This requires changing quite a few places that were depending on
sizeof(HeapTupleHeaderData), but it seems for the best.
Michael Paquier, some adjustments by me
If we expect batching at the very beginning, we size nbuckets for
"full work_mem" (see how many tuples we can get into work_mem,
while not breaking NTUP_PER_BUCKET threshold).
If we expect to be fine without batching, we start with the 'right'
nbuckets and track the optimal nbuckets as we go (without actually
resizing the hash table). Once we hit work_mem (considering the
optimal nbuckets value), we keep the value.
At the end of the first batch, we check whether (nbuckets !=
nbuckets_optimal) and resize the hash table if needed. Also, we
keep this value for all batches (it's OK because it assumes full
work_mem, and it makes the batchno evaluation trivial). So the
resize happens only once.
There could be cases where it would improve performance to allow
the NTUP_PER_BUCKET threshold to be exceeded to keep everything in
one batch rather than spilling to a second batch, but attempts to
generate such a case have so far been unsuccessful; that issue may
be addressed with a follow-on patch after further investigation.
Tomas Vondra with minor format and comment cleanup by me
Reviewed by Robert Haas, Heikki Linnakangas, and Kevin Grittner
Since this makes the bucket headers use ~10x as much memory, properly
account for that memory when we figure out whether everything fits
in work_mem. This might result in some cases that previously used
only a single batch getting split into multiple batches, but it's
unclear as yet whether we need defenses against that case, and if so,
what the shape of those defenses should be.
It's worth noting that even in these edge cases, users should still be
no worse off than they would have been last week, because commit
45f6240a8f saved a big pile of memory
on exactly the same workloads.
Tomas Vondra, reviewed and somewhat revised by me.
Instead of palloc'ing each HashJoinTuple individually, allocate 32kB chunks
and pack the tuples densely in the chunks. This avoids the AllocChunk
header overhead, and the space wasted by standard allocator's habit of
rounding sizes up to the nearest power of two.
This doesn't contain any planner changes, because the planner's estimate of
memory usage ignores the palloc overhead. Now that the overhead is smaller,
the planner's estimates are in fact more accurate.
Tomas Vondra, reviewed by Robert Haas.
The dynahash code requires the number of buckets in a hash table to fit
in an int; but since we calculate the desired hash table size dynamically,
there are various scenarios where we might calculate too large a value.
The resulting overflow can lead to infinite loops, division-by-zero
crashes, etc. I (tgl) had previously installed some defenses against that
in commit 299d171652, but that covered only one
call path. Moreover it worked by limiting the request size to work_mem,
but in a 64-bit machine it's possible to set work_mem high enough that the
problem appears anyway. So let's fix the problem at the root by installing
limits in the dynahash.c functions themselves.
Trouble report and patch by Jeff Davis.
This reduces unnecessary exposure of other headers through htup.h, which
is very widely included by many files.
I have chosen to move the function prototypes to the new file as well,
because that means htup.h no longer needs to include tupdesc.h. In
itself this doesn't have much effect in indirect inclusion of tupdesc.h
throughout the tree, because it's also required by execnodes.h; but it's
something to explore in the future, and it seemed best to do the htup.h
change now while I'm busy with it.
This commit changes index-only scans so that data is read directly from the
index tuple without first generating a faux heap tuple. The only immediate
benefit is that indexes on system columns (such as OID) can be used in
index-only scans, but this is necessary infrastructure if we are ever to
support index-only scans on expression indexes. The executor is now ready
for that, though the planner still needs substantial work to recognize
the possibility.
To do this, Vars in index-only plan nodes have to refer to index columns
not heap columns. I introduced a new special varno, INDEX_VAR, to mark
such Vars to avoid confusion. (In passing, this commit renames the two
existing special varnos to OUTER_VAR and INNER_VAR.) This allows
ruleutils.c to handle them with logic similar to what we use for subplan
reference Vars.
Since index-only scans are now fundamentally different from regular
indexscans so far as their expression subtrees are concerned, I also chose
to change them to have their own plan node type (and hence, their own
executor source file).
This provides information about the numbers of tuples that were visited
but not returned by table scans, as well as the numbers of join tuples
that were considered and discarded within a join plan node.
There is still some discussion going on about the best way to report counts
for outer-join situations, but I think most of what's in the patch would
not change if we revise that, so I'm going to go ahead and commit it as-is.
Documentation changes to follow (they weren't in the submitted patch
either).
Marko Tiikkaja, reviewed by Marc Cousin, somewhat revised by Tom
This is advantageous first because it allows us to hash the smaller table
regardless of the outer-join type, and second because hash join can be more
flexible than merge join in dealing with arbitrary join quals in a FULL
join. For merge join all the join quals have to be mergejoinable, but hash
join will work so long as there's at least one hashjoinable qual --- the
others can be any condition. (This is true essentially because we don't
keep per-inner-tuple match flags in merge join, while hash join can do so.)
To do this, we need a has-it-been-matched flag for each tuple in the
hashtable, not just one for the current outer tuple. The key idea that
makes this practical is that we can store the match flag in the tuple's
infomask, since there are lots of bits there that are of no interest for a
MinimalTuple. So we aren't increasing the size of the hashtable at all for
the feature.
