rightmost on its tree level, we split 2/3 to the left and 1/3 to the
new right page, rather than the even split we use elsewhere. The idea
is that when faced with a steadily increasing series of inserted keys
(such as sequence or timestamp values), we'll end up with a btree that's
about 2/3ds full not 1/2 full, which is much closer to the desired
steady-state load for a btree. Per suggestion from Ann Harrison of
IBPhoenix.
existing lock manager and spinlocks: it understands exclusive vs shared
lock but has few other fancy features. Replace most uses of spinlocks
with lightweight locks. All remaining uses of spinlocks have very short
lock hold times (a few dozen instructions), so tweak spinlock backoff
code to work efficiently given this assumption. All per my proposal on
pghackers 26-Sep-01.
two performance improvements. I put an explanation of the
changes at
http://cs1.cs.nyu.edu/been/postgres-rtree.html
The performance improvements are quite significant.
All the changes are in the file src/backend/access/rtree/rtree.c
I was working with the 7.1.3 code.
I'm including the diff output as an attachment.
Kenneth Been
Define a new function, GetCurrentTransactionStartTimeUsec() to get the time
to this precision.
Allow now() and timestamp 'now' to use this higher precision result so
we now have fractional seconds in this "constant".
Add timestamp without time zone type.
Move previous timestamp type to timestamp with time zone.
Accept another ISO variant for date/time values: yyyy-mm-ddThh:mm:ss
(note the "T" separating the day from hours information).
Remove 'current' from date/time types; convert to 'now' in input.
Separate time and timetz regression tests.
Separate timestamp and timestamptz regression test.
Standardize on %X/%X as the formatting for XLOG position display --- we
had a couple of different formats before, and none of 'em were as useful
as hex offsets IMHO.
buffer manager with 'pg_clog', a specialized access method modeled
on pg_xlog. This simplifies startup (don't need to play games to
open pg_log; among other things, OverrideTransactionSystem goes away),
should improve performance a little, and opens the door to recycling
commit log space by removing no-longer-needed segments of the commit
log. Actual recycling is not there yet, but I felt I should commit
this part separately since it'd still be useful if we chose not to
do transaction ID wraparound.
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.
default, but OIDS are removed from many system catalogs that don't need them.
Some interesting side effects: TOAST pointers are 20 bytes not 32 now;
pg_description has a three-column key instead of one.
Bugs fixed in passing: BINARY cursors work again; pg_class.relhaspkey
has some usefulness; pg_dump dumps comments on indexes, rules, and
triggers in a valid order.
initdb forced.
(as proposed in http://fts.postgresql.org/db/mw/msg.html?mid=1028327)
2. support for 'pass-by-value' arguments - to test this
we used special opclass for int4 with values in range [0-2^15]
More testing will be done after resolving problem with
index_formtuple and implementation of B-tree using GiST
3. small patch to contrib modules (seg,cube,rtree_gist,intarray) -
mark functions as 'isstrict' where needed.
Oleg Bartunov
rather than deleting them only to have to create more. Steady state
is 2*CHECKPOINT_SEGMENTS + WAL_FILES + 1 segment files, which will
simply be renamed rather than constantly deleted and recreated.
To make this safe, added current XLOG file/offset number to page
header of XLOG pages, so that an un-overwritten page from an old
incarnation of a logfile can be reliably told from a valid page.
