bucket number, so as to ensure locality of access to the index during the
insertion step. Without this, building an index significantly larger than
available RAM takes a very long time because of thrashing. On the other
hand, sorting is just useless overhead when the index does fit in RAM.
We choose to sort when the initial index size exceeds effective_cache_size.
This is a revised version of work by Tom Raney and Shreya Bhargava.
deals with the queue, including locking etc, is all in sinvaladt.c. This means
that the struct definition of the queue, and the queue pointer, are now
internal "implementation details" inside sinvaladt.c.
Per my proposal dated 25-Jun-2007 and followup discussion.
two buckets at the start, we create a number of buckets appropriate for the
estimated size of the table. This avoids a lot of expensive bucket-split
actions during initial index build on an already-populated table.
This is one of the two core ideas of Tom Raney and Shreya Bhargava's patch
to reduce hash index build time. I'm committing it separately to make it
easier for people to test the effects of this separately from the effects
of their other core idea (pre-sorting the index entries by bucket number).
This accidentally failed to fail before 8.3, because the context we were
switching back to was long-lived anyway; but it sure looks risky as can be
now. Well spotted by Pavan Deolasee.
that are reported as "equal" by wcscoll() are checked to see if they really
are bitwise equal, and are sorted per strcmp() if not. We made this happen
a couple of years ago in the regular code path, but it unaccountably got
left out of the Windows/UTF8 case (probably brain fade on my part at the
time). As in the prior set of changes, affected users may need to reindex
indexes on textual columns.
Backpatch as far as 8.2, which is the oldest release we are still supporting
on Windows.
messages if the calling transaction aborts later on. Collapsing out line
pointer redirects is a done deal as soon as we complete the page update,
so syscache *must* be notified even if the VACUUM FULL as a whole doesn't
complete. To fix, add some functionality to inval.c to allow the pending
inval messages to be sent immediately while heap_page_prune is still
running. The implementation is a bit chintzy: it will only work in the
context of VACUUM FULL. But that's all we need now, and it can always be
extended later if needed. Per my trouble report of a week ago.
(probably NULL) before exiting. Up to now it's just left the variable as it
set it, which means that after we're done processing the current client
message, ActiveSnapshot is probably pointing at garbage (because this function
is typically run in MessageContext which will get reset). There doesn't seem
to have been any code path in which that mattered before 8.3, but now the
plancache module might try to use the stale value if the next client message
is a Bind for a prepared statement that is in need of replanning. Per report
from Alex Hunsaker.
pg_listener modifications commanded by LISTEN and UNLISTEN until the end
of the current transaction. This allows us to hold the ExclusiveLock on
pg_listener until after commit, with no greater risk of deadlock than there
was before. Aside from fixing the race condition, this gets rid of a
truly ugly kludge that was there before, namely having to ignore
HeapTupleBeingUpdated failures during NOTIFY. There is a small potential
incompatibility, which is that if a transaction issues LISTEN or UNLISTEN
and then looks into pg_listener before committing, it won't see any resulting
row insertion or deletion, where before it would have. It seems unlikely
that anyone would be depending on that, though.
This patch also disallows LISTEN and UNLISTEN inside a prepared transaction.
That case had some pretty undesirable properties already, such as possibly
allowing pg_listener entries to be made for PIDs no longer present, so
disallowing it seems like a better idea than trying to maintain the behavior.
before it goes groveling through the ProcArray. In situations where the same
recently-committed transaction ID is checked repeatedly by tqual.c, this saves
a lot of shared-memory searches. And it's cheap enough that it shouldn't
hurt noticeably when it doesn't help.
Concept and patch by Simon, some minor tweaking and comment-cleanup by Tom.
than dividing them into 1GB segments as has been our longtime practice. This
requires working support for large files in the operating system; at least for
the time being, it won't be the default.
Zdenek Kotala
variables to it. More need to be converted, but I wanted to get this in
before it conflicts with too much...
Other than just centralising the text-to-int conversion for parameters,
this allows the pg_settings view to contain a list of available options
and allows an error hint to show what values are allowed.
With the addition of multiple autovacuum workers, our choices were to delete
the check, document the interaction with autovacuum_max_workers, or complicate
the check to try to hide that interaction. Since this restriction has never
been adequate to ensure backends can't run out of pinnable buffers, it doesn't
really have enough excuse to live to justify the second or third choices.
Per discussion of a complaint from Andreas Kling (see also bug #3888).
This commit also removes several documentation references to this restriction,
but I'm not sure I got them all.
pattern-examination heuristic method to purely histogram-driven selectivity at
histogram size 100, we compute both estimates and use a weighted average.
