Our policy since commit ab77a5a45 has been that a plan node having
any initplans is automatically not parallel-safe. (This could be
relaxed, but not today.) clean_up_removed_plan_level neglected
this, and could attach initplans to a parallel-safe child plan
node without clearing the plan's parallel-safe flag. That could
lead to "subplan was not initialized" errors at runtime, in case
an initplan referenced another one and only the referencing one
got transmitted to parallel workers.
The fix in clean_up_removed_plan_level is trivial enough.
materialize_finished_plan also moves initplans from one node
to another, but it's okay because it already copies the source
node's parallel_safe flag. The other place that does this kind
of thing is standard_planner's hack to inject a top-level Gather
when debug_parallel_query is active. But that's actually dead
code given that we're correctly enforcing the "initplans aren't
parallel safe" rule, so just replace it with an Assert that
there are no initplans.
Also improve some related comments.
Normally we'd add a regression test case for this sort of bug.
The mistake itself is already reached by existing tests, but there
is accidentally no visible problem. The only known test case that
creates an actual failure seems too indirect and fragile to justify
keeping it as a regression test (not least because it fails to fail
in v11, though the bug is clearly present there too).
Per report from Justin Pryzby. Back-patch to all supported branches.
Discussion: https://postgr.es/m/ZDVt6MaNWkRDO1LQ@telsasoft.com
create_append_path() would only apply get_baserel_parampathinfo
when the path is for a partitioned table, but it's also potentially
useful for paths for UNION ALL appendrels. Specifically, that
supports building a Memoize path atop this one.
While we're in the vicinity, delete some dead code in
create_merge_append_plan(): there's no need for it to support
parameterized MergeAppend paths, and it doesn't look like that
is going to change anytime soon. It'll be easy enough to undo
this when/if it becomes useful.
Richard Guo
Discussion: https://postgr.es/m/CAMbWs4_ABSu4PWG2rE1q10tJugEXHWgru3U8dAgkoFvgrb6aEA@mail.gmail.com
Here remove some dead code from heapgettup() and heapgettup_pagemode()
which was trying to support NoMovementScanDirection scans. This code can
never be reached as standard_ExecutorRun() never calls ExecutePlan with
NoMovementScanDirection.
Additionally, plans which were scanning an unordered index would use
NoMovementScanDirection rather than ForwardScanDirection. There was no
real need for this, so here we adjust this so we use ForwardScanDirection
for unordered index scans. A comment in pathnodes.h claimed that
NoMovementScanDirection was used for PathKey reasons, but if that was
true, it no longer is, per code in build_index_paths().
This does change the non-text format of the EXPLAIN output so that
unordered index scans now have a "Forward" scan direction rather than
"NoMovement". The text format of EXPLAIN has not changed.
Author: Melanie Plageman
Reviewed-by: Tom Lane, David Rowley
Discussion: https://postgr.es/m/CAAKRu_bvkhka0CZQun28KTqhuUh5ZqY=_T8QEqZqOL02rpi2bw@mail.gmail.com
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
A couple of places weren't up to speed for this. By sheer good
luck, we didn't fail but just selected a non-memoized join plan,
at least in the test case we have. Nonetheless, it's a bug,
and I'm not quite sure that it couldn't have worse consequences
in other examples. So back-patch to v14 where Memoize came in.
Richard Guo
Discussion: https://postgr.es/m/CAMbWs48GkNom272sfp0-WeD6_0HSR19BJ4H1c9ZKSfbVnJsvRg@mail.gmail.com
It neglected to recurse to the subpath, meaning you'd get back
a path identical to the input. This could produce wrong query
results if the omission meant that the subpath fails to enforce
some join clause it should be enforcing. We don't have a test
case for this at the moment, but the code is obviously broken
and the fix is equally obvious. Back-patch to v14 where
Memoize was introduced.
Richard Guo
Discussion: https://postgr.es/m/CAMbWs4_R=ORpz=Lkn2q3ebPC5EuWyfZF+tmfCPVLBVK5W39mHA@mail.gmail.com
These two functions failed to cover MaterialPath. That's not a
fatal problem, but we can generate better plans in some cases
if we support it.
