logic operations during planning. Seems cleaner to create two new Path
node types, instead --- this avoids duplication of cost-estimation code.
Also, create an enable_bitmapscan GUC parameter to control use of bitmap
plans.
scans, using in-memory tuple ID bitmaps as the intermediary. The planner
frontend (path creation and cost estimation) is not there yet, so none
of this code can be executed. I have tested it using some hacked planner
code that is far too ugly to see the light of day, however. Committing
now so that the bulk of the infrastructure changes go in before the tree
drifts under me.
access: define new index access method functions 'amgetmulti' that can
fetch multiple TIDs per call. (The functions exist but are totally
untested as yet.) Since I was modifying pg_am anyway, remove the
no-longer-needed 'rel' parameter from amcostestimate functions, and
also remove the vestigial amowner column that was creating useless
work for Alvaro's shared-object-dependencies project.
Initdb forced due to changes in pg_am.
structs. There are many places in the planner where we were passing
both a rel and an index to subroutines, and now need only pass the
index struct. Notationally simpler, and perhaps a tad faster.
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 ...
In particular, there was a mathematical tie between the two possible
nestloop-with-materialized-inner-scan plans for a join (ie, we computed
the same cost with either input on the inside), resulting in a roundoff
error driven choice, if the relations were both small enough to fit in
sort_mem. Add a small cost factor to ensure we prefer materializing the
smaller input. This changes several regression test plans, but with any
luck we will now have more stability across platforms.
1. Solve the problem of not having TOAST references hiding inside composite
values by establishing the rule that toasting only goes one level deep:
a tuple can contain toasted fields, but a composite-type datum that is
to be inserted into a tuple cannot. Enforcing this in heap_formtuple
is relatively cheap and it avoids a large increase in the cost of running
the tuptoaster during final storage of a row.
2. Fix some interesting problems in expansion of inherited queries that
reference whole-row variables. We never really did this correctly before,
but it's now relatively painless to solve by expanding the parent's
whole-row Var into a RowExpr() selecting the proper columns from the
child.
If you dike out the preventive check in CheckAttributeType(),
composite-type columns now seem to actually work. However, we surely
cannot ship them like this --- without I/O for composite types, you
can't get pg_dump to dump tables containing them. So a little more
work still to do.
In the past, we used a 'Lispy' linked list implementation: a "list" was
merely a pointer to the head node of the list. The problem with that
design is that it makes lappend() and length() linear time. This patch
fixes that problem (and others) by maintaining a count of the list
length and a pointer to the tail node along with each head node pointer.
A "list" is now a pointer to a structure containing some meta-data
about the list; the head and tail pointers in that structure refer
to ListCell structures that maintain the actual linked list of nodes.
The function names of the list API have also been changed to, I hope,
be more logically consistent. By default, the old function names are
still available; they will be disabled-by-default once the rest of
the tree has been updated to use the new API names.
That particular corner case is not exactly compelling, but given 7.4's
ability to discard redundant join clauses, it is possible for the situation
to arise from queries that are not so obviously silly. Per bug report
of 6-Apr-04.
corner cases that could stand improvement, but it does all the basic
stuff. A byproduct is that the selectivity routines are no longer
constrained to working on simple Vars; we might in future be able to
improve the behavior for subexpressions that don't match indexes.
Make btree index creation and initial validation of foreign-key constraints
use maintenance_work_mem rather than work_mem as their memory limit.
Add some code to guc.c to allow these variables to be referenced by their
old names in SHOW and SET commands, for backwards compatibility.