To write this without turning the hash code into even more of a pile of
spaghetti than it already was, I rewrote ExecHashJoin in a state-machine
style, similar to ExecMergeJoin. Other than that decision, it was pretty
straightforward.
relation using the general PARAM_EXEC executor parameter mechanism, rather
than the ad-hoc kluge of passing the outer tuple down through ExecReScan.
The previous method was hard to understand and could never be extended to
handle parameters coming from multiple join levels. This patch doesn't
change the set of possible plans nor have any significant performance effect,
but it's necessary infrastructure for future generalization of the concept
of an inner indexscan plan.
ExecReScan's second parameter is now unused, so it's removed.
The purpose of this change is to eliminate the need for every caller
of SearchSysCache, SearchSysCacheCopy, SearchSysCacheExists,
GetSysCacheOid, and SearchSysCacheList to know the maximum number
of allowable keys for a syscache entry (currently 4). This will
make it far easier to increase the maximum number of keys in a
future release should we choose to do so, and it makes the code
shorter, too.
Design and review by Tom Lane.
We show the number of buckets, the number of batches (and also the original
number if it has changed), and the peak space used by the hash table. Minor
executor changes to track peak space used.
"column < constant", and the comparison value is in the first or last
histogram bin or outside the histogram entirely, try to fetch the actual
column min or max value using an index scan (if there is an index on the
column). If successful, replace the lower or upper histogram bound with
that value before carrying on with the estimate. This limits the
estimation error caused by moving min/max values when the comparison
value is close to the min or max. Per a complaint from Josh Berkus.
It is tempting to consider using this mechanism for mergejoinscansel as well,
but that would inject index fetches into main-line join estimation not just
endpoint cases. I'm refraining from that until we can get a better handle
on the costs of doing this type of lookup.
and teach ANALYZE to compute such stats for tables that have subclasses.
Per my proposal of yesterday.
autovacuum still needs to be taught about running ANALYZE on parent tables
when their subclasses change, but the feature is useful even without that.
The original coding ensured nbuckets and nbatch didn't exceed INT_MAX,
which while not insane on its own terms did nothing to protect subsequent
code like "palloc(nbatch * sizeof(BufFile *))". Since enormous join size
estimates might well be planner error rather than reality, it seems best
to constrain the initial sizes to be not more than work_mem/sizeof(pointer),
thus ensuring the allocated arrays don't exceed work_mem. We will allow
nbatch to get bigger than that during subsequent ExecHashIncreaseNumBatches
calls, but we should still guard against integer overflow in those palloc
requests. Per bug #5145 from Bernt Marius Johnsen.
Although the given test case only seems to fail back to 8.2, previous
releases have variants of this issue, so patch all supported branches.
distribution, by creating a special fast path for the (first few) most common
values of the outer relation. Tuples having hashvalues matching the MCVs
are effectively forced to be in the first batch, so that we never write
them out to the batch temp files.
Bryce Cutt and Ramon Lawrence, with some editorialization by me.
for each temp file, rather than once per sort or hashjoin; this allows
spreading the data of a large sort or join across multiple tablespaces.
(I remain dubious that this will make any difference in practice, but certain
people insisted.) Arrange to cache the results of parsing the GUC variable
instead of recomputing from scratch on every demand, and push usage of the
cache down to the bottommost fd.c level.
tablespace(s) in which to store temp tables and temporary files. This is a
list to allow spreading the load across multiple tablespaces (a random list
element is chosen each time a temp object is to be created). Temp files are
not stored in per-database pgsql_tmp/ directories anymore, but per-tablespace
directories.
Jaime Casanova and Albert Cervera, with review by Bernd Helmle and Tom Lane.
selecting power-of-2, rather than prime, numbers of buckets in hash joins.
If the hash functions are doing their jobs properly by making all hash bits
equally random, this is good enough, and it saves expensive integer division
and modulus operations.
Hashing for aggregation purposes still needs work, so it's not time to
mark any cross-type operators as hashable for general use, but these cases
work if the operators are so marked by hand in the system catalogs.
match because they contain a null join key (and the join operator is
known strict). Improves performance significantly when the inner
relation contains a lot of nulls, as per bug #2930.
any use in the past many years, we'd have made some effort to include
them in all executor node types; but in fact they were only in
nodeAppend.c and nodeIndexscan.c, up until I copied nodeIndexscan.c's
occurrence into the new bitmap node types. Remove some other unused
macros in execdebug.h, too. Some day the whole header probably ought to
go away in favor of better-designed facilities.
bits indicating which optional capabilities can actually be exercised
at runtime. This will allow Sort and Material nodes, and perhaps later
other nodes, to avoid unnecessary overhead in common cases.
This commit just adds the infrastructure and arranges to pass the correct
flag values down to plan nodes; none of the actual optimizations are here
yet. I'm committing this separately in case anyone wants to measure the
added overhead. (It should be negligible.)