The weight put on the heuristic estimate decreases linearly with histogram
size, dropping to zero for 100 or more histogram entries.
Likewise in ltreeparentsel(). After a patch by Greg Stark, though I
reorganized the logic a bit to give the caller of histogram_selectivity()
more control.
of the generated range condition var >= 'foo' AND var < 'fop' as being less
than what eqsel() would estimate for var = 'foo'. This is intuitively
reasonable and it gets rid of the need for some entirely ad-hoc coding we
formerly used to reject bogus estimates. The basic problem here is that
if the prefix is more than a few characters long, the two boundary values
are too close together to be distinguishable by comparison to the column
histogram, resulting in a selectivity estimate of zero, which is often
not very sane. Change motivated by an example from Peter Eisentraut.
Arguably this is a bug fix, but I'll refrain from back-patching it
for the moment.
it accumulates the set of changes to be made and then applies them. It had
to accumulate the set of changes anyway to prepare a WAL record for the
pruning action, so this isn't an enormous change; the only new complexity is
to not doubly mark tuples that are visited twice in the scan. The main
advantage is that we can substantially reduce the scope of the critical
section in which the changes are applied, thus avoiding PANIC in foreseeable
cases like running out of memory in inval.c. A nice secondary advantage is
that it is now far clearer that WAL replay will actually do the same thing
that the original pruning did.
This commit doesn't do anything about the open problem that
CacheInvalidateHeapTuple doesn't have the right semantics for a CTID change
caused by collapsing out a redirect pointer. But whatever we do about that,
it'll be a good idea to not do it inside a critical section.
available output buffer when presented with corrupt input. Some testing
suggests that this slows the decompression loop about 1%, which seems an
acceptable price to pay for more robustness. (Curiously, the penalty
seems to be *less* on not-very-compressible data, which I didn't expect
since the overhead per output byte ought to be more in the literal-bytes
path.)
Patch from Zdenek Kotala. I fixed a corner case and did some renaming
of variables to make the routine more readable.
were discussed last year, but we felt it was too late in the 8.3 cycle to
change the code immediately. Specifically, the patch:
* Reduces the minimum datum size to be considered for compression from
256 to 32 bytes, as suggested by Greg Stark.
* Increases the required compression rate for compressed storage from
20% to 25%, again per Greg's suggestion.
* Replaces force_input_size (size above which compression is forced)
with a maximum size to be considered for compression. It was agreed
that allowing large inputs to escape the minimum-compression-rate
requirement was not bright, and that indeed we'd rather have a knob
that acted in the other direction. I set this value to 1MB for the
moment, but it could use some performance studies to tune it.
* Adds an early-failure path to the compressor as suggested by Jan:
if it's been unable to find even one compressible substring in the
first 1KB (parameterizable), assume we're looking at incompressible
input and give up. (Possibly this logic can be improved, but I'll
commit it as-is for now.)
* Improves the toasting heuristics so that when we have very large
fields with attstorage 'x' or 'e', we will push those out to toast
storage before considering inline compression of shorter fields.
This also responds to a suggestion of Greg's, though my original
proposal for a solution was a bit off base because it didn't fix
the problem for large 'e' fields.
There was some discussion in the earlier threads of exposing some
of the compression knobs to users, perhaps even on a per-column
basis. I have not done anything about that here. It seems to me
that if we are changing around the parameters, we'd better get some
experience and be sure we are happy with the design before we set
things in stone by providing user-visible knobs.
TopMemoryContext, rather than scattered through executor per-query contexts.
This poses no danger of memory leak since the ResourceOwner mechanism
guarantees release of no-longer-needed items. It is needed because the
per-query context might already be released by the time we try to clean up
the hash scan list. Report by ykhuang, diagnosis by Heikki.
Back-patch to 8.0, where the ResourceOwner-based cleanup was introduced.
The given test case does not fail before 8.2, probably because we rearranged
transaction abort processing somehow; but this coding is undoubtedly risky
so I'll patch 8.0 and 8.1 anyway.
a unused memory holes in tsquery.
Per report by Richard Huxton <dev@archonet.com>.
It was working well because in fact tsquery->size is not used for any
kind of operation except comparing tsqueries. So, in HEAD it's enough to
fix to_tsquery function, but for previous version it's needed to
remove optimization in CompareTSQ to prevent requirement of renew all
stored tsquery.
caches that we don't actually need to touch. This saves some trivial
number of cycles and avoids certain cases of deadlock when doing concurrent
VACUUM FULL on system catalogs. Per report from Gavin Roy.