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.
with index qual clauses in the Path representation. This saves a little
work during createplan and (probably more importantly) allows reuse of
cached selectivity estimates during indexscan planning. Also fix latent
bug: wrong plan would have been generated for a 'special operator' used
in a nestloop-inner-indexscan join qual, because the special operator
would not have gotten into the list of quals to recheck. This bug is
only latent because at present the special-operator code could never
trigger on a join qual, but sooner or later someone will want to do it.
join conditions in which each OR subclause includes a constraint on
the same relation. This implements the other useful side-effect of
conversion to CNF format, without its unpleasant side-effects. As
per pghackers discussion of a few weeks ago.
teaching the latter to accept either RestrictInfo nodes or bare
clause expressions; and cache the selectivity result in the RestrictInfo
node when possible. This extends the caching behavior of approx_selectivity
to many more contexts, and should reduce duplicate selectivity
calculations.
----------------------------------------------------------------------
/*
* relation_byte_size
* Estimate the storage space in bytes for a given number of tuples
* of a given width (size in bytes).
*/
static double
relation_byte_size(double tuples, int width)
{
return tuples * (MAXALIGN(width) + MAXALIGN(sizeof(HeapTupleData)));
}
----------------------------------------------------------------------
Shouldn't this be HeapTupleHeaderData and not HeapTupleData ?
(Of course, from a costing perspective these shouldn't be very different but ...)
Sailesh Krishnamurthy
and hash bucket-size estimation. Issue has been there awhile but is more
critical in 7.4 because it affects varchar columns. Per report from
Greg Stark.
node emits only those vars that are actually needed above it in the
plan tree. (There were comments in the code suggesting that this was
done at some point in the dim past, but for a long time we have just
made join nodes emit everything that either input emitted.) Aside from
being marginally more efficient, this fixes the problem noted by Peter
Eisentraut where a join above an IN-implemented-as-join might fail,
because the subplan targetlist constructed in the latter case didn't
meet the expectation of including everything.
Along the way, fix some places that were O(N^2) in the targetlist
length. This is not all the trouble spots for wide queries by any
means, but it's a step forward.
'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.
passed to join selectivity estimators. Make use of this in eqjoinsel
to derive non-bogus selectivity for IN clauses. Further tweaking of
cost estimation for IN.
initdb forced because of pg_proc.h changes.
Try to model the effect of rescanning input tuples in mergejoins;
account for JOIN_IN short-circuiting where appropriate. Also, recognize
that mergejoin and hashjoin clauses may now be more than single operator
calls, so we have to charge appropriate execution costs.
There are two implementation techniques: the executor understands a new
JOIN_IN jointype, which emits at most one matching row per left-hand row,
or the result of the IN's sub-select can be fed through a DISTINCT filter
and then joined as an ordinary relation.
Along the way, some minor code cleanup in the optimizer; notably, break
out most of the jointree-rearrangement preprocessing in planner.c and
put it in a new file prep/prepjointree.c.
containing a volatile function), rather than only on 'Var = Var' clauses
as before. This makes it practical to do flatten_join_alias_vars at the
start of planning, which in turn eliminates a bunch of klugery inside the
planner to deal with alias vars. As a free side effect, we now detect
implied equality of non-Var expressions; for example in
SELECT ... WHERE a.x = b.y and b.y = 42
we will deduce a.x = 42 and use that as a restriction qual on a. Also,
we can remove the restriction introduced 12/5/02 to prevent pullup of
subqueries whose targetlists contain sublinks.
Still TODO: make statistical estimation routines in selfuncs.c and costsize.c
smarter about expressions that are more complex than plain Vars. The need
for this is considerably greater now that we have to be able to estimate
the suitability of merge and hash join techniques on such expressions.
costs for expression evaluation, not only per-tuple cost as before.
This extension is needed in order to deal realistically with hashed or
materialized sub-selects.
computation: reduce the bucket number mod nbatch. This changes the
association between original bucket numbers and batches, but that
doesn't matter. Minor other cleanups in hashjoin code to help
centralize decisions.
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.
so that all executable expression nodes inherit from a common supertype
Expr. This is somewhat of an exercise in code purity rather than any
real functional advance, but getting rid of the extra Oper or Func node
formerly used in each operator or function call should provide at least
a little space and speed improvement.
initdb forced by changes in stored-rules representation.
