is using mark/restore but not rewind or backward-scan capability. Insert a
materialize plan node between a mergejoin and its inner child if the inner
child is a sort that is expected to spill to disk. The materialize shields
the sort from the need to do mark/restore and thereby allows it to perform
its final merge pass on-the-fly; while the materialize itself is normally
cheap since it won't spill to disk unless the number of tuples with equal
key values exceeds work_mem.
Greg Stark, with some kibitzing from Tom Lane.
recompute the limit/offset immediately, so that the updated values are
available when the child's ReScan function is invoked. Add a regression
test for this, too. Bug is new in HEAD (due to the bounded-sorting patch)
so no need for back-patch.
I did not do anything about merging this signaling with chgParam processing,
but if we were to do that we'd still need to compute the updated values
at this point rather than during the first ProcNode call.
Per observation and test case from Greg Stark, though I didn't use his patch.
need be returned. We keep a heap of the current best N tuples and sift-up
new tuples into it as we scan the input. For M input tuples this means
only about M*log(N) comparisons instead of M*log(M), not to mention a lot
less workspace when N is small --- avoiding spill-to-disk for large M
is actually the most attractive thing about it. Patch includes planner
and executor support for invoking this facility in ORDER BY ... LIMIT
queries. Greg Stark, with some editorialization by moi.
is in progress on the same hashtable. This seems the least invasive way to
fix the recently-recognized problem that a split could cause the scan to
visit entries twice or (with much lower probability) miss them entirely.
The only field-reported problem caused by this is the "failed to re-find
shared lock object" PANIC in COMMIT PREPARED reported by Michel Dorochevsky,
which was caused by multiply visited entries. However, it seems certain
that mdsync() is vulnerable to missing required fsync's due to missed
entries, and I am fearful that RelationCacheInitializePhase2() might be at
risk as well. Because of that and the generalized hazard presented by this
bug, back-patch all the supported branches.
Along the way, fix pg_prepared_statement() and pg_cursor() to not assume
that the hashtables they are examining will stay static between calls.
This is risky regardless of the newly noted dynahash problem, because
hash_seq_search() has never promised to cope with deletion of table entries
other than the just-returned one. There may be no bug here because the only
supported way to call these functions is via ExecMakeTableFunctionResult()
which will cycle them to completion before doing anything very interesting,
but it seems best to get rid of the assumption. This affects 8.2 and HEAD
only, since those functions weren't there earlier.
seen by code inspecting the expression. The best way to do this seems
to be to drop the original representation as a function invocation, and
instead make a special expression node type that represents applying
the element-type coercion function to each array element. In this way
the element function is exposed and will be checked for volatility.
Per report from Guillaume Smet.
parent query's EState. Now that there's a single flat rangetable for both
the main plan and subplans, there's no need anymore for a separate EState,
and removing it allows cleaning up some crufty code in nodeSubplan.c and
nodeSubqueryscan.c. Should be a tad faster too, although any difference
will probably be hard to measure. This is the last bit of subsidiary
mop-up work from changing to a flat rangetable.
useless substructure for its RangeTblEntry nodes. (I chose to keep using the
same struct node type and just zero out the link fields for unneeded info,
rather than making a separate ExecRangeTblEntry type --- it seemed too
fragile to have two different rangetable representations.)
Along the way, put subplans into a list in the toplevel PlannedStmt node,
and have SubPlan nodes refer to them by list index instead of direct pointers.
Vadim wanted to do that years ago, but I never understood what he was on about
until now. It makes things a *whole* lot more robust, because we can stop
worrying about duplicate processing of subplans during expression tree
traversals. That's been a constant source of bugs, and it's finally gone.
There are some consequent simplifications yet to be made, like not using
a separate EState for subplans in the executor, but I'll tackle that later.
storing mostly-redundant Query trees in prepared statements, portals, etc.
To replace Query, a new node type called PlannedStmt is inserted by the
planner at the top of a completed plan tree; this carries just the fields of
Query that are still needed at runtime. The statement lists kept in portals
etc. now consist of intermixed PlannedStmt and bare utility-statement nodes
--- no Query. This incidentally allows us to remove some fields from Query
and Plan nodes that shouldn't have been there in the first place.
