We can't support these properly, since once the relation's physical files
are unlinked, there's no way to roll back the transaction. I suppose
we could postpone the unlink till transaction commit, but then what of
BEGIN; DROP TABLE foo; CREATE TABLE foo; ?
The code does allow dropping a table/index created in the current
transaction block, however, since the post-abort state would be that
the table doesn't exist anyway.
real name before doing lookup. We only want to index temp tables by their
real names in the relcache, to ensure there's not more than one relcache
entry for them.
has positive refcount, it is rebuilt from pg_class data. This ensures
that relcache entries will track changes made by other backends. Formerly,
a shared inval report would just be ignored if it happened to arrive while
the relcache entry was in use. Also, fix relcache to reset ref counts
to zero during transaction abort. Finally, change LockRelation() so that
it checks for shared inval reports after obtaining the lock. In this way,
once any kind of lock has been obtained on a rel, we can trust the relcache
entry to be up-to-date.
the SInval spinlock while it is calling the passed invalFunction or
resetFunction. This is necessary to avoid deadlock with lmgr change;
InvalidateSharedInvalid can be called recursively now. It should be
a good performance improvement anyway --- holding a spinlock for more
than a very short interval is a no-no.
insight that RelationFlushRelation ought to invoke smgrclose, and that the
way to make that work is to ensure that mdclose doesn't fail if the relation
is already closed (or unlinked, if we are looking at a DROP TABLE). While
I was testing that, I was able to identify several problems that we had
with multiple-segment relations. The system is now able to do initdb and
pass the regression tests with a very small segment size (I had it set to
64Kb per segment for testing). I don't believe that ever worked before.
File descriptor leaks seem to be gone too.
I have partially addressed the concerns we had about mdtruncate(), too.
On a Win32 or NFS filesystem it is not possible to unlink a file that
another backend is holding open, so what md.c now does is to truncate
unwanted files to zero length before trying to unlink them. The other
backends will be forced to close their open files by relation cache
invalidation --- but I think it would take considerable work to make
that happen before vacuum truncates the relation rather than after.
Leaving zero-length files lying around seems a usable compromise.
references or CASE expressions, didn't parenthesize complex expressions
properly. Also, always output variable references as fully qualified
names to eliminate ambiguity bug recently reported. (This could be
smarter, but reliability comes first.)
Most parts of the planner should ignore, or indeed never even see, uplevel
Vars because they will be or have been replaced by Params. There were a
couple of places that got it wrong though, probably my fault from recent
changes...
last loop which would return the *first* surviving-to-that-point candidate
regardless of which one actually passed the test. This was producing
such curious results as 'oid % 2' getting translated to 'int2(oid) % 2'.
documented intepretation of the lefthand and oper fields. Fix a number of
obscure problems while at it --- for example, the old code failed if the parser
decided to insert a type-coercion function just below the operator of a
SubLink.
CAUTION: this will break stored rules that contain subplans. You may
need to initdb.
ALLOC_BIGCHUNK_LIMIT are always allocated as separate malloc() blocks,
and are free()d immediately upon pfree(). Also, if such a chunk is enlarged
with repalloc(), translate the operation into a realloc() so as to
minimize memory usage. Of course, these large chunks still get freed
automatically if the alloc set is reset.
I have set ALLOC_BIGCHUNK_LIMIT at 64K for now, but perhaps another
size would be better?
match then it tried for a self-commutative operator with the reversed input
data types. This is pretty silly; there could never be such an operator,
except maybe in binary-compatible-type scenarios, and we have oper_inexact
for that. Besides which, the oprsanity regress test would complain about
such an operator. Remove nonfunctional code and simplify routine calling
convention accordingly.
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 ...
sort order down into planner, instead of handling it only at the very top
level of the planner. This fixes many things. An explicit sort is now
avoided if there is a cheaper alternative (typically an indexscan) not
only for ORDER BY, but also for the internal sort of GROUP BY. It works
even when there is no other reason (such as a WHERE condition) to consider
the indexscan. It works for indexes on functions. It works for indexes
on functions, backwards. It's just so cool...