Tom Lane and Richard Guo
Discussion: https://postgr.es/m/1854233.1669949723@sss.pgh.pa.us
This reverts commit db0d67db24 and
several follow-on fixes. The idea of making a cost-based choice
of the order of the sorting columns is not fundamentally unsound,
but it requires cost information and data statistics that we don't
really have. For example, relying on procost to distinguish the
relative costs of different sort comparators is pretty pointless
so long as most such comparator functions are labeled with cost 1.0.
Moreover, estimating the number of comparisons done by Quicksort
requires more than just an estimate of the number of distinct values
in the input: you also need some idea of the sizes of the larger
groups, if you want an estimate that's good to better than a factor of
three or so. That's data that's often unknown or not very reliable.
Worse, to arrive at estimates of the number of calls made to the
lower-order-column comparison functions, the code needs to make
estimates of the numbers of distinct values of multiple columns,
which are necessarily even less trustworthy than per-column stats.
Even if all the inputs are perfectly reliable, the cost algorithm
as-implemented cannot offer useful information about how to order
sorting columns beyond the point at which the average group size
is estimated to drop to 1.
Close inspection of the code added by db0d67db2 shows that there
are also multiple small bugs. These could have been fixed, but
there's not much point if we don't trust the estimates to be
accurate in-principle.
Finally, the changes in cost_sort's behavior made for very large
changes (often a factor of 2 or so) in the cost estimates for all
sorting operations, not only those for multi-column GROUP BY.
That naturally changes plan choices in many situations, and there's
precious little evidence to show that the changes are for the better.
Given the above doubts about whether the new estimates are really
trustworthy, it's hard to summon much confidence that these changes
are better on the average.
Since we're hard up against the release deadline for v15, let's
revert these changes for now. We can always try again later.
Note: in v15, I left T_PathKeyInfo in place in nodes.h even though
it's unreferenced. Removing it would be an ABI break, and it seems
a bit late in the release cycle for that.
Discussion: https://postgr.es/m/TYAPR01MB586665EB5FB2C3807E893941F5579@TYAPR01MB5866.jpnprd01.prod.outlook.com
The present implementations of adjust_appendrel_attrs_multilevel and
its sibling adjust_child_relids_multilevel are very messy, because
they work by reconstructing the relids of the child's immediate
parent and then seeing if that's bms_equal to the relids of the
target parent. Aside from being quite inefficient, this will not
work with planned future changes to make joinrels' relid sets
contain outer-join relids in addition to baserels.
The whole thing can be solved at a stroke by adding explicit parent
and top_parent links to child RelOptInfos, and making these functions
work with RelOptInfo pointers instead of relids. Doing that is
simpler for most callers, too.
In my original version of this patch, I got rid of
RelOptInfo.top_parent_relids on the grounds that it was now redundant.
However, that adds a lot of code churn in places that otherwise would
not need changing, and arguably the extra indirection needed to fetch
top_parent->relids in those places costs something. So this version
leaves that field in place.
Discussion: https://postgr.es/m/553080.1657481916@sss.pgh.pa.us
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
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 evaluating a query with a multi-column GROUP BY clause using sort,
the cost may be heavily dependent on the order in which the keys are
compared when building the groups. Grouping does not imply any ordering,
so we're allowed to compare the keys in arbitrary order, and a Hash Agg
leverages this. But for Group Agg, we simply compared keys in the order
as specified in the query. This commit explores alternative ordering of
the keys, trying to find a cheaper one.
In principle, we might generate grouping paths for all permutations of
the keys, and leave the rest to the optimizer. But that might get very
expensive, so we try to pick only a couple interesting orderings based
on both local and global information.
When planning the grouping path, we explore statistics (number of
distinct values, cost of the comparison function) for the keys and
reorder them to minimize comparison costs. Intuitively, it may be better
to perform more expensive comparisons (for complex data types etc.)
last, because maybe the cheaper comparisons will be enough. Similarly,
the higher the cardinality of a key, the lower the probability we’ll
need to compare more keys. The patch generates and costs various
orderings, picking the cheapest ones.
The ordering of group keys may interact with other parts of the query,
some of which may not be known while planning the grouping. E.g. there
may be an explicit ORDER BY clause, or some other ordering-dependent
operation, higher up in the query, and using the same ordering may allow
using either incremental sort or even eliminate the sort entirely.
