITY_UNKNOWN again.
contain_volatile_functions() is the only code in charge of setting the
cache value to either VOLATILITY_VOLATILE or VOLATILITY_NOVOLATILE.
Some existing code does benefit from this additional caching, however,
this change is mainly aimed at an upcoming patch that must check for
volatility during the join search. Repeated volatility checks in that
case can become very expensive when the join search contains more than a
few relations.
Author: David Rowley
Discussion: https://postgr.es/m/3795226.1614059027@sss.pgh.pa.us
d2d8a229bc introduced a new function generate_useful_gather_paths to
be used as a replacement for generate_gather_paths, but forgot to update
a couple of places that referenced the older function.
This is possibly not 100% complete (ref. create_ordered_paths), but it's
better than not changing anything.
Author: "Hou, Zhijie" <houzj.fnst@cn.fujitsu.com>
Reviewed-by: Tomas Vondra <tomas.vondra@enterprisedb.com>
Discussion: https://postgr.es/m/4ce1d5116fe746a699a6d29858c6a39a@G08CNEXMBPEKD05.g08.fujitsu.local
It turns out that the calculation of [Merge]AppendPath.partitioned_rels
in allpaths.c is faulty and sometimes omits relevant non-leaf partitions,
allowing an assertion added by commit a929e17e5a to trigger. Rather
than fix that, it seems better to get rid of those fields altogether.
We don't really need the info until create_plan time, and calculating
it once for the selected plan should be cheaper than calculating it
for each append path we consider.
The preceding two commits did away with all use of the partitioned_rels
values; this commit just mechanically removes the fields and the code
that calculated them.
Discussion: https://postgr.es/m/87sg8tqhsl.fsf@aurora.ydns.eu
Discussion: https://postgr.es/m/CAJKUy5gCXDSmFs2c=R+VGgn7FiYcLCsEFEuDNNLGfoha=pBE_g@mail.gmail.com
Commit ebb7ae839d ensured we ignore pathkeys with volatile expressions
when considering adding a sort below a Gather Merge. Turns out we need
to care about parallel safety of the pathkeys too, otherwise we might
try sorting e.g. on results of a correlated subquery (as demonstrated
by a report from Luis Roberto).
Initial investigation by Tom Lane, patch by James Coleman. Backpatch
to 13, where the code was instroduced (as part of Incremental Sort).
Reported-by: Luis Roberto
Author: James Coleman
Reviewed-by: Tomas Vondra
Backpatch-through: 13
Discussion: https://postgr.es/m/622580997.37108180.1604080457319.JavaMail.zimbra%40siscobra.com.br
Discussion: https://postgr.es/m/CAAaqYe8cK3g5CfLC4w7bs=hC0mSksZC=H5M8LSchj5e5OxpTAg@mail.gmail.com
generate_useful_gather_paths used to skip unsorted paths (without any
pathkeys), but that is unnecessary - the later code actually can handle
such paths just fine by adding a Sort node. This is clearly a thinko,
preventing construction of useful plans.
Backpatch to 13, where Incremental Sort was introduced.
Author: James Coleman
Reviewed-by: Tomas Vondra
Backpatch-through: 13
Discussion: https://postgr.es/m/CAAaqYe8cK3g5CfLC4w7bs=hC0mSksZC=H5M8LSchj5e5OxpTAg@mail.gmail.com
When considering Incremental Sort below a Gather Merge, we need to be
a bit more careful when matching pathkeys to EC members. It's not enough
to find a member whose Vars are all in the current relation's target;
volatile expressions in particular need to be contained in the target,
otherwise it's too early to use the pathkey.
Reported-by: Jaime Casanova
Author: James Coleman
Reviewed-by: Tomas Vondra
Backpatch-through: 13, where the incremental sort code was added
Discussion: https://postgr.es/m/CAJGNTeNaxpXgBVcRhJX%2B2vSbq%2BF2kJqGBcvompmpvXb7pq%2BoFA%40mail.gmail.com
Previously we only tagged on the required information to allow the
executor to perform run-time partition pruning for Append/MergeAppend
nodes belonging to base relations. It was thought that nested
Append/MergeAppend nodes were just about always pulled up into the
top-level Append/MergeAppend and that making the run-time pruning info for
any sub Append/MergeAppend nodes was a waste of time. However, that was
likely badly thought through.