Still to do: simplify the execution-time range table; at the moment the
range table passed to the executor still contains Query trees for subqueries.
initdb forced due to change of stored rules.
that aren't turned into true joins). Since this is the last missing bit of
infrastructure, go ahead and fill out the hash integer_ops and float_ops
opfamilies with cross-type operators. The operator family project is now
DONE ... er, except for documentation ...
involving HashAggregate over SubqueryScan (this is the known case, there
may well be more). The bug is only latent in releases before 8.2 since they
didn't try to access tupletable slots' descriptors during ExecDropTupleTable.
The least bogus fix seems to be to make subqueries share the parent query's
memory context, so that tupdescs they create will have the same lifespan as
those of the parent query. There are comments in the code envisioning going
even further by not having a separate child EState at all, but that will
require rethinking executor access to range tables, which I don't want to
tackle right now. Per bug report from Jean-Pierre Pelletier.
the XmlExpr code in various lists, use a representation that has some hope
of reverse-listing correctly (though it's still a de-escaping function
shy of correctness), generally try to make it look more like Postgres
coding conventions.
by name on each and every row processed. Profiling suggests this may
buy a percent or two for simple UPDATE scenarios, which isn't huge,
but when it's so easy to get ...
the SQL spec, viz IS NULL is true if all the row's fields are null, IS NOT
NULL is true if all the row's fields are not null. The former coding got
this right for a limited number of cases with IS NULL (ie, those where it
could disassemble a ROW constructor at parse time), but was entirely wrong
for IS NOT NULL. Per report from Teodor.
I desisted from changing the behavior for arrays, since on closer inspection
it's not clear that there's any support for that in the SQL spec. This
probably needs more consideration.
plpgsql support to come later. Along the way, convert execMain's
SELECT INTO support into a DestReceiver, in order to eliminate some ugly
special cases.
Jonah Harris and Tom Lane
that's shorter-lived than the expression state being evaluated in it really
doesn't work :-( --- we end up with fn_extra caches getting deleted while
still in use. Rather than abandon the notion of caching expression state
across domain_in calls altogether, I chose to make domain_in a bit cozier
with ExprContext. All we really need for evaluating variable-free
expressions is an ExprContext, not an EState, so I invented the notion of a
"standalone" ExprContext. domain_in can prevent resource leakages by doing
a ReScanExprContext on this rather than having to free it entirely; so we
can make the ExprContext have the same lifespan (and particularly the same
per_query memory context) as the expression state structs.
temporary context that can be reset when advancing to the next sublist.
This is faster and more thorough at recovering space than the previous
method; moreover it will do the right thing if something in the sublist
tries to register an expression context callback.
(e.g. "INSERT ... VALUES (...), (...), ...") and elsewhere as allowed
by the spec. (e.g. similar to a FROM clause subselect). initdb required.
Joe Conway and Tom Lane.
the opportunity to treat COUNT(*) as a zero-argument aggregate instead
of the old hack that equated it to COUNT(1); this is materially cleaner
(no more weird ANYOID cases) and ought to be at least a tiny bit faster.
Original patch by Sergey Koposov; review, documentation, simple regression
tests, pg_dump and psql support by moi.
tuple hash table entries. This addresses the problem previously noted
that use of a 'physical tlist' in the input scan node could bloat the
hash table entries far beyond what the planner expects. It's a better
answer than my previous thought of undoing the physical tlist optimization,
because we can also remove columns that are needed to compute the aggregate
functions but aren't part of the grouping column set.
support both FOR UPDATE and FOR SHARE in one command, as well as both
NOWAIT and normal WAIT behavior. The more general code is actually
simpler and cleaner.
(previously we only did = and <> correctly). Also, allow row comparisons
with any operators that are in btree opclasses, not only those with these
specific names. This gets rid of a whole lot of indefensible assumptions
about the behavior of particular operators based on their names ... though
it's still true that IN and NOT IN expand to "= ANY". The patch adds a
RowCompareExpr expression node type, and makes some changes in the
representation of ANY/ALL/ROWCOMPARE SubLinks so that they can share code
with RowCompareExpr.