This change means that if you try to restart postmaster in a CVS-tip
database after installing the change, you'll get a complaint about
bad XLOG page magic number. If you don't want to initdb, run
contrib/pg_resetxlog (and be sure you shut down the old postmaster
cleanly).
in GetSnapshotData, GetNewTransactionId, CommitTransaction, AbortTransaction,
etc. Correct race condition in transaction status testing in
HeapTupleSatisfiesVacuum --- this wasn't important for old VACUUM with
exclusive lock on its table, but it sure is important now. All per
pghackers discussion 7/11/01 and 7/12/01.
per previous discussion on pghackers. Most of the duplicate code in
different AMs' ambuild routines has been moved out to a common routine
in index.c; this means that all index types now do the right things about
inserting recently-dead tuples, etc. (I also removed support for EXTEND
INDEX in the ambuild routines, since that's about to go away anyway, and
it cluttered the code a lot.) The retail indextuple deletion routines have
been replaced by a "bulk delete" routine in which the indexscan is inside
the access method. I haven't pushed this change as far as it should go yet,
but it should allow considerable simplification of the internal bookkeeping
for deletions. Also, add flag columns to pg_am to eliminate various
hardcoded tests on AM OIDs, and remove unused pg_am columns.
Fix rtree and gist index types to not attempt to store NULLs; before this,
gist usually crashed, while rtree managed not to crash but computed wacko
bounding boxes for NULL entries (which might have had something to do with
the performance problems we've heard about occasionally).
Add AtEOXact routines to hash, rtree, and gist, all of which have static
state that needs to be reset after an error. We discovered this need long
ago for btree, but missed the other guys.
Oh, one more thing: concurrent VACUUM is now the default.
validity checking rules for VACUUM. Make some other rearrangements of the
VACUUM code to allow more code to be shared between full and lazy VACUUM.
Minor code cleanups and added comments for TransactionId manipulations.
stub) into the rest of the system. Adopt a cleaner approach to preventing
deadlock in concurrent heap_updates: allow RelationGetBufferForTuple to
select any page of the rel, and put the onus on it to lock both buffers
in a consistent order. Remove no-longer-needed isExtend hack from
API of ReleaseAndReadBuffer.
do anything yet, but it has the necessary connections to initialization
and so forth. Make some gestures towards allowing number of blocks in
a relation to be BlockNumber, ie, unsigned int, rather than signed int.
(I doubt I got all the places that are sloppy about it, yet.) On the
way, replace the hardwired NLOCKS_PER_XACT fudge factor with a GUC
variable.
IS TRUE, etc, with some degree of verisimilitude. Split out
selectivity support functions from builtins.h into a new header
file selfuncs.h, so as to reduce the number of header files builtins.h
must depend on. Fix a few missing inclusions exposed thereby.
From Joe Conway, with some kibitzing from Tom Lane.
SI messages now include the relevant database OID, so that operations
in one database do not cause useless cache flushes in backends attached
to other databases. Declare SI messages properly using a union, to
eliminate the former assumption that Oid is the same size as int or Index.
Rewrite the nearly-unreadable code in inval.c, and document it better.
Arrange for catcache flushes at end of command/transaction to happen before
relcache flushes do --- this avoids loading a new tuple into the catcache
while setting up new relcache entry, only to have it be flushed again
immediately.
pg_database now has unique indexes on oid and on datname.
pg_shadow now has unique indexes on usename and on usesysid.
pg_am now has unique index on oid.
pg_opclass now has unique index on oid.
pg_amproc now has unique index on amid+amopclaid+amprocnum.
Remove pg_rewrite's unnecessary index on oid, delete unused RULEOID syscache.
Remove index on pg_listener and associated syscache for performance reasons
(caching rows that are certain to change before you need 'em again is
rather pointless).
Change pg_attrdef's nonunique index on adrelid into a unique index on
adrelid+adnum.
Fix various incorrect settings of pg_class.relisshared, make that the
primary reference point for whether a relation is shared or not.
IsSharedSystemRelationName() is now only consulted to initialize relisshared
during initial creation of tables and indexes. In theory we might now
support shared user relations, though it's not clear how one would get
entries for them into pg_class &etc of multiple databases.
Fix recently reported bug that pg_attribute rows created for an index all have
the same OID. (Proof that non-unique OID doesn't matter unless it's
actually used to do lookups ;-))
There's no need to treat pg_trigger, pg_attrdef, pg_relcheck as bootstrap
relations. Convert them into plain system catalogs without hardwired
entries in pg_class and friends.
