the format on Tuple(Numeric) and the format to calculate(NumericVar)
are different. I understood that to reduce I/O. However, when many
comparisons or calculations of NUMERIC are executed, the conversion
of Numeric and NumericVar becomes a bottleneck.
It is profile result when "create index on NUMERIC column" is executed:
% cumulative self self total
time seconds seconds calls s/call s/call name
17.61 10.27 10.27 34542006 0.00 0.00 cmp_numerics
11.90 17.21 6.94 34542006 0.00 0.00 comparetup_index
7.42 21.54 4.33 71102587 0.00 0.00 AllocSetAlloc
7.02 25.64 4.09 69084012 0.00 0.00 set_var_from_num
4.87 28.48 2.84 69084012 0.00 0.00 alloc_var
4.79 31.27 2.79 142205745 0.00 0.00 AllocSetFreeIndex
4.55 33.92 2.65 34542004 0.00 0.00 cmp_abs
4.07 36.30 2.38 71101189 0.00 0.00 AllocSetFree
3.83 38.53 2.23 69084012 0.00 0.00 free_var
The create index command executes many comparisons of Numeric values.
Functions other than comparetup_index spent a lot of cycles for
conversion from Numeric to NumericVar.
An attached patch enables the comparison of Numeric values without
executing conversion to NumericVar. The execution time of that SQL
becomes half.
o Test SQL (index_test table has 1,000,000 tuples)
create index index_test_idx on index_test(num_col);
o Test results (executed the test five times)
(1)PentiumIII
original: 39.789s 36.823s 36.737s 37.752s 37.019s
patched : 18.560s 19.103s 18.830s 18.408s 18.853s
4.07 36.30 2.38 71101189 0.00 0.00 AllocSetFree
3.83 38.53 2.23 69084012 0.00 0.00 free_var
The create index command executes many comparisons of Numeric values.
Functions other than comparetup_index spent a lot of cycles for
conversion from Numeric to NumericVar.
An attached patch enables the comparison of Numeric values without
executing conversion to NumericVar. The execution time of that SQL
becomes half.
o Test SQL (index_test table has 1,000,000 tuples)
create index index_test_idx on index_test(num_col);
o Test results (executed the test five times)
(1)PentiumIII
original: 39.789s 36.823s 36.737s 37.752s 37.019s
patched : 18.560s 19.103s 18.830s 18.408s 18.853s
(2)Pentium4
original: 16.349s 14.997s 12.979s 13.169s 12.955s
patched : 7.005s 6.594s 6.770s 6.740s 6.828s
(3)Itanium2
original: 15.392s 15.447s 15.350s 15.370s 15.417s
patched : 7.413s 7.330s 7.334s 7.339s 7.339s
(4)Ultra Sparc
original: 64.435s 59.336s 59.332s 58.455s 59.781s
patched : 28.630s 28.666s 28.983s 28.744s 28.595s
Atsushi Ogawa
would basically punt in all cases for 'foo <> ALL (array)', which resulted
in a performance regression for NOT IN compared to what we were doing in
8.1 and before. Per report from Pavel Stehule.
relations: fix the executor so that we can have an Append plan on the
inside of a nestloop and still pass down outer index keys to index scans
within the Append, then generate such plans as if they were regular
inner indexscans. This avoids the need to evaluate the outer relation
multiple times.
... in fact, it will be applied now in any query whatsoever. I'm still
a bit concerned about the cycles that might be expended in failed proof
attempts, but given that CE is turned off by default, it's the user's
choice whether to expend those cycles or not. (Possibly we should
change the simple bool constraint_exclusion parameter to something
more fine-grained?)
internally $$ strings are converted to single-quote strings.
In ecpg, output newlines in commands using standard C escapes, rather
than using literal newlines, which is not portable.
thereby sharing code with the inheritance case. This puts the UNION-ALL-view
approach to partitioned tables on par with inheritance, so far as constraint
exclusion is concerned: it works either way. (Still need to update the docs
to say so.) The definition of "simple UNION ALL" is a little simpler than
I would like --- basically the union arms can only be SELECT * FROM foo
--- but it's good enough for partitioned-table cases.
MemSet on AIX by setting MEMSET_LOOP_LIMIT to zero.
Add optimization to skip MemSet tests in MEMSET_LOOP_LIMIT == 0 case and
just call memset() directly.
it later. This fixes a problem where EXEC_BACKEND didn't have progname
set, causing a segfault if log_min_messages was set below debug2 and our
own snprintf.c was being used.
