資料庫新手該選 PostgreSQL 嗎？它能幫你扛住什麼

What's new in this version:

Changes:
* Require execute permission on the trigger function for CREATE TRIGGER 
- This missing check could allow another user to execute a trigger function with forged input data, by installing it on a table he owns. This is only of significance for trigger functions marked SECURITY DEFINER, since otherwise trigger functions run as the table owner anyway. 
* Remove arbitrary limitation on length of common name in SSL certificates 
- Both libpq and the server truncated the common name extracted from an SSL certificate at 32 bytes. Normally this would cause nothing worse than an unexpected verification failure, but there are some rather-implausible scenarios in which it might allow one certificate holder to impersonate another. The victim would have to have a common name exactly 32 bytes long, and the attacker would have to persuade a trusted CA to issue a certificate in which the common name has that string as a prefix. Impersonating a server would also require some additional exploit to redirect client connections. 
* Convert newlines to spaces in names written in pg_dump comments 
- pg_dump was incautious about sanitizing object names that are emitted within SQL comments in its output script. A name containing a newline would at least render the script syntactically incorrect. Maliciously crafted object names could present a SQL injection risk when the script is reloaded. 
* Fix btree index corruption from insertions concurrent with vacuuming 
- An index page split caused by an insertion could sometimes cause a concurrently-running VACUUM to miss removing index entries that it should remove. After the corresponding table rows are removed, the dangling index entries would cause errors (such as "could not read block N in file ...") or worse, silently wrong query results after unrelated rows are re-inserted at the now-free table locations. This bug has been present since release 8.2, but occurs so infrequently that it was not diagnosed until now. If you have reason to suspect that it has happened in your database, reindexing the affected index will fix things.
* Fix transient zeroing of shared buffers during WAL replay 
- The replay logic would sometimes zero and refill a shared buffer, so that the contents were transiently invalid. In hot standby mode this can result in a query that's executing in parallel seeing garbage data. Various symptoms could result from that, but the most common one seems to be "invalid memory alloc request size".
* Fix handling of data-modifying WITH subplans in READ COMMITTED rechecking 
- A WITH clause containing INSERT/UPDATE/DELETE would crash if the parent UPDATE or DELETE command needed to be re-evaluated at one or more rows due to concurrent updates in READ COMMITTED mode.
* Fix corner case in SSI transaction cleanup 
- When finishing up a read-write serializable transaction, a crash could occur if all remaining active serializable transactions are read-only.
* Fix postmaster to attempt restart after a hot-standby crash 
- A logic error caused the postmaster to terminate, rather than attempt to restart the cluster, if any backend process crashed while operating in hot standby mode.
* Fix CLUSTER/VACUUM FULL handling of toast values owned by recently-updated rows 
- This oversight could lead to "duplicate key value violates unique constraint" errors being reported against the toast table's index during one of these commands.
* Update per-column permissions, not only per-table permissions, when changing table owner 
- Failure to do this meant that any previously granted column permissions were still shown as having been granted by the old owner. This meant that neither the new owner nor a superuser could revoke the now-untraceable-to-table-owner permissions.
* Support foreign data wrappers and foreign servers in REASSIGN OWNED 
- This command failed with "unexpected classid" errors if it needed to change the ownership of any such objects.
* Allow non-existent values for some settings in ALTER USER/DATABASE SET 
- Allow default_text_search_config, default_tablespace, and temp_tablespaces to be set to names that are not known. This is because they might be known in another database where the setting is intended to be used, or for the tablespace cases because the tablespace might not be created yet. The same issue was previously recognized for search_path, and these settings now act like that one.
* Fix "unsupported node type" error caused by COLLATE in an INSERT expression 
* Avoid crashing when we have problems deleting table files post-commit 
- Dropping a table should lead to deleting the underlying disk files only after the transaction commits. In event of failure then (for instance, because of wrong file permissions) the code is supposed to just emit a warning message and go on, since it's too late to abort the transaction. This logic got broken as of release 8.4, causing such situations to result in a PANIC and an unrestartable database.
* Recover from errors occurring during WAL replay of DROP TABLESPACE 
- Replay will attempt to remove the tablespace's directories, but there are various reasons why this might fail (for example, incorrect ownership or permissions on those directories). Formerly the replay code would panic, rendering the database unrestartable without manual intervention. It seems better to log the problem and continue, since the only consequence of failure to remove the directories is some wasted disk space.
* Fix race condition in logging AccessExclusiveLocks for hot standby 
- Sometimes a lock would be logged as being held by "transaction zero". This is at least known to produce assertion failures on slave servers, and might be the cause of more serious problems.
* Track the OID counter correctly during WAL replay, even when it wraps around 
- Previously the OID counter would remain stuck at a high value until the system exited replay mode. The practical consequences of that are usually nil, but there are scenarios wherein a standby server that's been promoted to master might take a long time to advance the OID counter to a reasonable value once values are needed.