Unify global.bki and template1.bki into a single init script postgres.bki,
since the alleged distinction between them was misleading and pointless.
Not to mention that it didn't work for setting up indexes on shared
system relations.
Rationalize locking of pg_shadow, pg_group, pg_attrdef (no need to use
AccessExclusiveLock where ExclusiveLock or even RowExclusiveLock will do).
Also, hold locks until transaction commit where necessary.
appropriate pin-count manipulation, and instead use ReleaseAndReadBuffer.
Make use of the fact that the passed-in buffer (if there is one) must
be pinned to avoid grabbing the bufmgr spinlock when we are able to
return this same buffer. Eliminate unnecessary 'previous tuple' and
'next tuple' fields of HeapScanDesc and IndexScanDesc, thereby removing
a whole lot of bookkeeping from heap_getnext() and related routines.
checkpoint's redo pointer, not its undo pointer, per discussion in
pghackers a few days ago. No point in hanging onto undo information
until we have the ability to do something with it --- and this solves
a rather large problem with log space for long-running transactions.
Also, change all calls of write() to detect the case where write
returned a count less than requested, but failed to set errno.
Presume that this situation indicates ENOSPC, and give the appropriate
error message, rather than a random message associated with the previous
value of errno.
report on old-style functions invoked by RI triggers. We had a number of
other places that were being sloppy about which memory context FmgrInfo
subsidiary data will be allocated in. Turns out none of them actually
cause a problem in 7.1, but this is for arcane reasons such as the fact
that old-style triggers aren't supported anyway. To avoid getting burnt
later, I've restructured the trigger support so that we don't keep trigger
FmgrInfo structs in relcache memory. Some other related cleanups too:
it's not really necessary to call fmgr_info at all while setting up
the index support info in relcache entries, because those ScanKeyEntry
structs are never used to invoke the functions. This should speed up
relcache initialization a tiny bit.
Python) to support shared extension modules, I have learned that Guido
prefers the style of the attached patch to solve the above problem.
I feel that this solution is particularly appropriate in this case
because the following:
PglargeType
PgType
PgQueryType
are already being handled in the way that I am proposing for PgSourceType.
Jason Tishler
PageGetFreeSpace() was being called while not holding the buffer lock, which
not only could yield a garbage answer, but even if it's the right answer there
might be less space available after we reacquire the buffer lock.
Also repair potential deadlock introduced by my recent performance improvement
in RelationGetBufferForTuple(): it was possible for two heap_updates to try to
lock two buffers in opposite orders. The fix creates a global rule that
buffers of a single heap relation should be locked in decreasing block number
order. Currently, this only applies to heap_update; VACUUM can get away with
ignoring the rule since it holds exclusive lock on the whole relation anyway.
However, if we try to implement a VACUUM that can run in parallel with other
transactions, VACUUM will also have to obey the lock order rule.
with many NULLs ( inserting of NULL into indexed field cause
ERROR: MemoryContextAlloc: invalid request size)
As a workaround 'vacuum analyze' could be used.
This patch resolves the problem, please upply to 7.1.1 sources and
current cvs tree.
Oleg Bartunov
when we need to move to a new page; as long as we can insert the new
tuple on the same page as before, we only need LockBuffer and not the
expensive stuff. Also, twiddle bufmgr interfaces to avoid redundant
lseeks in RelationGetBufferForTuple and BufferAlloc. Successive inserts
now require one lseek per page added, rather than one per tuple with
several additional ones at each page boundary as happened before.
Lock contention when multiple backends are inserting in same table
is also greatly reduced.
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.
give consistent results for all datatypes. Types float4, float8, and
numeric were broken for NaN values; abstime, timestamp, and interval
were broken for INVALID values; timetz was just plain broken (some
possible pairs of values were neither < nor = nor >). Also clean up
text, bpchar, varchar, and bit/varbit to eliminate duplicate code and
thereby reduce the probability of similar inconsistencies arising in
the future.