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.
parameter to allow it to be forced off for comparison purposes.
Add ORDER BY clauses to a bunch of regression test queries that will
otherwise produce randomly-ordered output in the new regime.
Reused the Expr node to hold DISTINCT which strongly resembles
the existing OP info. Define DISTINCT_EXPR which strongly resembles
the existing OPER_EXPR opType, but with handling for NULLs required
by SQL99.
We have explicit support for single-element DISTINCT comparisons
all the way through to the executor. But, multi-element DISTINCTs
are handled by expanding into a comparison tree in gram.y as is done for
other row comparisons. Per discussions, it might be desirable to move
this into one or more purpose-built nodes to be handled in the backend.
Define the optional ROW keyword and token per SQL99.
This allows single-element row constructs, which were formerly disallowed
due to shift/reduce conflicts with parenthesized a_expr clauses.
Define the SQL99 TREAT() function. Currently, use as a synonym for CAST().
comments on one of the optimizer functions a lot more
clear, adds a summary of the recent KSQO discussion to the
comments in the code, adds regression tests for the bug with
sequence state Tom fixed recently and another reg. test, and
removes some PostQuel legacy stuff: ExecAppend -> ExecInsert,
ExecRetrieve -> ExecSelect, etc.
Error messages remain unchanged until a vote.
Neil Conway
comments on one of the optimizer functions a lot more
clear, adds a summary of the recent KSQO discussion to the
comments in the code, adds regression tests for the bug with
sequence state Tom fixed recently and another reg. test, and
removes some PostQuel legacy stuff: ExecAppend -> ExecInsert,
ExecRetrieve -> ExecSelect, etc. This was changed because the
elog() messages from this routine are user-visible, so we
should be using the SQL terms.
Neil Conway
some kibitzing from Tom Lane. Not everything works yet, and there's
no documentation or regression test, but let's commit this so Joe
doesn't need to cope with tracking changes in so many files ...
now has an RTE of its own, and references to its outputs now are Vars
referencing the JOIN RTE, rather than CASE-expressions. This allows
reverse-listing in ruleutils.c to use the correct alias easily, rather
than painfully reverse-engineering the alias namespace as it used to do.
Also, nested FULL JOINs work correctly, because the result of the inner
joins are simple Vars that the planner can cope with. This fixes a bug
reported a couple times now, notably by Tatsuo on 18-Nov-01. The alias
Vars are expanded into COALESCE expressions where needed at the very end
of planning, rather than during parsing.
Also, beginnings of support for showing plan qualifier expressions in
EXPLAIN. There are probably still cases that need work.
initdb forced due to change of stored-rule representation.
both input streams to the end. If one variable's range is much less
than the other, an indexscan-based merge can win by not scanning all
of the other table. Per example from Reinhard Max.
pgsql-hackers. pg_opclass now has a row for each opclass supported by each
index AM, not a row for each opclass name. This allows pg_opclass to show
directly whether an AM supports an opclass, and furthermore makes it possible
to store additional information about an opclass that might be AM-dependent.
pg_opclass and pg_amop now store "lossy" and "haskeytype" information that we
previously expected the user to remember to provide in CREATE INDEX commands.
Lossiness is no longer an index-level property, but is associated with the
use of a particular operator in a particular index opclass.
Along the way, IndexSupportInitialize now uses the syscaches to retrieve
pg_amop and pg_amproc entries. I find this reduces backend launch time by
about ten percent, at the cost of a couple more special cases in catcache.c's
IndexScanOK.
