The original design used an array of offsets into the variable-length
portion of a JSONB container. However, such an array is basically
uncompressible by simple compression techniques such as TOAST's LZ
compressor. That's bad enough, but because the offset array is at the
front, it tended to trigger the give-up-after-1KB heuristic in the TOAST
code, so that the entire JSONB object was stored uncompressed; which was
the root cause of bug #11109 from Larry White.
To fix without losing the ability to extract a random array element in O(1)
time, change this scheme so that most of the JEntry array elements hold
lengths rather than offsets. With data that's compressible at all, there
tend to be fewer distinct element lengths, so that there is scope for
compression of the JEntry array. Every N'th entry is still an offset.
To determine the length or offset of any specific element, we might have
to examine up to N preceding JEntrys, but that's still O(1) so far as the
total container size is concerned. Testing shows that this cost is
negligible compared to other costs of accessing a JSONB field, and that
the method does largely fix the incompressible-data problem.
While at it, rearrange the order of elements in a JSONB object so that
it's "all the keys, then all the values" not alternating keys and values.
This doesn't really make much difference right at the moment, but it will
allow providing a fast path for extracting individual object fields from
large JSONB values stored EXTERNAL (ie, uncompressed), analogously to the
existing optimization for substring extraction from large EXTERNAL text
values.
Bump catversion to denote the incompatibility in on-disk format.
We will need to fix pg_upgrade to disallow upgrading jsonb data stored
with 9.4 betas 1 and 2.
Heikki Linnakangas and Tom Lane
Without this fix, parallel restore of a schema-only dump can deadlock,
because when the dump is schema-only, the dependency will still be
pointing at the TABLE item rather than the TABLE DATA item.
Robert Haas and Tom Lane
We removed a similar ban on this in json_object recently, but the ban in
datum_to_json was left, which generate4d sprutious errors in othee json
generators, notable json_build_object.
Along the way, add an assertion that datum_to_json is not passed a null
key. All current callers comply with this rule, but the assertion will
catch any possible future misbehaviour.
Some compilers don't automatically search the current directory for
included files. 9cc2c182fc fixed that for builds from tarballs by
adding an include to the source directory. But that doesn't work when
the scanner is generated in the VPATH directory. Use the same search
path as the other parsers in the tree.
One compiler that definitely was affected is solaris' sun cc.
Backpatch to 9.1 which introduced using an actual parser for
replication commands.
This function created new Vars with varno different from varnoold, which
is a condition that should never prevail before setrefs.c does the final
variable-renumbering pass. The created Vars could not be seen as equal()
to normal Vars, which among other things broke equivalence-class processing
for them. The consequences of this were indeed visible in the regression
tests, in the form of failure to propagate constants as one would expect.
I stumbled across it while poking at bug #11457 --- after intentionally
disabling join equivalence processing, the security-barrier regression
tests started falling over with fun errors like "could not find pathkey
item to sort", because of failure to match the corrupted Vars to normal
ones.
When the number of allowed iterations is limited (either a "?" quantifier
or a bound expression), the last sub-match has to reach to the end of the
target string. The previous coding here first tried the shortest possible
match (one character, usually) and then gave up and back-tracked if that
didn't work, typically leading to failure to match overall, as shown in
bug #11478 from Christoph Berg. The minimum change to fix that would be to
not decrement k before "goto backtrack"; but that would be a pretty stupid
solution, because we'd laboriously try each possible sub-match length
before finally discovering that only ending at the end can work. Instead,
force the sub-match endpoint limit up to the end for even the first
shortest() call if we cannot have any more sub-matches after this one.
Bug introduced in my rewrite that added the iterdissect logic, commit
173e29aa5d. The shortest-first search code
was too closely modeled on the longest-first code, which hasn't got this
issue since it tries a match reaching to the end to start with anyway.
Back-patch to all affected branches.
Per discussion in bug #11350, log ALTER SYSTEM commands at the
log_statement=ddl level, rather than at the log_statement=all level.
Pointed out by Tomonari Katsumata.
Back-patch to 9.4 where ALTER SYSTEM was introduced.
While withCheckOption exprs had been handled in many cases by
happenstance, they need to be handled during set_plan_references and
more specifically down in set_plan_refs for ModifyTable plan nodes.
This is to ensure that the opfuncid's are set for operators referenced
in the withCheckOption exprs.
Identified as an issue by Thom Brown
Patch by Dean Rasheed
Back-patch to 9.4, where withCheckOption was introduced.
This has been broken since commit af7914c662,
which added the EXPLAIN (TIMING) option. Although that commit included
updates to auto_explain, they evidently weren't tested very carefully,
because the code failed to print node timings even when it should, due to
failure to set es.timing in the ExplainState struct. Reported off-list by
Neelakanth Nadgir of Salesforce.