Also alway strdup() progname.
Backpatch to 8.1.X and 8.0.X.
inheritance trees on-the-fly, which pretty well constrained us to considering
only one way of planning inheritance, expand inheritance sets during the
planner prep phase, and build a side data structure that can be consulted
later to find which RTEs are members of which inheritance sets. As proof of
concept, use the data structure to plan joins against inheritance sets more
efficiently: we can now use indexes on the set members in inner-indexscan
joins. (The generated plans could be improved further, but it'll take some
executor changes.) This data structure will also support handling UNION ALL
subqueries in the same way as inheritance sets, but that aspect of it isn't
finished yet.
to avoid sharing substructure with the lower-level indexquals. This is
currently only an issue if there are SubPlans in the indexquals, which is
uncommon but not impossible --- see bug #2218 reported by Nicholas Vinen.
We use the same kluge for indexqual vs indexqualorig in the index scans
themselves ... would be nice to clean this up someday.
requested sort order. It was assuming that build_index_pathkeys always
generates a pathkey per index column, which was not true if implied equality
deduction had determined that two index columns were effectively equated to
each other. Simplest fix seems to be to install an option that causes
build_index_pathkeys to support this behavior as well as the original one.
Per report from Brian Hirt.
memory in the executor's per-query memory context. It also inefficient:
it invokes get_call_result_type() and TupleDescGetAttInMetadata() for
every call to return_next, rather than invoking them once (per PL/Perl
function call) and memoizing the result.
This patch makes the following changes:
- refactor the code to include all the "per PL/Perl function call" data
inside a single struct, "current_call_data". This means we don't need to
save and restore N pointers for every recursive call into PL/Perl, we
can just save and restore one.
- lookup the return type metadata needed by plperl_return_next() once,
and then stash it in "current_call_data", so as to avoid doing the
lookup for every call to return_next.
- create a temporary memory context in which to evaluate the return
type's input functions. This memory context is reset for each call to
return_next.
The patch appears to fix the memory leak, and substantially reduces
the overhead imposed by return_next.
one 'creating subdirectories' message instead of one per subdirectory.
The original decision to print something for each subdirectory was made
when there were only one or two of 'em; we have way too many now.
Per discussion.
While we normally prefer the notation "foo.*" for a whole-row Var, that does
not work at SELECT top level, because in that context the parser will assume
that what is wanted is to expand the "*" into a list of separate target
columns, yielding behavior different from a whole-row Var. We have to emit
just "foo" instead in that context. Per report from Sokolov Yura.
and rely exclusively on the SQL type system to tell the difference between
the types. Prevent creation of invalid CIDR values via casting from INET
or set_masklen() --- both of these operations now silently zero any bits
to the right of the netmask. Remove duplicate CIDR comparison operators,
letting the type rely on the INET operators instead.
just refer to btree index entries as plain IndexTuples, which is what
they have been for a very long time. This is mostly just an exercise
in removing extraneous notation, but it does save a palloc/pfree cycle
per index insertion.
because pqSendSome will absorb input data anytime it'd be forced to block.
Avoiding a kernel call per PQputCopyData call helps COPY speed materially.
Alon Goldshuv
and non-required keys in a btree index scan, mark the required scankeys
with private flag bits SK_BT_REQFWD and/or SK_BT_REQBKWD. This seems
at least marginally clearer to me, and it eliminates a wired-into-the-
data-structure assumption that required keys are consecutive. Even though
that assumption will remain true for the foreseeable future, having it
in there makes the code seem more complex than necessary.
and DELETE. If specified, the alias must be used instead of the full
table name. Also, the alias currently cannot be used in the SET clause
of UPDATE.
Patch from Atsushi Ogawa, various editorialization by Neil Conway.
Along the way, make the rowtypes regression test pass if add_missing_from
is enabled, and add a new (skeletal) regression test for DELETE.
to try to create a log segment file concurrently, but the code erroneously
specified O_EXCL to open(), resulting in a needless failure. Before 7.4,
it was even a PANIC condition :-(. Correct code is actually simpler than
what we had, because we can just say O_CREAT to start with and not need a
second open() call. I believe this accounts for several recent reports of
hard-to-reproduce "could not create file ...: File exists" errors in both
pg_clog and pg_subtrans.
Continue to support GRANT ON [TABLE] for sequences for backward
compatibility; issue warning for invalid sequence permissions.