I have not yet done anything about making RowCompareExpr an indexable
operator, but will look at that soon.
initdb forced due to changes in stored rules.
relation if it's already been locked by execMain.c as either a result
relation or a FOR UPDATE/SHARE relation. This avoids an extra trip to
the shared lock manager state. Per my suggestion yesterday.
it's worth probing the outer relation for emptiness before building the
hash table. To wit, if we're rescanning a join previously performed,
remember whether we found it nonempty the previous time, and don't bother
with the probe if it was nonempty. This buys back the performance lost
in examples like Mario Weilguni's.
"ctid IN (list)" will still work after we convert IN to ScalarArrayOpExpr.
Make some minor efficiency improvements while at it, such as ensuring that
multiple TIDs are fetched in physical heap order. And fix EXPLAIN so that
it shows what's really going on for a TID scan.
when we first read the page, rather than checking them one at a time.
This allows us to take and release the buffer content lock just once
per page, instead of once per tuple. Since it's a shared lock the
contention penalty for holding the lock longer shouldn't be too bad.
We can safely do this only when using an MVCC snapshot; else the
assumption that visibility won't change over time is uncool. Therefore
there are now two code paths depending on the snapshot type. I also
made the same change in nodeBitmapHeapscan.c, where it can be done always
because we only support MVCC snapshots for bitmap scans anyway.
Also make some incidental cleanups in the APIs of these functions.
Per a suggestion from Qingqing Zhou.
qualification when the underlying operator is indexable and useOr is true.
That is, indexkey op ANY (ARRAY[...]) is effectively translated into an
OR combination of one indexscan for each array element. This only works
for bitmap index scans, of course, since regular indexscans no longer
support OR'ing of scans. There are still some loose ends to clean up
before changing 'x IN (list)' to translate as a ScalarArrayOpExpr;
for instance predtest.c ought to be taught about it. But this gets the
basic functionality in place.
comment line where output as too long, and update typedefs for /lib
directory. Also fix case where identifiers were used as variable names
in the backend, but as typedefs in ecpg (favor the backend for
indenting).
Backpatch to 8.1.X.
slot of the topmost plan node when a trigger returns a modified tuple.
These appear to be the only places where a plan node's caller did not
treat the result slot as read-only, which is an assumption that nodeUnique
makes as of 8.1. Fixes trigger-vs-DISTINCT bug reported by Frank van Vugt.
outer relation is empty did not work, per test case from Patrick Welche.
It tried to use nodeHashjoin.c's high-level mechanisms for fetching an
outer-relation tuple, but that code expected the hash table to be filled
already. As patched, the code failed in corner cases such as having no
outer-relation tuples for the first hash batch. Revert and rewrite.
(a/k/a SELECT INTO). Instead, flush and fsync the whole relation before
committing. We do still need the WAL log when PITR is active, however.
Simon Riggs and Tom Lane.
work if either of the join relations are empty. The logic is:
(1) if the inner relation's startup cost is less than the outer
relation's startup cost and this is not an outer join, read
a single tuple from the inner relation via ExecHash()
- if NULL, we're done
(2) read a single tuple from the outer relation
- if NULL, we're done
(3) build the hash table on the inner relation
- if hash table is empty and this is not an outer join,
we're done
(4) otherwise, do hash join as usual
The implementation uses the new MultiExecProcNode API, per a
suggestion from Tom: invoking ExecHash() now produces the first
tuple from the Hash node's child node, whereas MultiExecHash()
builds the hash table.
I had to put in a bit of a kludge to get the row count returned
for EXPLAIN ANALYZE to be correct: since ExecHash() is invoked to
return a tuple, and then MultiExecHash() is invoked, we would
return one too many tuples to EXPLAIN ANALYZE. I hacked around
this by just manually detecting this situation and subtracting 1
from the EXPLAIN ANALYZE row count.
When one side of the join has a NULL, we don't want to uselessly try
to match it against every remaining tuple of the other side. While
at it, rewrite the comparison machinery to avoid multiple evaluations
of the left and right input expressions and to use a btree comparator
where available, instead of double operator calls. Also revise the
state machine to eliminate redundant comparisons and hopefully make it
more readable too.
startup to end, rather than re-opening it in each MultiExecBitmapIndexScan
call. I had foolishly thought that opening/closing wouldn't be much
more expensive than a rescan call, but that was sheer brain fade.