The patch generates orderings and picks those minimizing the comparison
cost (for various pathkeys), and then adds orderings that might be
useful for operations higher up in the plan (ORDER BY, etc.). Finally,
it always keeps the ordering specified in the query, on the assumption
the user might have additional insights.
This introduces a new GUC enable_group_by_reordering, so that the
optimization may be disabled if needed.
The original patch was proposed by Teodor Sigaev, and later improved and
reworked by Dmitry Dolgov. Reviews by a number of people, including me,
Andrey Lepikhov, Claudio Freire, Ibrar Ahmed and Zhihong Yu.
Author: Dmitry Dolgov, Teodor Sigaev, Tomas Vondra
Reviewed-by: Tomas Vondra, Andrey Lepikhov, Claudio Freire, Ibrar Ahmed, Zhihong Yu
Discussion: https://postgr.es/m/7c79e6a5-8597-74e8-0671-1c39d124c9d6%40sigaev.ru
Discussion: https://postgr.es/m/CA%2Bq6zcW_4o2NC0zutLkOJPsFt80megSpX_dVRo6GK9PC-Jx_Ag%40mail.gmail.com
MERGE performs actions that modify rows in the target table using a
source table or query. MERGE provides a single SQL statement that can
conditionally INSERT/UPDATE/DELETE rows -- a task that would otherwise
require multiple PL statements. For example,
MERGE INTO target AS t
USING source AS s
ON t.tid = s.sid
WHEN MATCHED AND t.balance > s.delta THEN
UPDATE SET balance = t.balance - s.delta
WHEN MATCHED THEN
DELETE
WHEN NOT MATCHED AND s.delta > 0 THEN
INSERT VALUES (s.sid, s.delta)
WHEN NOT MATCHED THEN
DO NOTHING;
MERGE works with regular tables, partitioned tables and inheritance
hierarchies, including column and row security enforcement, as well as
support for row and statement triggers and transition tables therein.
MERGE is optimized for OLTP and is parameterizable, though also useful
for large scale ETL/ELT. MERGE is not intended to be used in preference
to existing single SQL commands for INSERT, UPDATE or DELETE since there
is some overhead. MERGE can be used from PL/pgSQL.
MERGE does not support targetting updatable views or foreign tables, and
RETURNING clauses are not allowed either. These limitations are likely
fixable with sufficient effort. Rewrite rules are also not supported,
but it's not clear that we'd want to support them.
Author: Pavan Deolasee <pavan.deolasee@gmail.com>
Author: Álvaro Herrera <alvherre@alvh.no-ip.org>
Author: Amit Langote <amitlangote09@gmail.com>
Author: Simon Riggs <simon.riggs@enterprisedb.com>
Reviewed-by: Peter Eisentraut <peter.eisentraut@enterprisedb.com>
Reviewed-by: Andres Freund <andres@anarazel.de> (earlier versions)
Reviewed-by: Peter Geoghegan <pg@bowt.ie> (earlier versions)
Reviewed-by: Robert Haas <robertmhaas@gmail.com> (earlier versions)
Reviewed-by: Japin Li <japinli@hotmail.com>
Reviewed-by: Justin Pryzby <pryzby@telsasoft.com>
Reviewed-by: Tomas Vondra <tomas.vondra@enterprisedb.com>
Reviewed-by: Zhihong Yu <zyu@yugabyte.com>
Discussion: https://postgr.es/m/CANP8+jKitBSrB7oTgT9CY2i1ObfOt36z0XMraQc+Xrz8QB0nXA@mail.gmail.com
Discussion: https://postgr.es/m/CAH2-WzkJdBuxj9PO=2QaO9-3h3xGbQPZ34kJH=HukRekwM-GZg@mail.gmail.com
Discussion: https://postgr.es/m/20201231134736.GA25392@alvherre.pgsql
Memoize would always use the hash equality operator for the cache key
types to determine if the current set of parameters were the same as some
previously cached set. Certain types such as floating points where -0.0
and +0.0 differ in their binary representation but are classed as equal by
the hash equality operator may cause problems as unless the join uses the
same operator it's possible that whichever join operator is being used
would be able to distinguish the two values. In which case we may
accidentally return in the incorrect rows out of the cache.
To fix this here we add a binary mode to Memoize to allow it to the
current set of parameters to previously cached values by comparing
bit-by-bit rather than logically using the hash equality operator. This
binary mode is always used for LATERAL joins and it's used for normal
joins when any of the join operators are not hashable.