[Backward compatibility warning message.]
Add USAGE permission for sequences that allows only currval() and
nextval(), not setval().
Mention object name in grant/revoke warnings because of possible
multi-object operations.
temp table not only our own process' tables. It's not real important
since vacuum.c will skip temp tables anyway, but might as well make the
code do what it claims to do.
index's support-function cache (in index_getprocinfo). Since none of that
data can change for an index that's in active use, it seems sufficient to
treat all open indexes the same way we were treating "nailed" system indexes
--- that is, just re-read the pg_class row and leave the rest of the relcache
entry strictly alone. The pg_class re-read might not be strictly necessary
either, but since the reltablespace and relfilenode can change in normal
operation it seems safest to do it. (We don't support changing any of the
other info about an index at all, at the moment.)
Back-patch as far as 8.0. It might be possible to adapt the patch to 7.4,
but it would take more work than I care to expend for such a low-probability
problem. 7.3 is out of luck for sure.
occurs when it tries to heap_open pg_tablespace. When control returns to
smgrcreate, that routine will be holding a dangling pointer to a closed
SMgrRelation, resulting in mayhem. This is of course a consequence of
the violation of proper module layering inherent in having smgr.c call
a tablespace command routine, but the simplest fix seems to be to change
the locking mechanism. There's no real need for TablespaceCreateDbspace
to touch pg_tablespace at all --- it's only opening it as a way of locking
against a parallel DROP TABLESPACE command. A much better answer is to
create a special-purpose LWLock to interlock these two operations.
This drops TablespaceCreateDbspace quite a few layers down the food chain
and makes it something reasonably safe for smgr to call.
This is utterly insignificant in normal operation, but it becomes a
problem during cache inval stress testing. The original coding in fact
had no leak --- the 8.0 List rewrite created the issue. I wonder whether
list_concat should pfree the discarded header?
files: avoid creating stats hashtable entries for tables that aren't being
touched except by vacuum/analyze, ensure that entries for dropped tables are
removed promptly, and tweak the data layout to avoid storing useless struct
padding. Also improve the performance of pgstat_vacuum_tabstat(), and make
sure that autovacuum invokes it exactly once per autovac cycle rather than
multiple times or not at all. This should cure recent complaints about 8.1
showing much higher stats I/O volume than was seen in 8.0. It'd still be a
good idea to revisit the design with an eye to not re-writing the entire
stats dataset every half second ... but that would be too much to backpatch,
I fear.
cursors. Patch from Joachim Wieland, review and ediorialization by Neil
Conway. The view lists cursors defined by DECLARE CURSOR, using SPI, or
via the Bind message of the frontend/backend protocol. This means the
view does not list the unnamed portal or the portal created to implement
EXECUTE. Because we do list SPI portals, there might be more rows in
this view than you might expect if you are using SPI implicitly (e.g.
via a procedural language).
Per recent discussion on -hackers, the query string included in the
view for cursors defined by DECLARE CURSOR is based on
debug_query_string. That means it is not accurate if multiple queries
separated by semicolons are submitted as one query string. However,
there doesn't seem a trivial fix for that: debug_query_string
is better than nothing. I also changed SPI_cursor_open() to include
the source text for the portal it creates: AFAICS there is no reason
not to do this.
Update the documentation and regression tests, bump the catversion.
are two basically different kinds of scankeys, and we ought to try harder
to indicate which is used in each place in the code. I've chosen the names
"search scankey" and "insertion scankey", though you could make about
as good an argument for "operator scankey" and "comparison function
scankey".
an array of regtype, rather than an array of OIDs. This is likely to
be more useful to user, and the type OID can easily be obtained by
casting a regtype value to OID. Per suggestion from Tom.
Update the documentation and regression tests, and bump the catversion.
a va_list. Christof Petig's previous patch made this change, but neglected
to update ecpglib/descriptor.c, resulting in a compiler warning (and a
likely runtime crash) on AMD64 and PPC.
data type is unspecified or is declared to be "unknown", the type will
be inferred from the context in which the parameter is used. This was
already possible for protocol-level prepared statements.
isn't being used anywhere anymore, and there seems no point in a generic
index_keytest() routine when two out of three remaining access methods
aren't using it. Also, add a comment documenting a convention for
letting access methods define private flag bits in ScanKey sk_flags.