Some examples of cases we're unable to pullup nested Append/MergeAppends
are: 1) Parallel Append nodes with a mix of parallel and non-parallel
paths into a Parallel Append. 2) When planning an ordered Append scan a
sub-partition which is unordered may require a nested MergeAppend path to
ensure sub-partitions don't mix up the order of tuples being fed into the
top-level Append.
Unfortunately, it was not just as simple as removing the lines in
createplan.c which were purposefully not building the run-time pruning
info for anything but RELOPT_BASEREL relations. The code in
add_paths_to_append_rel() was far too sloppy about which partitioned_rels
it included for the Append/MergeAppend paths. The original code there
would always assume accumulate_append_subpath() would pull each sub-Append
and sub-MergeAppend path into the top-level path. While it does not
appear that there were any actual bugs caused by having the additional
partitioned table RT indexes recorded, what it did mean is that later in
planning, when we built the run-time pruning info that we wasted effort
and built PartitionedRelPruneInfos for partitioned tables that we had no
subpaths for the executor to run-time prune.
Here we tighten that up so that partitioned_rels only ever contains the RT
index for partitioned tables which actually have subpaths in the given
Append/MergeAppend. We can now Assert that every PartitionedRelPruneInfo
has a non-empty present_parts. That should allow us to catch any weird
corner cases that have been missed.
In passing, it seems there is no longer a good reason to have the
AppendPath and MergeAppendPath's partitioned_rel fields a List of IntList.
We can simply have a List of Relids instead. This is more compact in
memory and faster to add new members to. We still know which is the root
level partition as these always have a lower relid than their children.
Previously this field was used for more things, but run-time partition
pruning now remains the only user of it and it has no need for a List of
IntLists.
Here we also get rid of the RelOptInfo partitioned_child_rels field. This
is what was previously used to (sometimes incorrectly) set the
Append/MergeAppend path's partitioned_rels field. That was the only usage
of that field, so we can happily just remove it.
I also couldn't resist changing some nearby code to make use of the newly
added for_each_from macro so we can skip the first element in the list
without checking if the current item was the first one on each
iteration.
A bug report from Andreas Kretschmer prompted all this work, however,
after some consideration, I'm not personally classing this as a bug fix.
So no backpatch. In Andreas' test case, it just wasn't that clear that
there was a nested Append since the top-level Append just had a single
sub-path which was pulled up a level, per 8edd0e794.
Author: David Rowley
Reviewed-by: Amit Langote
Discussion: https://postgr.es/m/flat/CAApHDvqSchs%2BubdybcfFaSPB%2B%2BEA7kqMaoqajtP0GtZvzOOR3g%40mail.gmail.com
Historically, we've considered the state with relpages and reltuples
both zero as indicating that we do not know the table's tuple density.
This is problematic because it's impossible to distinguish "never yet
vacuumed" from "vacuumed and seen to be empty". In particular, a user
cannot use VACUUM or ANALYZE to override the planner's normal heuristic
that an empty table should not be believed to be empty because it is
probably about to get populated. That heuristic is a good safety
measure, so I don't care to abandon it, but there should be a way to
override it if the table is indeed intended to stay empty.
Hence, represent the initial state of ignorance by setting reltuples
to -1 (relpages is still set to zero), and apply the minimum-ten-pages
heuristic only when reltuples is still -1. If the table is empty,
VACUUM or ANALYZE (but not CREATE INDEX) will override that to
reltuples = relpages = 0, and then we'll plan on that basis.
This requires a bunch of fiddly little changes, but we can get rid of
some ugly kluges that were formerly needed to maintain the old definition.
One notable point is that FDWs' GetForeignRelSize methods will see
baserel->tuples = -1 when no ANALYZE has been done on the foreign table.
That seems like a net improvement, since those methods were formerly
also in the dark about what baserel->tuples = 0 really meant. Still,
it is an API change.