CAUTION: I have changed the representation of SortClause nodes, therefore
THIS UPDATE BREAKS STORED RULES. You will need to initdb.
above a Sort or Materialize node. As far as I can tell, the only place
that actually needed that was set_tlist_references, which was being lazy
about checking to see if it had a noname node to fix or not...
store all ordering information in pathkeys lists (which are now lists of
lists of PathKeyItem nodes, not just lists of lists of vars). This was
a big win --- the code is smaller and IMHO more understandable than it
was, even though it handles more cases. I believe the node changes will
not force an initdb for anyone; planner nodes don't show up in stored
rules.
commuted (ie, the index var appears on the right). These are now handled
the same way as merge and hash join quals that need to be commuted: the
actual reversing of the clause only happens if we actually choose the path
and generate a plan from it. Furthermore, the clause is only reversed in
the 'indexqual' field of the plan, not in the 'indxqualorig' field. This
allows the clause to still be recognized and removed from qpquals of upper
level join plans. Also, simplify and generalize match_clause_to_indexkey;
now it recognizes binary-compatible indexes for join as well as restriction
clauses.
contains much code that looks like it will handle indexquals with the index
key on either side of the operator, in fact indexquals must have the index
key on the left because of limitations of the ScanKey machinery. Perhaps
someone will be motivated to fix that someday...
work under a wider range of scenarios than it did --- it formerly did not
handle a multi-pass inner scan, nor cases in which the inner scan's
indxqualorig or non-index qual contained outer var references. I am not
sure that these limitations could be hit in the existing optimizer, but
they need to be fixed for future expansion.
> >
> > was implemented by Jan Wieck.
> > His work is for ascending order cases.
> >
> > Here is a patch to prevent sorting also in descending
> > order cases.
> > Because I had already changed _bt_first() to position
> > backward correctly before v6.5,this patch would work.
> >
Hiroshi Inoue
Inoue@tpf.co.jp
multi-scan indexscan plans; it tried to use the same table-to-index
attribute mapping for all the scans, even if they used different indexes.
It would klugily work as long as OR indexquals never used multikey indexes,
but that's not likely to hold up much longer...
to go along with expression_tree_walker. (_walker is not suitable for
routines that need to alter the tree structure significantly.) Other minor
cleanups in clauses.c.
Also, move responsibility for calling vc_abort into main xact.c list of
things-to-call-at-abort. What in the world was it doing down inside of
TransactionIdAbort()?
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.
(it should just call the given operator, not look up an = operator).
Fix intltsel() so that all numeric data types are converted to double
before trying to estimate where the given comparison value is in the
known range of column values. intltsel() still needs work, or replacement,
for non-numeric data types ... but for nonintegral numeric types it
should now be delivering reasonable estimates.
neqsel now behave as per my suggestions in pghackers a few days ago.
selectivity for < > <= >= should work OK for integral types as well, but
still need work for nonintegral types. Since these routines have never
actually executed before :-(, this may result in some significant changes
in the optimizer's choices of execution plans. Let me know if you see
any serious misbehavior.
CAUTION: THESE CHANGES REQUIRE INITDB. pg_statistic table has changed.
so that Case works in WHERE join clauses. Temporary patch --- this routine
is one of many that ought to be changed to use centralized expression-tree-
walking logic.
rels that the inner path needs to join to, but it was only checking for
the first one. Failure could only have been observed with an OR-clause
that mentions 3 or more tables, and then only if the bogus path was
actually selected as cheapest ...
optimizer rather than parser. This has many advantages, such as not
getting fooled by chance uses of operator names ~ and ~~ (the operators
are identified by OID now), and not creating useless comparison operations
in contexts where the comparisons will not actually be used as indexquals.
The new code also recognizes exact-match LIKE and regex patterns, and
produces an = indexqual instead of >= and <=.
This change does NOT fix the problem with non-ASCII locales: the code
still doesn't know how to generate an upper bound indexqual for non-ASCII
collation order. But it's no worse than before, just the same deficiency
in a different place...
Also, dike out loc_restrictinfo fields in Plan nodes. These were doing
nothing useful in the absence of 'expensive functions' optimization,
and they took a considerable amount of processing to fill in.
The only place it was being used was as temporary storage in indxpath.c,
and the logic was wrong: the same restrictinfo node could get chosen to
carry the info for two different joins. Right fix is to return a second
list of unjoined-relids parallel to the list of clause groups.
identified by Hiroshi (incorrect cost attributed to OR clauses
after multiple passes through set_rest_selec()). I think the code
was trying to allow selectivities of OR subclauses to be passed in
from outside, but noplace was actually passing any useful data, and
set_rest_selec() was passing wrong data.