Reported-by: Tom Lane
Author: David Rowley
Discussion: https://postgr.es/m/3004308.1632952496@sss.pgh.pa.us
Backpatch-through: 14, where Memoize was added
Several mistakes have piled in the code comments over the time,
including incorrect grammar, function names and simple typos. This
commit takes care of a portion of these.
No backpatch is done as this is only cosmetic.
Author: Justin Pryzby
Discussion: https://postgr.es/m/20210924215827.GS831@telsasoft.com
All the code paths simplified here were already using a boolean or used
an expression that led to zero or one, making the extra bits
unnecessary.
Author: Justin Pryzby
Reviewed-by: Tom Lane, Michael Paquier, Peter Smith
Discussion: https://postgr.es/m/20210428182936.GE27406@telsasoft.com
The point of introducing the hash_mem_multiplier GUC was to let users
reproduce the old behavior of hash aggregation, i.e. that it could use
more than work_mem at need. However, the implementation failed to get
the job done on Win64, where work_mem is clamped to 2GB to protect
various places that calculate memory sizes using "long int". As
written, the same clamp was applied to hash_mem. This resulted in
severe performance regressions for queries requiring a bit more than
2GB for hash aggregation, as they now spill to disk and there's no
way to stop that.
Getting rid of the work_mem restriction seems like a good idea, but
it's a big job and could not conceivably be back-patched. However,
there's only a fairly small number of places that are concerned with
the hash_mem value, and it turns out to be possible to remove the
restriction there without too much code churn or any ABI breaks.
So, let's do that for now to fix the regression, and leave the
larger task for another day.
This patch does introduce a bit more infrastructure that should help
with the larger task, namely pg_bitutils.h support for working with
size_t values.
Per gripe from Laurent Hasson. Back-patch to v13 where the
behavior change came in.
Discussion: https://postgr.es/m/997817.1627074924@sss.pgh.pa.us
Discussion: https://postgr.es/m/MN2PR15MB25601E80A9B6D1BA6F592B1985E39@MN2PR15MB2560.namprd15.prod.outlook.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
create_projection_plan contains a hidden assumption (here made
explicit by an Assert) that a projection-capable Path will yield a
projection-capable Plan. Unfortunately, that assumption is violated
only a few lines away, by create_projection_plan itself. This means
that two stacked ProjectionPaths can yield an outcome where we try to
jam the upper path's tlist into a non-projection-capable child node,
resulting in an invalid plan.
There isn't any good reason to have stacked ProjectionPaths; indeed the
whole concept is faulty, since the set of Vars/Aggs/etc needed by the
upper one wouldn't necessarily be available in the output of the lower
one, nor could the lower one create such values if they weren't
available from its input. Hence, we can fix this by adjusting
create_projection_path to strip any top-level ProjectionPath from the
subpath it's given. (This amounts to saying "oh, we changed our
minds about what we need to project here".)
The test case added here only fails in v13 and HEAD; before that, we
don't attempt to shove the Sort into the parallel part of the plan,
for reasons that aren't entirely clear to me. However, all the
directly-related code looks generally the same as far back as v11,
where the hazard was introduced (by d7c19e62a). So I've got no faith
that the same type of bug doesn't exist in v11 and v12, given the
right test case. Hence, back-patch the code changes, but not the
irrelevant test case, into those branches.
Per report from Bas Poot.
Discussion: https://postgr.es/m/534fca83789c4a378c7de379e9067d4f@politie.nl
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 add a new output parameter to estimate_num_groups() to allow it to
inform the caller of additional, possibly useful information about the
estimation.
The new output parameter is a struct that currently contains just a single
field with a set of flags. This was done rather than having the flags as
an output parameter to allow future fields to be added without having to
change the signature of the function at a later date when we want to pass
back further information that might not be suitable to store in the flags
field.
It seems reasonable that one day in the future that the planner would want
to know more about the estimation. For example, how many individual sets
of statistics was the estimation generated from? The planner may want to
take that into account if we ever want to consider risks as well as costs
when generating plans.
For now, there's only 1 flag we set in the flags field. This is to
indicate if the estimation fell back on using the hard-coded constants in
any part of the estimation. Callers may like to change their behavior if
this is set, and this gives them the ability to do so. Callers may pass
the flag pointer as NULL if they have no interest in obtaining any
additional information about the estimate.