Initial work by Oleg Bartunov and Teodor Sigaev, further hacking by Tom Lane.
initdb forced.
should be computed from total number of distinct values in whole
relation, not # distinct values we expect to have after restriction
clauses are applied.
of costsize.c routines to pass Query root, so that costsize can figure
more things out by itself and not be so dependent on its callers to tell
it everything it needs to know. Use selectivity of hash or merge clause
to estimate number of tuples processed internally in these joins
(this is more useful than it would've been before, since eqjoinsel is
somewhat more accurate than before).
create_index_paths are not immediately discarded, but are available for
subsequent planner work. This allows avoiding redundant syscache lookups
in several places. Change interface to operator selectivity estimation
procedures to allow faster and more flexible estimation.
Initdb forced due to change of pg_proc entries for selectivity functions!
collected by ANALYZE. Also, add some modest amount of intelligence to
guesses that are used for varlena columns in the absence of any ANALYZE
statistics. The 'width' reported by EXPLAIN is finally something less
than totally bogus for varlena columns ... and, in consequence, hashjoin
estimating should be a little better ...
a separate statement (though it can still be invoked as part of VACUUM, too).
pg_statistic redesigned to be more flexible about what statistics are
stored. ANALYZE now collects a list of several of the most common values,
not just one, plus a histogram (not just the min and max values). Random
sampling is used to make the process reasonably fast even on very large
tables. The number of values and histogram bins collected is now
user-settable via an ALTER TABLE command.
There is more still to do; the new stats are not being used everywhere
they could be in the planner. But the remaining changes for this project
should be localized, and the behavior is already better than before.
A not-very-related change is that sorting now makes use of btree comparison
routines if it can find one, rather than invoking '<' twice.
avoid repeated evaluations in cost_qual_eval(). This turns out to save
a useful fraction of planning time. No change to external representation
of RestrictInfo --- although that node type doesn't appear in stored
rules anyway.
(Don't forget that an alias is required.) Views reimplemented as expanding
to subselect-in-FROM. Grouping, aggregates, DISTINCT in views actually
work now (he says optimistically). No UNION support in subselects/views
yet, but I have some ideas about that. Rule-related permissions checking
moved out of rewriter and into executor.
INITDB REQUIRED!
materialized tupleset is small enough) instead of a temporary relation.
This was something I was thinking of doing anyway for performance, and Jan
says he needs it for TOAST because he doesn't want to cope with toasting
noname relations. With this change, the 'noname table' support in heap.c
is dead code, and I have accordingly removed it. Also clean up 'noname'
plan handling in planner --- nonames are either sort or materialize plans,
and it seems less confusing to handle them separately under those names.
That means you can now set your options in either or all of $PGDATA/configuration,
some postmaster option (--enable-fsync=off), or set a SET command. The list of
options is in backend/utils/misc/guc.c, documentation will be written post haste.
pg_options is gone, so is that pq_geqo config file. Also removed were backend -K,
-Q, and -T options (no longer applicable, although -d0 does the same as -Q).
Added to configure an --enable-syslog option.
changed all callers from TPRINTF to elog(DEBUG)
table for an average of NTUP_PER_BUCKET tuples/bucket, but cost_hashjoin
was assuming a target load of one tuple/bucket. This was causing a
noticeable underestimate of hashjoin costs.
to next integer. Previously, if selectivity was small, we could compute
very tiny scan cost on the basis of estimating that only 0.001 tuple
would be fetched, which is silly. This naturally led to some rather
silly plans...
costs using the inner path's parent->rows count as the number of tuples
processed per inner scan iteration. This is wrong when we are using an
inner indexscan with indexquals based on join clauses, because the rows
count in a Relation node reflects the selectivity of the restriction
clauses for that rel only. Upshot was that if join clause was very
selective, we'd drastically overestimate the true cost of the join.
Fix is to calculate correct output-rows estimate for an inner indexscan
when the IndexPath node is created and save it in the path node.
Change of path node doesn't require initdb, since path nodes don't
appear in saved rules.
accesses versus sequential accesses, a (very crude) estimate of the
effects of caching on random page accesses, and cost to evaluate WHERE-
clause expressions. Export critical parameters for this model as SET
variables. Also, create SET variables for the planner's enable flags
(enable_seqscan, enable_indexscan, etc) so that these can be controlled
more conveniently than via PGOPTIONS.