Simon Riggs and Tom Lane
generate their output tuple descriptors from their target lists (ie, using
ExecAssignResultTypeFromTL()). We long ago fixed things so that all node
types have minimally valid tlists, so there's no longer any good reason to
have two different ways of doing it. This change is needed to fix bug
reported by Hayden James: the fix of 2005-11-03 to emit the correct column
names after optimizing away a SubqueryScan node didn't work if the new
top-level plan node used ExecAssignResultTypeFromOuterPlan to generate its
tupdesc, since the next plan node down won't have the correct column labels.
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.
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.
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.
return just a single tuple at a time. Currently the only such node
type is Hash, but I expect we will soon have indexscans that can return
tuple bitmaps. A side benefit is that EXPLAIN ANALYZE now shows the
correct tuple count for a Hash node.
old comment in the code claimed that this was necessary. Since it is not
actually necessary any more, it is clearer to remove the comment and
just return NULL instead -- the return value of ExecHash() is not used.
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 ...
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.
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.
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.
the hashclauses field of the parent HashJoin. This avoids problems with
duplicated links to SubPlans in hash clauses, as per report from
Andrew Holm-Hansen.
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.
(materialization into a tuple store) discussed on pgsql-hackers earlier.
I've updated the documentation and the regression tests.
Notes on the implementation:
- I needed to change the tuple store API slightly -- it assumes that it
won't be used to hold data across transaction boundaries, so the temp
files that it uses for on-disk storage are automatically reclaimed at
end-of-transaction. I added a flag to tuplestore_begin_heap() to control
this behavior. Is changing the tuple store API in this fashion OK?
- in order to store executor results in a tuple store, I added a new
CommandDest. This works well for the most part, with one exception: the
current DestFunction API doesn't provide enough information to allow the
Executor to store results into an arbitrary tuple store (where the
particular tuple store to use is chosen by the call site of
ExecutorRun). To workaround this, I've temporarily hacked up a solution
that works, but is not ideal: since the receiveTuple DestFunction is
passed the portal name, we can use that to lookup the Portal data
structure for the cursor and then use that to get at the tuple store the
Portal is using. This unnecessarily ties the Portal code with the
tupleReceiver code, but it works...
The proper fix for this is probably to change the DestFunction API --
Tom suggested passing the full QueryDesc to the receiveTuple function.
In that case, callers of ExecutorRun could "subclass" QueryDesc to add
any additional fields that their particular CommandDest needed to get
access to. This approach would work, but I'd like to think about it for
a little bit longer before deciding which route to go. In the mean time,
the code works fine, so I don't think a fix is urgent.
- (semi-related) I added a NO SCROLL keyword to DECLARE CURSOR, and
adjusted the behavior of SCROLL in accordance with the discussion on
-hackers.
- (unrelated) Cleaned up some SGML markup in sql.sgml, copy.sgml
Neil Conway
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.
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.
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.
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.
value '-2' is used to indicate a variable-width type whose width is
computed as strlen(datum)+1. Everything that looks at typlen is updated
except for array support, which Joe Conway is working on; at the moment
it wouldn't work to try to create an array of cstring.
PostgreSQL. This hash function replaces the one used by hash indexes and
the catalog cache. Hash joins use a different, relatively poor-quality
hash function, but I'll fix that later.
As suggested by Tom Lane, this patch also changes the size of the fixed
hash table used by the catalog cache to be a power-of-2 (instead of a
prime: I chose 256 instead of 257). This allows the catcache to lookup
hash buckets using a simple bitmask. This should improve the performance
of the catalog cache slightly, since the previous method (modulo a
prime) was slow.
In my tests, this improves the performance of hash indexes by between 4%
and 8%; the performance when using btree indexes or seqscans is
basically unchanged.
Neil Conway <neilconway@rogers.com>
maintained for each cache entry. A cache entry will not be freed until
the matching ReleaseSysCache call has been executed. This eliminates
worries about cache entries getting dropped while still in use. See
my posting to pg-hackers of even date for more info.
for example, an SQL function can be used in a functional index. (I make
no promises about speed, but it'll work ;-).) Clean up and simplify
handling of functions returning sets.
right circumstances a hash join executed as a DECLARE CURSOR/FETCH
query would crash the backend. Problem as seen in current sources was
that the hash tables were stored in a context that was a child of
TransactionCommandContext, which got zapped at completion of the FETCH
command --- but cursor cleanup executed at COMMIT expected the tables
to still be valid. I haven't chased down the details as seen in 7.0.*
but I'm sure it's the same general problem.
There's now only one transition value and transition function.
NULL handling in aggregates is a lot cleaner. Also, use Numeric
accumulators instead of integer accumulators for sum/avg on integer
datatypes --- this avoids overflow at the cost of being a little slower.
Implement VARIANCE() and STDDEV() aggregates in the standard backend.
Also, enable new LIKE selectivity estimators by default. Unrelated
change, but as long as I had to force initdb anyway...
Tom Lane
Bruce Momjian
Kevin Grittner
Heikki Linnakangas
Robert Haas
Alvaro Herrera
Magnus Hagander
Neil Conway
PostgreSQL Daemon