We're not adding any actual usages of these flags here. Some follow-up
commits will make use of this feature. Additionally, we're also not
making any changes to add support for clauselist_selectivity() and
clauselist_selectivity_ext(). However, if this is required in the future
then the same struct being added here should be fine to use as a new
output argument for those functions too.
Author: David Rowley
Discussion: https://postgr.es/m/CAApHDvqQqpk=1W-G_ds7A9CsXX3BggWj_7okinzkLVhDubQzjA@mail.gmail.com
This adds a new executor node named TID Range Scan. The query planner
will generate paths for TID Range scans when quals are discovered on base
relations which search for ranges on the table's ctid column. These
ranges may be open at either end. For example, WHERE ctid >= '(10,0)';
will return all tuples on page 10 and over.
To support this, two new optional callback functions have been added to
table AM. scan_set_tidrange is used to set the scan range to just the
given range of TIDs. scan_getnextslot_tidrange fetches the next tuple
in the given range.
For AMs were scanning ranges of TIDs would not make sense, these functions
can be set to NULL in the TableAmRoutine. The query planner won't
generate TID Range Scan Paths in that case.
Author: Edmund Horner, David Rowley
Reviewed-by: David Rowley, Tomas Vondra, Tom Lane, Andres Freund, Zhihong Yu
Discussion: https://postgr.es/m/CAMyN-kB-nFTkF=VA_JPwFNo08S0d-Yk0F741S2B7LDmYAi8eyA@mail.gmail.com
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
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.
This patch undoes a couple of very minor uses of the partitioned_rels
values.
createplan.c was testing for nil-ness to optimize away the preparatory
work for make_partition_pruneinfo(). That is worth doing if the check
is nigh free, but it's not worth going to any great lengths to avoid.
create_append_path() was testing for nil-ness as part of deciding how
to set up ParamPathInfo for an AppendPath. I replaced that with a
check for the appendrel's parent rel being partitioned. That's not
quite the same thing but should cover most cases. If we note any
interesting loss of optimizations, we can dumb this down to just
always use the more expensive method when the parent is a baserel.
Discussion: https://postgr.es/m/87sg8tqhsl.fsf@aurora.ydns.eu
Discussion: https://postgr.es/m/CAJKUy5gCXDSmFs2c=R+VGgn7FiYcLCsEFEuDNNLGfoha=pBE_g@mail.gmail.com
For debugging purposes, Path nodes are supposed to have outfuncs
support, but this was overlooked in the original incremental sort patch.
While at it, clean up a couple other minor oversights, as well as
bizarre choice of return type for create_incremental_sort_path().
(All the existing callers just cast it to "Path *" immediately, so
they don't care, but some future caller might care.)
outfuncs.c fix by Zhijie Hou, the rest by me
Discussion: https://postgr.es/m/324c4d81d8134117972a5b1f6cdf9560@G08CNEXMBPEKD05.g08.fujitsu.local
In the past, we always estimated that a ModifyTable node would emit the
same number of rows as its subpaths. Without a RETURNING clause, the
correct estimate is zero. Fix, in preparation for a proposed parallel
write patch that is sensitive to that number.
A remaining problem is that for RETURNING queries, the estimated width
is based on subpath output rather than the RETURNING tlist.
Reviewed-by: Greg Nancarrow <gregn4422@gmail.com>
Discussion: https://postgr.es/m/CAJcOf-cXnB5cnMKqWEp2E2z7Mvcd04iLVmV%3DqpFJrR3AcrTS3g%40mail.gmail.com
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.
reparameterize_path_by_child() failed to reparameterize BitmapAnd
and BitmapOr paths. This matters only if such a path is chosen as
the inside of a nestloop partition-wise join, where we have to pass
in parameters from the outside of the nestloop. If that did happen,
we generated a bad plan that would likely lead to crashes at execution.
This is not entirely reparameterize_path_by_child()'s fault though;
it's the victim of an ancient decision (my ancient decision, I think)
to not bother filling in param_info in BitmapAnd/Or path nodes. That
caused the function to believe that such nodes and their children
contain no parameter references and so need not be processed.