Planner now estimates both startup cost (cost before retrieving
first tuple) and total cost of each path, so it can optimize queries
with LIMIT on a reasonable basis by interpolating between these costs.
Same facility is a win for EXISTS(...) subqueries and some other cases.
Redesign pathkey representation to achieve a major speedup in planning
(I saw as much as 5X on a 10-way join); also minor changes in planner
to reduce memory consumption by recycling discarded Path nodes and
not constructing unnecessary lists.
Minor cleanups to display more-plausible costs in some cases in
EXPLAIN output.
Initdb forced by change in interface to index cost estimation
functions.
fields in JoinPaths --- turns out that we do need that after all :-(.
Also, rearrange planner so that only one RelOptInfo is created for a
particular set of joined base relations, no matter how many different
subsets of relations it can be created from. This saves memory and
processing time compared to the old method of making a bunch of RelOptInfos
and then removing the duplicates. Clean up the jointree iteration logic;
not sure if it's better, but I sure find it more readable and plausible
now, particularly for the case of 'bushy plans'.
pghackers discussion of 5-Jan-2000. The amopselect and amopnpages
estimators are gone, and in their place is a per-AM amcostestimate
procedure (linked to from pg_am, not pg_amop).
and fix_opids processing to a single recursive pass over the plan tree
executed at the very tail end of planning, rather than haphazardly here
and there at different places. Now that tlist Vars do not get modified
until the very end, it's possible to get rid of the klugy var_equal and
match_varid partial-matching routines, and just use plain equal()
throughout the optimizer. This is a step towards allowing merge and
hash joins to be done on expressions instead of only Vars ...
hashjoinable clause, not one path for a randomly-chosen element of each
set of clauses with the same join operator. That is, if you wrote
SELECT ... WHERE t1.f1 = t2.f2 and t1.f3 = t2.f4,
and both '=' ops were the same opcode (say, all four fields are int4),
then the system would either consider hashing on f1=f2 or on f3=f4,
but it would *not* consider both possibilities. Boo hiss.
Also, revise estimation of hashjoin costs to include a penalty when the
inner join var has a high disbursion --- ie, the most common value is
pretty common. This tends to lead to badly skewed hash bucket occupancy
and way more comparisons than you'd expect on average.
I imagine that the cost calculation still needs tweaking, but at least
it generates a more reasonable plan than before on George Young's example.
==========================================
What follows is a set of diffs that cleans up the usage of BLCKSZ.
As a side effect, the person compiling the code can change the
value of BLCKSZ _at_their_own_risk_. By that, I mean that I've
tried it here at 4096 and 16384 with no ill-effects. A value
of 4096 _shouldn't_ affect much as far as the kernel/file system
goes, but making it bigger than 8192 can have severe consequences
if you don't know what you're doing. 16394 worked for me, _BUT_
when I went to 32768 and did an initdb, the SCSI driver broke and
the partition that I was running under went to hell in a hand
basket. Had to reboot and do a good bit of fsck'ing to fix things up.
The patch can be safely applied though. Just leave BLCKSZ = 8192
and everything is as before. It basically only cleans up all of the
references to BLCKSZ in the code.
If this patch is applied, a comment in the config.h file though above
the BLCKSZ define with warning about monkeying around with it would
be a good idea.
Darren darrenk@insightdist.com
(Also cleans up some of the #includes in files referencing BLCKSZ.)
==========================================
2. PageWeights are variables now.
3. Fixed using ceil((double)selec*indextuples) as estimation
for expected heap pages: ceil((double)selec*relpages) now.
/usr/include/limits.h (which quiets the costsize.c warnings)...under
FreeBSD, /usr/include/limits.h *includes* machine/limits.h, while under
Solaris, there is no such things as /usr/include/machine...
Problem with Solaris pointed out by Mark Wahl