Backpatch to 8.2. In earlier versions, CatalogCacheInitializeCache didn't
lock the relation so there's no deadlock risk (though that certainly had
plenty of risks of its own).
temporary table; we can't support that because there's no way to clean up the
source backend's internal state if the eventual COMMIT PREPARED is done by
another backend. This was checked correctly in 8.1 but I broke it in 8.2 :-(.
Patch by Heikki Linnakangas, original trouble report by John Smith.
"struct varlena" would be at least word-aligned. Per buildfarm results
from gypsy_moth. I did a little bit of trawling for other instances of
this coding pattern, and didn't find any; but if we turn up any more
of them I think we'd better revert the "char [4]" patch and find another
way of making tuptoaster.c alignment-safe.
to explicitly cast the output back to char before comparing it to a char
value, else we get the wrong result for high-bit-set characters. Found by
Rolf Jentsch. Also, fix several places where <ctype.h> functions were being
called without casting the argument to unsigned char; this is likewise
unportable, but we keep making that mistake :-(. These found by buildfarm
member salamander, which I will desperately miss if it ever goes belly-up.
left in the code though it was not meant to be provided. It represents a
security hole because unprivileged users could use it to look at (at least the
first line of) any file readable by the backend. Fortunately, this is only
possible if the backend was built with XML support, so the damage is at least
mitigated; and 8.3 probably hasn't propagated into any security-critical uses
yet anyway. Per report from Sergey Burladyan.
is also licensed to put a local variable declared that way at an unaligned
address. Which will not work if the variable is then manipulated with
SET_VARSIZE or other macros that assume alignment. So the previous patch
is not an unalloyed good, but on balance I think it's still a win, since
we have very few places that do that sort of thing. Fix the one place in
tuptoaster.c that does it. Per buildfarm results from gypsy_moth
(I'm a bit surprised that only one machine showed a failure).
"multi_call_ctx" to be a distinct sub-context of the EState's per-query
context, and delete the multi_call_ctx as soon as the SRF finishes
execution. This avoids leaking SRF memory until the end of the current
query, which is particularly egregious when the SRF is scanned
multiple times. This change also fixes a leak of the fields of the
AttInMetadata struct in shutdown_MultiFuncCall().
Also fix a leak of the SRF result TupleDesc when rescanning a
FunctionScan node. The TupleDesc is allocated in the per-query context
for every call to ExecMakeTableFunctionResult(), so we should free it
after calling that function. Since the SRF might choose to return
a non-expendable TupleDesc, we only free the TupleDesc if it is
not being reference-counted.
Backpatch to 8.3 and 8.2 stable branches.
So only rebuild when a makefile changes (which presumably defines the
file list somewhere), and only touch the file if an object changed. The
touch is necessary so the parent make knows something changed and
ultimately rebuilds postgres.
values into \nnn octal escape sequences. When the database encoding is
multibyte this is *necessary* to avoid generating invalidly encoded text.
Even in a single-byte encoding, the old behavior seems very hazardous ---
consider for example what happens if the text is transferred to another
database with a different encoding. Decoding would then yield some other
bytea value than what was encoded, which is surely undesirable. Per gripe
from Hernan Gonzalez.
Backpatch to 8.3, but not further. This is a bit of a judgment call, but I
make it on these grounds: pre-8.3 we don't really have much encoding safety
anyway because of the convert() function family, and we would also have much
higher risk of breaking existing apps that may not be expecting this behavior.
8.3 is still new enough that we can probably get away with making this change
in the function's behavior.
(then it means 2000 AD). Formerly we silently interpreted this as 1 BC,
which at best is unwarranted familiarity with the implementation.
It's barely possible that some app somewhere expects the old behavior,
though, so we won't back-patch this into existing release branches.
Formerly, DecodeDate attempted to verify the day-of-the-month exactly, but
it was under the misapprehension that it would know whether we were looking
at a BC year or not. In reality this check can't be made until the calling
function (eg DecodeDateTime) has processed all the fields. So, split the
BC adjustment and validity checks out into a new function ValidateDate that
is called only after processing all the fields. In passing, this patch
makes DecodeTimeOnly work for BC inputs, which it never did before.
(The historical veracity of all this is nonexistent, of course, but if
we're going to say we support proleptic Gregorian calendar then we should
do it correctly. In any case the unpatched code is broken because it could
emit dates that it would then reject on re-inputting.)
Per report from Bernd Helmle. Back-patch as far as 8.0; in 7.x we were
not using our own calendar support and so this seems a bit too risky
to put into 7.4.
Tom Lane
Alvaro Herrera
Magnus Hagander
Andrew Dunstan
Teodor Sigaev
Bruce Momjian
Peter Eisentraut
Neil Conway