In passing, clean up the documentation for auto_explain's options a
little bit, including re-ordering them into what seems to me a more
logical order.
x86's memory barrier assembly was marked as clobbering "memory" but
not "cc" even though 'addl' sets various flags. As it turns out gcc on
x86 implicitly assumes "cc" on every inline assembler statement, so
it's not a bug. But as that's poorly documented and might get copied
to architectures or compilers where that's not the case, it seems
better to be precise.
Discussion: 20140919100016.GH4277@alap3.anarazel.de
To keep the code common, backpatch to 9.2 where explicit memory
barriers were introduced.
Add some quotes in the makefile snippet that creates the temporary
installation, so that it can handle spaces in the directory name and
possibly some other oddities.
They were marked to return a boolean, but they actually return a
GinTernaryValue, which is more like a "char". It makes no practical
difference, as the triConsistent functions cannot be called directly from
SQL because they have "internal" arguments, but this nevertheless seems
more correct.
Also fix the GinTernaryValue name in the documentation. I renamed the enum
earlier, but neglected the docs.
Alexander Korotkov. This is new in 9.4, so backpatch there.
In psql, expanded mode was not being displayed correctly when using
the normal ascii or unicode linestyles and border set to '3'. Now,
per the documentation, border '3' is really only sensible for HTML
and LaTeX formats, however, that's no excuse for ascii/unicode to
break in that case, and provisions had been made for psql to cleanly
handle this case (and it did, in non-expanded mode).
This was broken when ascii/unicode was initially added a good five
years ago because print_aligned_vertical_line wasn't passed in the
border setting being used by print_aligned_vertical but instead was
given the whole printTableContent. There really isn't a good reason
for vertical_line to have the entire printTableContent structure, so
just pass in the printTextFormat and border setting (similar to how
this is handled in horizontal_line).
Pointed out by Pavel Stehule, fix by me.
Back-patch to all currently-supported versions.
This patch allows us to execute ALTER SYSTEM RESET command to
remove the configuration entry from postgresql.auto.conf.
Vik Fearing, reviewed by Amit Kapila and me.
The code that tried to split a page at 75/25 ratio, when appending to the
end of an index, was buggy in two ways. First, there was a silly typo that
caused it to just fill the left page as full as possible. But the logic as
it was intended wasn't correct either, and would actually have given a ratio
closer to 60/40 than 75/25.
Gaetano Mendola spotted the typo. Backpatch to 9.4, where this code was added.
The code for raising a NUMERIC value to an integer power wasn't very
careful about large powers. It got an outright wrong answer for an
exponent of INT_MIN, due to failure to consider overflow of the Abs(exp)
operation; which is fixable by using an unsigned rather than signed
exponent value after that point. Also, even though the number of
iterations of the power-computation loop is pretty limited, it's easy for
the repeated squarings to result in ridiculously enormous intermediate
values, which can take unreasonable amounts of time/memory to process,
or even overflow the internal "weight" field and so produce a wrong answer.
We can forestall misbehaviors of that sort by bailing out as soon as the
weight value exceeds what will fit in int16, since then the final answer
must overflow (if exp > 0) or underflow (if exp < 0) the packed numeric
format.
Per off-list report from Pavel Stehule. Back-patch to all supported
branches.
When running vacuumdb --analyze-in-stages --all, it needs to run the
first stage across all databases before the second one, instead of
running all stages in a database before processing the next one.
Also respect the --quiet option with --analyze-in-stages.
The previous coding first generated a uniform random value between 0.0 and
1.0, then converted that to an integer between 1 and 10000, and divided that
again by 10000. Those conversions are unnecessary; we can use the double
value that pg_erand48() returns directly. While we're at it, put the logic
into a helper function, getPoissonRand().
The largest delay generated by the old coding was about 9.2 times the
average, because of the way the uniformly distributed value used for the
calculation was truncated to 1/10000 granularity. The new coding doesn't
have such clamping. With my laptop's DBL_MIN value, the maximum delay with
the new coding is about 700x the average. That seems acceptable - any
reasonable pgbench session should last long enough to average that out.
Backpatch to 9.4.
The reported latency values now include the "schedule lag" time, that is,
the time between the transaction's scheduled start time and the time it
actually started. This relates better to a model where requests arrive at a
certain rate, and we are interested in the response time to the end user or
application, rather than the response time of the database itself.
Also, when --rate is used, include the schedule lag time in the log output.
The --rate option is new in 9.4, so backpatch to 9.4. It seems better to
make this change in 9.4, while we're still in the beta period, than ship a
9.4 version that calculates the values differently than 9.5.
Apparently, older versions of "prove" (couldn't identify the exact
version from the changelog) don't look into the t/ directory for tests
by default, so specify it explicitly.
The code I added in commit f343a880d5 was
careless about preserving AND/OR flatness: it could create a structure with
an OR node directly underneath another one. That breaks an assumption
that's fairly important for planning efficiency, not to mention triggering
various Asserts (as reported by Benjamin Smith). Add a trifle more logic
to handle the case properly.