Restructure representation of "indexqual" in IndexPath nodes so that
it is the same as for indxqual in completed IndexScan nodes: namely,
a toplevel list with an entry for each pass of the index scan, having
sublists that are implicitly-ANDed index qual conditions for that pass.
You don't want to know what the old representation was :-(
Improve documentation of OR-clause indexscan functions.
Remove useless 'notclause' field from RestrictInfo nodes. (This might
force an initdb for anyone who has stored rules containing RestrictInfos,
but I do not think that RestrictInfo ever appears in completed plans.)
support, but which the grammar was accepting. Also, fix several bugs
having to do with failure to copy fields up from a subselect to a select
or insert node.
of the SELECT part of the statement is just like a plain SELECT. All
INSERT-specific processing happens after the SELECT parsing is done.
This eliminates many problems, e.g. INSERT ... SELECT ... GROUP BY using
the wrong column labels. Ensure that DEFAULT clauses are coerced to
the target column type, whether or not stored clause produces the right
type. Substantial cleanup of parser's array support.
creates a reduce/reduce conflict, which I resolved by changing the
'AexprConst -> Typename Sconst' rule to 'AexprConst -> SimpleTypename Sconst'.
In other words, a subscripted type declaration can't be used in that
syntax any longer. This seems a small price to pay for not having to
qualify subscripted columns anymore.
Other cleanups: rename res_target_list to update_target_list, and remove
productions for variants that are not legal in an UPDATE target list;
rename res_target_list2 to plain target_list; delete position_expr
in favor of using b_expr in that production; merge opt_indirection
into attr nonterminal, since there are no places where an unsubscripted
attr is wanted; fix typos in Param support; change case_arg so that
an arbitrary a_expr is allowed, not only a column name.
care of equal-key cases, eliminating bt_firsteq(). The linear search
formerly done by bt_firsteq() took a lot of time in the case where many
equal keys appear on the same page.
that contain null fields. Old code would produce erratic sort results
because comparisons of tuples containing nulls could produce inconsistent
answers.
> the DTK_MICROSEC case is just like the DTK_MILLISEC case.
> I think this is wrong and it ought to look like
> fsec = rint(fsec * 1000000) / 1000000;
> no?
Tom Lane.
"HAS_LONG_LONG" is defined based on the assumption that
strtol() would return ERANGE if a platform does not support
64-bit integers. In current PostgreSQL 6.5 (and 6.4.2)
distribution, "HAS_LONG_LONG" is defined only if platform
is "alpha". (See include/port/alpha.h) I think the int4
range check should apply to linux_alpha as well. (I have
not tested yet but I guess this might be applicable to
newer Linux/i386 distributions which includes new GCC which
implements long int as 64-bit int.)
with expression_tree_walker-based code. The former failed to cope with
expressions containing SubLinks, and the latter returned TRUE for both
SubLinks and Aggrefs (cut-and-paste bug?). There is a lot more scope for
using expression_tree_walker in this module, but I'll restrain myself
until the 6.6 split occurs from touching not-demonstrably-broken code.
is parse_aggs.c. This fixes its failure to cope with (at least) CaseExpr
and ArrayRef nodes, which is the reason why both of these fail in 6.5:
select coalesce(f1,0) from int4_tbl group by f1;
ERROR: Illegal use of aggregates or non-group column in target list
select sentence.words[0] from sentence group by sentence.words[0];
ERROR: Illegal use of aggregates or non-group column in target list
The array case still fails, but at least it's not parse_agg's fault
anymore ... considering that we now support CASE officially, I think
it's important to fix the first example ...
will gradually replace all of the boilerplate tree-walk-recursion code that
currently exists in O(N) slightly different forms in N subroutines.