I bumped catversion because code predating this change would get confused
by seeing reltuples = -1.
Discussion: https://postgr.es/m/F02298E0-6EF4-49A1-BCB6-C484794D9ACC@thebuild.com
The trouble with doing this is that an apparently-constant subquery
output column isn't really constant if it is a grouping column that
appears in only some of the grouping sets. A qual using such a
column would be subject to incorrect const-folding after push-down,
as seen in bug #16585 from Paul Sivash.
To fix, just disable qual pushdown altogether if the sub-query has
nonempty groupingSets. While we could imagine far less restrictive
solutions, there is not much point in working harder right now,
because subquery_planner() won't move HAVING clauses to WHERE within
such a subquery. If the qual stays in HAVING it's not going to be
a lot more useful than if we'd kept it at the outer level.
Having said that, this restriction could be removed if we used a
parsetree representation that distinguished such outputs from actual
constants, which is something I hope to do in future. Hence, make
the patch a minimal addition rather than integrating it more tightly
(e.g. by renumbering the existing items in subquery_is_pushdown_safe's
comment).
Back-patch to 9.5 where grouping sets were introduced.
Discussion: https://postgr.es/m/16585-9d8c340d23ade8c1@postgresql.org
The qual pushdown logic assumed that all Vars in a restriction clause
must be Vars referencing subquery outputs; but since we introduced
LATERAL, it's possible for such a Var to be a lateral reference instead.
This led to an assertion failure in debug builds. In a non-debug
build, there might be no ill effects (if qual_is_pushdown_safe decided
the qual was unsafe anyway), or we could get failures later due to
construction of an invalid plan. I've not gone to much length to
characterize the possible failures, but at least segfaults in the
executor have been observed.
Given that this has been busted since 9.3 and it took this long for
anybody to notice, I judge that the case isn't worth going to great
lengths to optimize. Hence, fix by just teaching qual_is_pushdown_safe
that such quals are unsafe to push down, matching the previous behavior
when it accidentally didn't fail.
Per report from Tom Ellis. Back-patch to all supported branches.
Discussion: https://postgr.es/m/20200713175124.GQ8220@cloudinit-builder
After running GetForeignRelSize for a foreign table, adjust rel->tuples
to be at least as large as rel->rows. This prevents bizarre behavior
in estimate_num_groups() and perhaps other places, especially in the
scenario where rel->tuples is zero because pg_class.reltuples is
(suggesting that ANALYZE has never been run for the table). As things
stood, we'd end up estimating one group out of any GROUP BY on such a
table, whereas the default group-count estimate is more likely to result
in a sane plan.
Also, clarify in the documentation that GetForeignRelSize has the option
to override the rel->tuples value if it has a better idea of what to use
than what is in pg_class.reltuples.
Per report from Jeff Janes. Back-patch to all supported branches.
Patch by me; thanks to Etsuro Fujita for review
Discussion: https://postgr.es/m/CAMkU=1xNo9cnan+Npxgz0eK7394xmjmKg-QEm8wYG9P5-CcaqQ@mail.gmail.com
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.
Commit d2d8a229bc introduced Incremental Sort, but it was considered
only in create_ordered_paths() as an alternative to regular Sort. There
are many other places that require sorted input and might benefit from
considering Incremental Sort too.
This patch modifies a number of those places, but not all. The concern
is that just adding Incremental Sort to any place that already adds
Sort may increase the number of paths considered, negatively affecting
planning time, without any benefit. So we've taken a more conservative
approach, based on analysis of which places do affect a set of queries
that did seem practical. This means some less common queries may not
benefit from Incremental Sort yet.
Author: Tomas Vondra
Reviewed-by: James Coleman
Discussion: https://postgr.es/m/CAPpHfds1waRZ=NOmueYq0sx1ZSCnt+5QJvizT8ndT2=etZEeAQ@mail.gmail.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
David Rowley
Tom Lane
Bruce Momjian
Thomas Munro
Tomas Vondra
Peter Eisentraut
Alvaro Herrera
Michael Paquier
Amit Kapila