This seems to fix about half of the performance lossage reported by
Sergey Koposov. I'm still not sure where the other half went.
to eliminate unnecessary deadlocks. This commit adds SELECT ... FOR SHARE
paralleling SELECT ... FOR UPDATE. The implementation uses a new SLRU
data structure (managed much like pg_subtrans) to represent multiple-
transaction-ID sets. When more than one transaction is holding a shared
lock on a particular row, we create a MultiXactId representing that set
of transactions and store its ID in the row's XMAX. This scheme allows
an effectively unlimited number of row locks, just as we did before,
while not costing any extra overhead except when a shared lock actually
has to be shared. Still TODO: use the regular lock manager to control
the grant order when multiple backends are waiting for a row lock.
Alvaro Herrera and Tom Lane.
node, as this behavior is now better done as a bitmap OR indexscan.
This allows considerable simplification in nodeIndexscan.c itself as
well as several planner modules concerned with indexscan plan generation.
Also we can improve the sharing of code between regular and bitmap
indexscans, since they are now working with nigh-identical Plan nodes.
but just to open and close it during MultiExecBitmapIndexScan. This
avoids acquiring duplicate resources (eg, multiple locks on the same
relation) in a tree with many bitmap scans. Also, don't bother to
lock the parent heap at all here, since we must be underneath a
BitmapHeapScan node that will be holding a suitable lock.
scans, using in-memory tuple ID bitmaps as the intermediary. The planner
frontend (path creation and cost estimation) is not there yet, so none
of this code can be executed. I have tested it using some hacked planner
code that is far too ugly to see the light of day, however. Committing
now so that the bulk of the infrastructure changes go in before the tree
drifts under me.
executing a statement that fires triggers. Formerly this time was
included in "Total runtime" but not otherwise accounted for.
As a side benefit, we avoid re-opening relations when firing non-deferred
AFTER triggers, because the trigger code can re-use the main executor's
ResultRelInfo data structure.
of tuples when passing data up through multiple plan nodes. A slot can now
hold either a normal "physical" HeapTuple, or a "virtual" tuple consisting
of Datum/isnull arrays. Upper plan levels can usually just copy the Datum
arrays, avoiding heap_formtuple() and possible subsequent nocachegetattr()
calls to extract the data again. This work extends Atsushi Ogawa's earlier
patch, which provided the key idea of adding Datum arrays to TupleTableSlots.
(I believe however that something like this was foreseen way back in Berkeley
days --- see the old comment on ExecProject.) A test case involving many
levels of join of fairly wide tables (about 80 columns altogether) showed
about 3x overall speedup, though simple queries will probably not be
helped very much.
I have also duplicated some code in heaptuple.c in order to provide versions
of heap_formtuple and friends that use "bool" arrays to indicate null
attributes, instead of the old convention of "char" arrays containing either
'n' or ' '. This provides a better match to the convention used by
ExecEvalExpr. While I have not made a concerted effort to get rid of uses
of the old routines, I think they should be deprecated and eventually removed.
on-the-fly, and thereby avoid blowing out memory when the planner has
underestimated the hash table size. Hash join will now obey the
work_mem limit with some faithfulness. Per my recent proposal
(hash aggregate part isn't done yet though).
Also performed an initial run through of upgrading our Copyright date to
extend to 2005 ... first run here was very simple ... change everything
where: grep 1996-2004 && the word 'Copyright' ... scanned through the
generated list with 'less' first, and after, to make sure that I only
picked up the right entries ...
of an inheritance child table is binary-compatible with the rowtype of
its parent, invent an expression node type that does the conversion
correctly. Fixes the new bug exhibited by Kris Shannon as well as a
lot of old bugs that would only show up when using multiple inheritance
or after altering the parent table.
columns. The returned tuple needs to have appropriate NULL columns
inserted so that it actually matches the declared rowtype. It seemed
convenient to use a JunkFilter for this, so I made some cleanups and
simplifications in the JunkFilter code to allow it to support this
additional functionality. (That in turn exposed a latent bug in
nodeAppend.c, which is that it was returning a tuple slot whose
descriptor didn't match its data.) Also, move check_sql_fn_retval
out of pg_proc.c and into functions.c, where it seems to more naturally
belong.