I've had it with adding missing cases to these subroutines...
special hack to ensure it would close its output file even after failure
due to elog(ERROR) partway through the copy. This is now unnecessary
because fd.c takes care of cleaning up open files at transaction abort;
worse, after fd.c closed the file copy.c would try to do so *again* at
the start of the next COPY command. This would result in havoc in most
implementations of stdio library.
tlist and qual are NULL. It ought to handle these the same as the cases
where tlist contains only constant expressions, ie, be willing to generate
a Result-node plan. This did not use to matter, but it does now because
union_planner will flatten the tlist when aggregates are present. Thus,
'select count(1) from table' now causes query_planner to be given a null
tlist, and to duplicate 6.4's behavior we need it to give back a Result
plan rather than refusing the query. 6.4 was arguably doing the Wrong
Thing for this query, but I'm not going to open a semantics issue right
before 6.5 release ... can revisit that problem later.
returned NULL, which it will do in some cases where an elog(ERROR) would
probably be more appropriate. For the moment, generate a not-very-
informative error message rather than proceeding to certain coredump.
Probably ought to think about making query_planner elog instead of
returning NULL, but this is at least a safe change for now.
pointer to palloc'd but uninitialized memory. This is not cool; anyone looking
at the returned 'tuple' would at best coredump and at worst behave in a
bizarre and irreproducible way. Fix it to return a predictable value,
namely a correctly-set-up palloc'd tuple containing zero attributes.
I believe this fix is both safe and critical.
this one could be useful for people experiencing out-of-memory crashes while
executing queries which retrieve or use a very large number of tuples.
The problem happens when storage is allocated for functions results used in
a large query, for example:
select upper(name) from big_table;
select big_table.array[1] from big_table;
select count(upper(name)) from big_table;
This patch is a dirty hack that fixes the out-of-memory problem for the most
common cases, like the above ones. It is not the final solution for the
problem but it can work for some people, so I'm posting it.
The patch should be safe because all changes are under #ifdef. Furthermore
the feature can be enabled or disabled at runtime by the `free_tuple_memory'
options in the pg_options file. The option is disabled by default and must
be explicitly enabled at runtime to have any effect.
To enable the patch add the follwing line to Makefile.custom:
CUSTOM_COPT += -DFREE_TUPLE_MEMORY
To enable the option at runtime add the following line to pg_option:
free_tuple_memory=1
Massimo
/*
* Read above about cases when !ItemIdIsUsed(Citemid)
* (child item is removed)... Due to the fact that
* at the moment we don't remove unuseful part of
* update-chain, it's possible to get too old
* parent row here. Like as in the case which
* caused this problem, we stop shrinking here.
* I could try to find real parent row but want
* not to do it because of real solution will
* be implemented anyway, latter, and we are too
* close to 6.5 release. - vadim 06/11/99
*/
if (Ptp.t_data->t_xmax != tp.t_data->t_xmin)
...
and possibly for other cases too:
DO NOT cache status of transaction in unknown state
(i.e. non-committed and non-aborted ones)
Example:
T1 reads row updated/inserted by running T2 and cache T2 status.
T2 commits.
Now T1 reads a row updated by T2 and with HEAP_XMAX_COMMITTED
in t_infomask (so cached T2 status is not changed).
Now T1 EvalPlanQual gets updated row version without HEAP_XMIN_COMMITTED
-> TransactionIdDidCommit(t_xmin) and TransactionIdDidAbort(t_xmin)
return FALSE and T2 decides that t_xmin is not committed and gets
ERROR above.
It's too late to find more smart way to handle such cases and so
I just changed xact status caching and got rid TransactionIdFlushCache()
from code.
Changed: transam.c, xact.c, lmgr.c and transam.h - last three
just because of TransactionIdFlushCache() is removed.
2. heapam.c:
T1 marked a row for update. T2 waits for T1 commit/abort.
T1 commits. T3 updates the row before T2 locks row page.
Now T2 sees that new row t_xmax is different from xact id (T1)
T2 was waiting for. Old code did Assert here. New one goes to
HeapTupleSatisfiesUpdate. Obvious changes too.
3. Added Assert to vacuum.c
4. bufmgr.c: break
Assert(buf->r_locks == 0 && !buf->ri_lock)
into two Asserts.
after ExecEndNode. It must be done! Or we'll be out of free
tuple slots very soon, though slots are freed by ExecEndNode
and ready for reusing.
We didn't see this problem before because of
int nSlots = ExecCountSlotsNode(plan);
TupleTable tupleTable = ExecCreateTupleTable(nSlots + 10);
/* why add ten? - jolly */
code in InitPlan - i.e. extra 10 slots. Simple select uses
3 slots and so it was possible to re-use evaluation plan
3 additional times and didn't get
elog(NOTICE, "Plan requires more slots than are available");
elog(ERROR, "send mail to your local executor guru to fix this");
Changes are obvious and shouldn't be problems with them.