O_SYNC, or O_DSYNC (as available on a given platform). Add GUC parameter
to control sync method.
Also, add defense to XLogWrite to prevent it from going nuts if passed
a target write position that's past the end of the buffers so far filled
by XLogInsert.
detect case that next page in log came from an older run than the prior
page. This avoids the necessity to re-zero the log after recovery from
a crash, which is good because we need not risk destroying valuable log
information.
This forces another initdb since yesterday :-(. Need to get that log
reset utility done...
* Store two past checkpoint locations, not just one, in pg_control.
On startup, we fall back to the older checkpoint if the newer one
is unreadable. Also, a physical copy of the newest checkpoint record
is kept in pg_control for possible use in disaster recovery (ie,
complete loss of pg_xlog). Also add a version number for pg_control
itself. Remove archdir from pg_control; it ought to be a GUC
parameter, not a special case (not that it's implemented yet anyway).
* Suppress successive checkpoint records when nothing has been entered
in the WAL log since the last one. This is not so much to avoid I/O
as to make it actually useful to keep track of the last two
checkpoints. If the things are right next to each other then there's
not a lot of redundancy gained...
* Change CRC scheme to a true 64-bit CRC, not a pair of 32-bit CRCs
on alternate bytes. Polynomial borrowed from ECMA DLT1 standard.
* Fix XLOG record length handling so that it will work at BLCKSZ = 32k.
* Change XID allocation to work more like OID allocation. (This is of
dubious necessity, but I think it's a good idea anyway.)
* Fix a number of minor bugs, such as off-by-one logic for XLOG file
wraparound at the 4 gig mark.
* Add documentation and clean up some coding infelicities; move file
format declarations out to include files where planned contrib
utilities can get at them.
* Checkpoint will now occur every CHECKPOINT_SEGMENTS log segments or
every CHECKPOINT_TIMEOUT seconds, whichever comes first. It is also
possible to force a checkpoint by sending SIGUSR1 to the postmaster
(undocumented feature...)
* Defend against kill -9 postmaster by storing shmem block's key and ID
in postmaster.pid lockfile, and checking at startup to ensure that no
processes are still connected to old shmem block (if it still exists).
* Switch backends to accept SIGQUIT rather than SIGUSR1 for emergency
stop, for symmetry with postmaster and xlog utilities. Clean up signal
handling in bootstrap.c so that xlog utilities launched by postmaster
will react to signals better.
* Standalone bootstrap now grabs lockfile in target directory, as added
insurance against running it in parallel with live postmaster.
succeeds or not. Revise rtree page split algorithm to take care about
making a feasible split --- ie, will the incoming tuple actually fit?
Failure to make a feasible split, combined with failure to notice the
failure, account for Jim Stone's recent bug report. I suspect that
hash and gist indices may have the same type of bug, but at least now
we'll get error messages rather than silent failures if so. Also clean
up rtree code to use Datum rather than char* where appropriate.
only if at least N other backends currently have open transactions. This
is not a great deal of intelligence about whether a delay might be
profitable ... but it beats no intelligence at all. Note that the default
COMMIT_DELAY is still zero --- this new code does nothing unless that
setting is changed.
Also, mark ENABLEFSYNC as a system-wide setting. It's no longer safe to
allow that to be set per-backend, since we may be relying on some other
backend's fsync to have synced the WAL log.
waste of cycles on single-CPU machines, and of dubious utility on multi-CPU
machines too.
Tweak s_lock_stuck so that caller can specify timeout interval, and
increase interval before declaring stuck spinlock for buffer locks and XLOG
locks.