As a side effect, cause subscripts in INSERT targetlists to do something
more or less sensible; previously we evaluated such subscripts and then
effectively ignored them. Another side effect is that UPDATE-ing an
element or slice of an array value that is NULL now produces a non-null
result, namely an array containing just the assigned-to positions.
rather than allowing them only in a few special cases as before. In
particular you can now pass a ROW() construct to a function that accepts
a rowtype parameter. Internal generation of RowExprs fixes a number of
corner cases that used to not work very well, such as referencing the
whole-row result of a JOIN or subquery. This represents a further step in
the work I started a month or so back to make rowtype values into
first-class citizens.
results with tuples as ordinary varlena Datums. This commit does not
in itself do much for us, except eliminate the horrid memory leak
associated with evaluation of whole-row variables. However, it lays the
groundwork for allowing composite types as table columns, and perhaps
some other useful features as well. Per my proposal of a few days ago.
so that the 'val' is computed only once, per recent discussion. The
speedup is not much when 'val' is just a simple variable, but could be
significant for larger expressions. More importantly this avoids issues
with multiple evaluations of a volatile 'val', and it allows the CASE
expression to be reverse-listed in its original form by ruleutils.c.
directly to the appropriate per-node execution function, using a function
pointer stored by ExecInitExpr. This speeds things up by eliminating one
level of function call. The function-pointer technique also enables further
small improvements such as only making one-time tests once (and then
changing the function pointer). Overall this seems to gain about 10%
on evaluation of simple expressions, which isn't earthshaking but seems
a worthwhile gain for a relatively small hack. Per recent discussion
on pghackers.
when scanning a table that we need all the columns from. In case of
SELECT INTO, we have to check that the hasoids flag matches the desired
output type, too. Per report from Mike Mascari.
regular qpqual ('filter condition'), add special-purpose code to
nodeIndexscan.c to recheck them. This ends being almost no net addition
of code, because the removal of planner code balances out the extra
executor code, but it is significantly more efficient when a lossy
operator is involved in an OR indexscan. The old implementation had
to recheck the entire indexqual in such cases.
discussion on pgsql-hackers: in READ COMMITTED mode we just have to force
a QuerySnapshot update in the trigger, but in SERIALIZABLE mode we have
to run the scan under a current snapshot and then complain if any rows
would be updated/deleted that are not visible in the transaction snapshot.
to allow es_snapshot to be set to SnapshotNow rather than a query snapshot.
This solves a bug reported by Wade Klaver, wherein triggers fired as a
result of RI cascade updates could misbehave.
handling many-way scans: instead of re-evaluating all prior indexscan
quals to see if a tuple has been fetched more than once, use a hash table
indexed by tuple CTID. But fall back to the old way if the hash table
grows to exceed SortMem.
as well as the hash function (formerly the comparison function was hardwired
as memcmp()). This makes it possible to eliminate the special-purpose
hashtable management code in execGrouping.c in favor of using dynahash to
manage tuple hashtables; which is a win because dynahash knows how to expand
a hashtable when the original size estimate was too small, whereas the
special-purpose code was too stupid to do that. (See recent gripe from
Stephan Szabo about poor performance when hash table size estimate is way
off.) Free side benefit: when using string_hash, the default comparison
function is now strncmp() instead of memcmp(). This should eliminate some
part of the overhead associated with larger NAMEDATALEN values.
'scalar op ALL (array)', where the operator is applied between the
lefthand scalar and each element of the array. The operator must
yield boolean; the result of the construct is the OR or AND of the
per-element results, respectively.
Original coding by Joe Conway, after an idea of Peter's. Rewritten
by Tom to keep the implementation strictly separate from subqueries.
specific hash functions used by hash indexes, rather than the old
not-datatype-aware ComputeHashFunc routine. This makes it safe to do
hash joining on several datatypes that previously couldn't use hashing.
The sets of datatypes that are hash indexable and hash joinable are now
exactly the same, whereas before each had some that weren't in the other.
of an index can now be a computed expression instead of a simple variable.