Though, I added Assert(epqstate->es_tupleTable->next == 0)
before EvalPlanQual():ExecInitNode and we'll notice if
something is still wrong. Is it better to change Assert
to elog(ERROR) ?
a non-leading % would be put into the >=/<= patterns. Also, repair
longstanding confusion about whether %% means a literal %%. The SQL92
doesn't say any such thing, and textlike() knows that, but gram.y didn't.
2. varsup.c:ReadNewTransactionId(): don't read nextXid from disk -
this func doesn't allocate next xid, so ShmemVariableCache->nextXid
may be used (but GetNewTransactionId() must be called first).
3. vacuum.c: change elog(ERROR, "Child item....") to elog(NOTICE) -
this is not ERROR, proper handling is just not implemented, yet.
4. s_lock.c: increase S_MAX_BUSY by 2 times.
5. shmem.c:GetSnapshotData(): have to call ReadNewTransactionId()
_after_ SpinAcquire(ShmemIndexLock).
the gettimeofday doesn't compile under Linux with glibc2 because
the DST_NONE constant is no more defined. It seems that this code
(written by me) has always be wrong but for some reason working.
From: Massimo Dal Zotto <dz@cs.unitn.it>
transactions will not assume that MyProc transaction was committed
before snapshot calculations. With old MyProc->xid assignment
(in xact.c:StartTransaction()) there was ability to see the same
row twice (I used gdb for this)!...
2. Assignments of InvalidTransactionId to MyProc->xid and MyProc->xmin
are moved from xact.c:CommitTransaction() to
xact.c:RecordTransactionCommit() - this invalidation must be done
before releasing transaction locks or bad (too high) XmaxRecent value
might be used by vacuum ("ERROR: Child itemid marked as unused"
reported by "Hiroshi Inoue" <Inoue@tpf.co.jp>; once again, gdb
allowed me reproduce this error).
2. Get rid of locking when updating statistics in vacuum.
3. Use QuerySnapshot in COPY TO and call SetQuerySnashot
in main tcop loop before FETCH and COPY TO.
redundant) SearchSysCache searches per table column in an INSERT, which
accounted for a good percentage of the CPU time for INSERT ... VALUES().
Now it only does two searches in the typical case.
through MAXBACKENDS array entries used to be fine when MAXBACKENDS = 64.
It's not so cool with MAXBACKENDS = 1024 (or more!), especially not in a
frequently-used routine like SIDelExpiredDataEntries. Repair by making
procState array size be the soft MaxBackends limit rather than the hard
limit, and by converting SIGetProcStateLimit() to a macro.
not be marked inFromCl any longer. Otherwise the planner gets confused
and joins over them where in fact it does not have to.
Adjust hasSubLinks now with a recursive lookup - could be wrong in
multi action rules because parse state isn't reset correctly and all
actions in the rule are marked hasSubLinks if one of them has.
Jan
aggregate functions, as in
select a, b from foo group by a;
The ungrouped reference to b is not kosher, but formerly we neglected to
check this unless there was an aggregate function somewhere in the query.
SelectStmt and CursorStmt tried to parse FOR UPDATE ... / FOR READ ONLY.
Cursor now checks that it is read only by looking at forUpdate of Query.
SelectStmt handles FOR READ ONLY too.
Jan
will pass through rather than spitting up. This is necessary to handle
cases where coerce_type causes a subexpression to be retransformed, as in
SELECT count(*) + 1.0 FROM table
remove optimizer's arbitrary limit on how large a join it will use hashing
for. (The limit was too large to prevent the problems we'd been seeing,
anyway...)
fixed-size hashtable. This should prevent 'hashtable out of memory' errors,
unless you really do run out of memory. Note: target size for hashtable
is now taken from -S postmaster switch, not -B, since it is local memory
in the backend rather than shared memory.
looks
like someone just didn't add support for multiple segments for
truncation.
The following patch seems to do the right thing, for me at least.
It passed my tests, my data looks right(no data that shouldn't be in
there) and regression is ok.
Ole Gjerde
Tom Lane
Bruce Momjian
Tatsuo Ishii
Thomas G. Lockhart
Vadim B. Mikheev
Marc G. Fournier
Jan Wieck