There are no such flags at the moment but I'm thinking about changing
btree's handling of "required keys" to use flag bits in the keys
rather than a count of required key positions. Also, if some AM did
still want SK_NEGATE then it would be reasonable to treat it as a private
flag bit.
transaction as aborted. Since we only call XactLockTableWait on XIDs
that we believe to be currently running, the odds of this code ever
actually firing are minimal. It's certainly unnecessary, since a
transaction that's not either running or committed will be presumed
aborted anyway. What's more, it's not hard to imagine scenarios where
this could result in corrupting pg_clog: for instance, if a bogus XID
somehow got passed to XactLockTableWait. I think the code probably
dates from the ancient era when we didn't have TransactionIdIsInProgress;
back then it may have been necessary, but now I think it's a waste of
cycles and potentially dangerous. Per discussion with Qingqing Zhou
and Karsten Hilbert.
permissions on the functions and operators contained in the opclass.
Since we already require superuser privilege to create an operator class,
there's no expansion-of-privilege hazard here, but if someone were to get
the idea of building an opclass containing functions that need security
restrictions, we'd better warn them off. Also, change the permission
checks from have-execute-privilege to have-ownership, and then comment
them all out since they're dead code anyway under the superuser restriction.
type definition. Because use of a type's I/O conversion functions isn't
access-checked, CREATE TYPE amounts to granting public execute permissions
on the functions, and so allowing it to anybody means that someone could
theoretically gain access to a function he's not supposed to be able to
execute. The parameter-type restrictions already enforced by CREATE TYPE
make it fairly unlikely that this oversight is meaningful in practice,
but still it seems like a good idea to plug the hole going forward.
Also, document the implicit grant just in case anybody gets the idea of
building I/O functions that might need security restrictions.
fmgr_info(), in the TopMemoryContext. I couldn't see that the code
actually leaked, but in general I think it's fragile to assume that
pfree'ing an FmgrInfo along with its fn_extra field is enough to
reclaim all the resources allocated by fmgr_info(). I changed the
code to do its allocations in a new child context of
TopMemoryContext, MbProcContext. When we want to release the
allocations we can just reset the context, which is cleaner.
our own command (or more generally, xmin = our xact and cmin >= current
command ID) should not be seen as good. Else we may try to update rows
we already updated. This error was inserted last August while fixing the
even bigger problem that the old coding wouldn't see *any* tuples inserted
by our own transaction as good. Per report from Euler Taveira de Oliveira.
rather than "return expr;" -- the latter style is used in most of the
tree. I kept the parentheses when they were necessary or useful because
the return expression was complex.
listed in the column's most-common-values statistics entry. This gives
us an exact selectivity result for the portion of the column population
represented by the MCV list, which can be a big leg up in accuracy if
that's a large fraction of the population. The heuristics involving
pattern contents and prefix are applied only to the part of the population
not included in the MCV list.
one argument at a time and then inserting the argument into a Python
list via PyList_SetItem(). This "steals" the reference to the argument:
that is, the reference to the new list member is now held by the Python
list itself. This works fine, except if an elog occurs. This causes the
function's PG_CATCH() block to be invoked, which decrements the
reference counts on both the current argument and the list of arguments.
If the elog happens to occur during the second or subsequent iteration
of the loop, the reference count on the current argument will be
decremented twice.
The fix is simple: set the local pointer to the current argument to NULL
immediately after adding it to the argument list. This ensures that the
Py_XDECREF() in the PG_CATCH() block doesn't double-decrement.
operator names. This is needed when dumping operator definitions that have
COMMUTATOR (or similar) links to operators in other schemas.
Apparently Daniel Whitter is the first person ever to try this :-(
and nail a couple more system indexes into cache. This doesn't make
any difference in normal system operation, but when forcing constant
cache resets it's difficult to get through the rules regression test
without these changes.
access information about the prepared statements that are available
in the current session. Original patch from Joachim Wieland, various
improvements by Neil Conway.
The "statement" column of the view contains the literal query string
sent by the client, without any rewriting or pretty printing. This
means that prepared statements created via SQL will be prefixed with
"PREPARE ... AS ", whereas those prepared via the FE/BE protocol will
not. That is unfortunate, but discussion on -patches did not yield an
efficient way to improve this, and there is some merit in returning
exactly what the client sent to the backend.
Catalog version bumped, regression tests updated.
Bruce Momjian
Peter Eisentraut
Tom Lane
Michael Meskes
Andrew Dunstan
Neil Conway