On systems that have fdatasync(), use that rather than fsync() to sync WAL
log writes. Ensure that WAL file is entirely allocated during XLogFileInit.
compressed storage works perfectly well. Might as well have a coherent
strategy for applying it, rather than the haphazard store-what-you-get
approach that was in the code before. The strategy I've set up here is
to attempt compression of any compressible index value exceeding
BLCKSZ/16, or about 500 bytes by default.
allocated by plan nodes are not leaked at end of query. This doesn't
really matter for normal queries, but it sure does for queries invoked
repetitively inside SQL functions. Clean up some other grotty code
associated with tupdescs, and fix a few other memory leaks exposed by
tests with simple SQL functions.
and new root page if old root one was splitted but new root page
wasn't created.
New code is protected by FixBTree bool flag setted to FALSE, so
nothing should be affected by this untested approach.
bothering to check the return value --- which meant that in case the
update or delete failed because of a concurrent update, you'd not find
out about it, except by observing later that the transaction produced
the wrong outcome. There are now subroutines simple_heap_update and
simple_heap_delete that should be used anyplace that you're not prepared
to do the full nine yards of coping with concurrent updates. In
practice, that seems to mean absolutely everywhere but the executor,
because *noplace* else was checking.
are treated more like 'cancel' interrupts: the signal handler sets a
flag that is examined at well-defined spots, rather than trying to cope
with an interrupt that might happen anywhere. See pghackers discussion
of 1/12/01.
are now critical sections, so as to ensure die() won't interrupt us while
we are munging shared-memory data structures. Avoid insecure intermediate
states in some code that proc_exit will call, like palloc/pfree. Rename
START/END_CRIT_CODE to START/END_CRIT_SECTION, since that seems to be
what people tend to call them anyway, and make them be called with () like
a function call, in hopes of not confusing pg_indent.
I doubt that this is sufficient to make SIGTERM safe anywhere; there's
just too much code that could get invoked during proc_exit().
1. Support of variable size keys - new algorithm of insertion to tree
(GLI - gist layrered insertion). Previous algorithm was implemented
as described in paper by Joseph M. Hellerstein et.al
"Generalized Search Trees for Database Systems". This (old)
algorithm was not suitable for variable size keys and could be
not effective ( walking up-down ) in case of multiple levels split
Bug fixed:
1. fixed bug in gistPageAddItem - key values were written to disk
uncompressed. This caused failure if decompression function
does real job.
2. NULLs handling - we keep NULLs in tree. Right way is to remove them,
but we don't know how to inform vacuum about index statistics. This is
just cosmetic warning message (like in case with R-Tree),
but I'm not sure how to recognize real problem if we remove NULLs
and suppress this warning as Tom suggested.
3. various memory leaks
This work was done by Teodor Sigaev (teodor@stack.net) and
Oleg Bartunov (oleg@sai.msu.su).
- no more elog(STOP) in StartupXLOG();
- both checkpoint' undo & redo are used to define
oldest on-line log file.
2. Ability to pre-allocate a few log files at checkpoint time
(wal_files option). Off by default.
before calling RelationInvalidateHeapTuple(), which is bad because the
latter needs to look at the tuple data, which is in the shared disk
buffer. If another backend manages to recycle the buffer while this
is going on, we will compute the wrong hashindex for the tuple or
maybe even crash outright. Must hold buffer refcount until afterwards.
(This bug is not in 7.0.*; seems to be have introduced during WAL changes.)
starting a new hashtable search no longer clobbers any other search
active anywhere in the system. Fix RelationCacheInvalidate() so that
it will not crash or go into an infinite loop if invoked recursively,
as for example by a second SI Reset message arriving while we are still
processing a prior one.
assume that TAS() will always succeed the first time, even if the lock
is known to be free. Also, make sure that code will eventually time out
and report a stuck spinlock, rather than looping forever. Small cleanups
in s_lock.h, too.
Tom Lane
Bruce Momjian
Thomas G. Lockhart
Peter Eisentraut
Hiroshi Inoue
Jan Wieck
Vadim B. Mikheev
Marc G. Fournier