Some Sparc CPUs can be run in various coherence models, ranging from
RMO (relaxed) over PSO (partial) to TSO (total). Solaris has always
run CPUs in TSO mode while in userland, but linux didn't use to and
the various *BSDs still don't. Unfortunately the sparc TAS/S_UNLOCK
were only correct under TSO. Fix that by adding the necessary memory
barrier instructions. On sparcv8+, which should be all relevant CPUs,
these are treated as NOPs if the current consistency model doesn't
require the barriers.
Discussion: 20140630222854.GW26930@awork2.anarazel.de
Will be backpatched to all released branches once a few buildfarm
cycles haven't shown up problems. As I've no access to sparc, this is
blindly written.
Now that ALTER TABLE .. ALL IN TABLESPACE has replaced the previous
ALTER TABLESPACE approach, it makes sense to move the calls down in
to ProcessUtilitySlow where the rest of ALTER TABLE is handled.
This also means that event triggers will support ALTER TABLE .. ALL
(which was the impetus for the original change, though it has other
good qualities also).
Álvaro Herrera
Back-patch to 9.4 as the original rework was.
psql's \s (print command history) doesn't work at all with recent libedit
versions when printing to the terminal, because libedit tries to do an
fchmod() on the target file which will fail if the target is /dev/tty.
(We'd already noted this in the context of the target being /dev/null.)
Even before that, it didn't work pleasantly, because libedit likes to
encode the command history file (to ensure successful reloading), which
renders it nigh unreadable, not to mention significantly different-looking
depending on exactly which libedit version you have. So let's forget using
write_history() for this purpose, and instead print the data ourselves,
using logic similar to that used to iterate over the history for newline
encoding/decoding purposes.
While we're at it, insert the ability to use the pager when \s is printing
to the terminal. This has been an acknowledged shortcoming of \s for many
years, so while you could argue it's not exactly a back-patchable bug fix
it still seems like a good improvement. Anyone who's seriously annoyed
at this can use "\s /dev/tty" or local equivalent to get the old behavior.
Experimentation with this showed that the history iteration logic was
actually rather broken when used with libedit. It turns out that with
libedit you have to use previous_history() not next_history() to advance
to more recent history entries. The easiest and most robust fix for this
seems to be to make a run-time test to verify which function to call.
We had not noticed this because libedit doesn't really need the newline
encoding logic: its own encoding ensures that command entries containing
newlines are reloaded correctly (unlike libreadline). So the effective
behavior with recent libedits was that only the oldest history entry got
newline-encoded or newline-decoded. However, because of yet other bugs in
history_set_pos(), some old versions of libedit allowed the existing loop
logic to reach entries besides the oldest, which means there may be libedit
~/.psql_history files out there containing encoded newlines in more than
just the oldest entry. To ensure we can reload such files, it seems
appropriate to back-patch this fix, even though that will result in some
incompatibility with older psql versions (ie, multiline history entries
written by a psql with this fix will look corrupted to a psql without it,
if its libedit is reasonably up to date).
Stepan Rutz and Tom Lane
The old claim is from my commit d06ebdb8d3 of
2000-07-17, but it seems to have been a plain old thinko; sum(float4) has
been distinct from sum(float8) since Berkeley days. Noted by KaiGai Kohei.
While at it, mention the existence of sum(money), which is also of
embarrassingly ancient vintage.
Update the tab completion for the changes made in
3c4cf08087, which rework 'MOVE ALL' to be
'ALTER .. ALL IN TABLESPACE'.
Fujii Masao
Back-patch to 9.4, as the original change was.
The link to the NIST web page about DES standards leads to nowhere, and
according to archive.org has been forwarded to an unrelated page for
many years. Therefore, just remove that link. More up to date
information can be found via Wikipedia, for example.
The syntax summary previously failed to clarify that the first
argument is also optional. The textual description did mention it,
but all the way at the bottom. It fits better with the command
overview, so move it there, and fix the summary also.
Back-patch to 9.4 where \pset without arguments was supported.
Dilip Kumar, reviewed by Fabien Coelho
The regression tests for contrib/test_decoding regularly failed on
postgres instances that were very slow. Either because the hardware
itself was slow or because very expensive debugging options like
CLOBBER_CACHE_ALWAYS were used.
The reason they failed was just that some additional transactions were
decoded. Analyze and vacuum, triggered by autovac.
To fix just add a option to test_decoding to only display transactions
in which a change was actually displayed. That's not pretty because it
removes information from the tests; but better than constantly failing
tests in very likely harmless ways.
Backpatch to 9.4 where logical decoding was introduced.
Discussion: 20140629142511.GA26930@awork2.anarazel.de
68a2e52bba has introduced LWLockAcquireCommon() containing the
previous contents of LWLockAcquire() plus added functionality. The
latter then calls it, just like LWLockAcquireWithVar(). Because the
majority of callers don't need the added functionality, declare the
common code as inline. The compiler then can optimize away the unused
code. Doing so is also useful when looking at profiles, to
differentiate the users.
Backpatch to 9.4, the first branch to contain LWLockAcquireCommon().