* Prevent emitting misleading "consistent recovery state reached" log message at the beginning of crash recovery 
* Fix initial value of pg_stat_replication.replay_location 
- Previously, the value shown would be wrong until at least one WAL record had been replayed.
* Fix regular expression back-references with * attached 
- Rather than enforcing an exact string match, the code would effectively accept any string that satisfies the pattern sub-expression referenced by the back-reference symbol.
- A similar problem still afflicts back-references that are embedded in a larger quantified expression, rather than being the immediate subject of the quantifier. This will be addressed in a future PostgreSQL release.
* Fix recently-introduced memory leak in processing of inet/cidr values 
- A patch in the December 2011 releases of PostgreSQL caused memory leakage in these operations, which could be significant in scenarios such as building a btree index on such a column.
* Fix planner's ability to push down index-expression restrictions through UNION ALL 
- This type of optimization was inadvertently disabled by a fix for another problem in 9.1.2.
* Fix planning of WITH clauses referenced in UPDATE/DELETE on an inherited table 
- This bug led to "could not find plan for CTE" failures.
* Fix GIN cost estimation to handle column IN (...) index conditions 
- This oversight would usually lead to crashes if such a condition could be used with a GIN index.
* Prevent assertion failure when exiting a session with an open, failed transaction 
- This bug has no impact on normal builds with asserts not enabled.
* Fix dangling pointer after CREATE TABLE AS/SELECT INTO in a SQL-language function 
- In most cases this only led to an assertion failure in assert-enabled builds, but worse consequences seem possible.
* Avoid double close of file handle in syslogger on Windows 
- Ordinarily this error was invisible, but it would cause an exception when running on a debug version of Windows.
* Fix I/O-conversion-related memory leaks in plpgsql 
- Certain operations would leak memory until the end of the current function.
* Work around bug in perl's SvPVutf8() function 
- This function crashes when handed a typeglob or certain read-only objects such as $^V. Make plperl avoid passing those to it.
* In pg_dump, don't dump contents of an extension's configuration tables if the extension itself is not being dumped 
* Improve pg_dump's handling of inherited table columns 
- pg_dump mishandled situations where a child column has a different default expression than its parent column. If the default is textually identical to the parent's default, but not actually the same (for instance, because of schema search path differences) it would not be recognized as different, so that after dump and restore the child would be allowed to inherit the parent's default. Child columns that are NOT NULL where their parent is not could also be restored subtly incorrectly.
* Fix pg_restore's direct-to-database mode for INSERT-style table data 
- Direct-to-database restores from archive files made with --inserts or --column-inserts options fail when using pg_restore from a release dated September or December 2011, as a result of an oversight in a fix for another problem. The archive file itself is not at fault, and text-mode output is okay.
* Teach pg_upgrade to handle renaming of plpython's shared library 
- Upgrading a pre-9.1 database that included plpython would fail because of this oversight.
* Allow pg_upgrade to process tables containing regclass columns 
- Since pg_upgrade now takes care to preserve pg_class OIDs, there was no longer any reason for this restriction.
* Make libpq ignore ENOTDIR errors when looking for an SSL client certificate file 
- This allows SSL connections to be established, though without a certificate, even when the user's home directory is set to something like /dev/null.
* Fix some more field alignment issues in ecpg's SQLDA area 
* Allow AT option in ecpg DEALLOCATE statements 
- The infrastructure to support this has been there for awhile, but through an oversight there was still an error check rejecting the case.
* Do not use the variable name when defining a varchar structure in ecpg 
* Fix contrib/auto_explain's JSON output mode to produce valid JSON 
- The output used brackets at the top level, when it should have used braces.
* Fix error in contrib/intarray's int[] & int[] operator 
- If the smallest integer the two input arrays have in common is 1, and there are smaller values in either array, then 1 would be incorrectly omitted from the result.
* Fix error detection in contrib/pgcrypto's encrypt_iv() and decrypt_iv() 
- These functions failed to report certain types of invalid-input errors, and would instead return random garbage values for incorrect input.
* Fix one-byte buffer overrun in contrib/test_parser 
- The code would try to read one more byte than it should, which would crash in corner cases. Since contrib/test_parser is only example code, this is not a security issue in itself, but bad example code is still bad.
* Use __sync_lock_test_and_set() for spinlocks on ARM, if available 
- This function replaces our previous use of the SWPB instruction, which is deprecated and not available on ARMv6 and later. Reports suggest that the old code doesn't fail in an obvious way on recent ARM boards, but simply doesn't interlock concurrent accesses, leading to bizarre failures in multiprocess operation.
* Use -fexcess-precision=standard option when building with gcc versions that accept it 
- This prevents assorted scenarios wherein recent versions of gcc will produce creative results.
* Allow use of threaded Python on FreeBSD 
- Our configure script previously believed that this combination wouldn't work; but FreeBSD fixed the problem, so remove that error check.
* Allow MinGW builds to use standardly-named OpenSSL libraries