In hindsight that decision looks pretty penny-wise and pound-foolish:
while it saves a few cycles during path node setup, we do commonly
need the information later. In particular, by reversing the decision
and requiring valid param_info data in all nodes of a bitmap path
tree, we can get rid of indxpath.c's get_bitmap_tree_required_outer()
function, which computed the data on-demand. It's not unlikely that
that nets out as a savings of cycles in many scenarios. A couple
of other things in indxpath.c can be simplified as well.
While here, get rid of some cases in reparameterize_path_by_child()
that are visibly dead or useless, given that we only care about
reparameterizing paths that can be on the inside of a parameterized
nestloop. This case reminds one of the maxim that untested code
probably does not work, so I'm unwilling to leave unreachable code
in this function. (I did leave the T_Gather case in place even
though it's not reached in the regression tests. It's not very
clear to me when the planner might prefer to put Gather below
rather than above a nestloop, but at least in principle the case
might be interesting.)
Per bug #16536, originally from Arne Roland but with a test case
by Andrew Gierth. Back-patch to v11 where this code came in.
Discussion: https://postgr.es/m/16536-2213ee0b3aad41fd@postgresql.org
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
Writing a trailing semicolon in a macro is almost never the right thing,
because you almost always want to write a semicolon after each macro
call instead. (Even if there was some reason to prefer not to, pgindent
would probably make a hash of code formatted that way; so within PG the
rule should basically be "don't do it".) Thus, if we have a semi inside
the macro, the compiler sees "something;;". Much of the time the extra
empty statement is harmless, but it could lead to mysterious syntax
errors at call sites. In perhaps an overabundance of neatnik-ism, let's
run around and get rid of the excess semicolons whereever possible.
The only thing worse than a mysterious syntax error is a mysterious
syntax error that only happens in the back branches; therefore,
backpatch these changes where relevant, which is most of them because
most of these mistakes are old. (The lack of reported problems shows
that this is largely a hypothetical issue, but still, it could bite
us in some future patch.)
John Naylor and Tom Lane
Discussion: https://postgr.es/m/CACPNZCs0qWTqJ2QUSGJ07B7uvAvzMb-KbG2q+oo+J3tsWN5cqw@mail.gmail.com
WITH TIES is an option to the FETCH FIRST N ROWS clause (the SQL
standard's spelling of LIMIT), where you additionally get rows that
compare equal to the last of those N rows by the columns in the
mandatory ORDER BY clause.
There was a proposal by Andrew Gierth to implement this functionality in
a more powerful way that would yield more features, but the other patch
had not been finished at this time, so we decided to use this one for
now in the spirit of incremental development.
Author: Surafel Temesgen <surafel3000@gmail.com>
Reviewed-by: Álvaro Herrera <alvherre@alvh.no-ip.org>
Reviewed-by: Tomas Vondra <tomas.vondra@2ndquadrant.com>
Discussion: https://postgr.es/m/CALAY4q9ky7rD_A4vf=FVQvCGngm3LOes-ky0J6euMrg=_Se+ag@mail.gmail.com
Discussion: https://postgr.es/m/87o8wvz253.fsf@news-spur.riddles.org.uk
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
In the wake of commit 1cff1b95a, the obvious way to sort a List
is to apply qsort() directly to the array of ListCells. list_qsort
was building an intermediate array of pointers-to-ListCells, which
we no longer need, but getting rid of it forces an API change:
the comparator functions need to do one less level of indirection.
Since we're having to touch the callers anyway, let's do two additional
changes: sort the given list in-place rather than making a copy (as
none of the existing callers have any use for the copying behavior),
and rename list_qsort to list_sort. It was argued that the old name
exposes more about the implementation than it should, which I find
pretty questionable, but a better reason to rename it is to be sure
we get the attention of any external callers about the need to fix
their comparator functions.
While we're at it, change four existing callers of qsort() to use
list_sort instead; previously, they all had local reinventions
of list_qsort, ie build-an-array-from-a-List-and-qsort-it.
(There are some other places where changing to list_sort perhaps
would be worthwhile, but they're less obviously wins.)
Discussion: https://postgr.es/m/29361.1563220190@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
Tom Lane
David Rowley
Bruce Momjian
Alvaro Herrera
Tomas Vondra
Michael Paquier
Peter Eisentraut
Thomas Munro
Peter Geoghegan
Jeff Davis
Amit Kapila