Restrictions on expressions are the same as for predicates (only immutable
functions, no sub-selects). This fixes problems recently introduced with
inlining SQL functions, because the inlining transformation is applied to
both expression trees so the planner can still match them up. Along the
way, improve efficiency of handling index predicates (both predicates and
index expressions are now cached by the relcache) and fix 7.3 oversight
that didn't record dependencies of predicate expressions.
expressions, ARRAY(sub-SELECT) expressions, some array functions.
Polymorphic functions using ANYARRAY/ANYELEMENT argument and return
types. Some regression tests in place, documentation is lacking.
Joe Conway, with some kibitzing from Tom Lane.
entire contents of the subplan into the tuplestore before we can return
any tuples. Instead, the tuplestore holds what we've already read, and
we fetch additional rows from the subplan as needed. Random access to
the previously-read rows works with the tuplestore, and doesn't affect
the state of the partially-read subplan. This is a step towards fixing
the problems with cursors over complex queries --- we don't want to
stick in Materialize nodes if they'll prevent quick startup for a cursor.
locParam lists can be converted to bitmapsets to speed updating. Also,
replace 'locParam' with 'allParam', which contains all the paramIDs
relevant to the node (i.e., the union of extParam and locParam); this
saves a step during SetChangedParamList() without costing anything
elsewhere.
startup, not in the parser; this allows ALTER DOMAIN to work correctly
with domain constraint operations stored in rules. Rod Taylor;
code review by Tom Lane.
Simplify SubLink by storing just a List of operator OIDs, instead of
a list of incomplete OpExprs --- that was a bizarre and bulky choice,
with no redeeming social value since we have to build new OpExprs
anyway when forming the plan tree.
given any malloc block until something is first allocated in it; but
thereafter, MemoryContextReset won't release that first malloc block.
This preserves the quick-reset property of the original policy, without
forcing 8K to be allocated to every context whether any of it is ever
used or not. Also, remove some more no-longer-needed explicit freeing
during ExecEndPlan.
a per-query memory context created by CreateExecutorState --- and destroyed
by FreeExecutorState. This provides a final solution to the longstanding
problem of memory leaked by various ExecEndNode calls.
in the planned representation of a subplan at all any more, only SubPlan.
This means subselect.c doesn't scribble on its input anymore, which seems
like a good thing; and there are no longer three different possible
interpretations of a SubLink. Simplify node naming and improve comments
in primnodes.h. No change to stored rules, though.
execution state trees, and ExecEvalExpr takes an expression state tree
not an expression plan tree. The plan tree is now read-only as far as
the executor is concerned. Next step is to begin actually exploiting
this property.
make VALUE a non-reserved word again, use less invasive method of passing
ConstraintTestValue into transformExpr, fix problems with nested constraint
testing, do correct thing with NULL result from a constraint expression,
remove memory leak. Domain checks still need much more work if we are going
to allow ALTER DOMAIN, however.
to plan nodes, not vice-versa. All executor state nodes now inherit from
struct PlanState. Copying of plan trees has been simplified by not
storing a list of SubPlans in Plan nodes (eliminating duplicate links).
The executor still needs such a list, but it can build it during
ExecutorStart since it has to scan the plan tree anyway.
No initdb forced since no stored-on-disk structures changed, but you
will need a full recompile because of node-numbering changes.
instead of only one. This should speed up planning (only one hash path
to consider for a given pair of relations) as well as allow more effective
hashing, when there are multiple hashable joinclauses.
one more row from the subplan than the COUNT would appear to require.
This costs a little more logic but a number of people have complained
about the old implementation.
node now does its own grouping of the input rows, and has no need for a
preceding GROUP node in the plan pipeline. This allows elimination of
the misnamed tuplePerGroup option for GROUP, and actually saves more code
in nodeGroup.c than it costs in nodeAgg.c, as well as being presumably
faster. Restructure the API of query_planner so that we do not commit to
using a sorted or unsorted plan in query_planner; instead grouping_planner
makes the decision. (Right now it isn't any smarter than query_planner
was, but that will change as soon as it has the option to select a hash-
based aggregation step.) Despite all the hackery, no initdb needed since
only in-memory node types changed.