E.2. Release 9.1.2

E.2.1. Migration to Version 9.1.2

A dump/restore is not required for those running 9.1.X.

However, a longstanding error was discovered in the definition of the information_schema.referential_constraints view. If you rely on correct results from that view, you should replace its definition as explained in the first changelog item below.

Also, if your installation was upgraded from a previous major release by running pg_upgrade, and it contains table columns of the citext data type, you should run CREATE EXTENSION citext FROM unpackaged. If you've already done that before upgrading to 9.1.2, you will instead need to do manual catalog updates as explained in the second changelog item.

E.2.2. Changes

• Fix bugs in information_schema.referential_constraints view (Tom Lane)

  This view was being insufficiently careful about matching the foreign-key constraint to the depended-on primary or unique key constraint. That could result in failure to show a foreign key constraint at all, or showing it multiple times, or claiming that it depends on a different constraint than the one it really does.

  Since the view definition is installed by initdb, merely upgrading will not fix the problem. If you need to fix this in an existing installation, you can (as a superuser) drop the information_schema schema then re-create it by sourcing SHAREDIR/information_schema.sql. (Run pg_config --sharedir if you're uncertain where SHAREDIR is.) This must be repeated in each database to be fixed.

• Make contrib/citext's upgrade script fix collations of citext columns and indexes (Tom Lane)

  Existing citext columns and indexes aren't correctly marked as being of a collatable data type during pg_upgrade from a pre-9.1 server. That leads to operations on them failing with errors such as "could not determine which collation to use for string comparison". This change allows them to be fixed by the same script that upgrades the citext module into a proper 9.1 extension during CREATE EXTENSION citext FROM unpackaged.

  If you have a previously-upgraded database that is suffering from this problem, and you already ran the CREATE EXTENSION command, you can manually run (as superuser) the UPDATE commands found at the end of SHAREDIR/extension/citext--unpackaged--1.0.sql. (Run pg_config --sharedir if you're uncertain where SHAREDIR is.)

• Fix possible crash during UPDATE or DELETE that joins to the output of a scalar-returning function (Tom Lane)

  A crash could only occur if the target row had been concurrently updated, so this problem surfaced only intermittently.

• Fix incorrect replay of WAL records for GIN index updates (Tom Lane)

  This could result in transiently failing to find index entries after a crash, or on a hot-standby server. The problem would be repaired by the next VACUUM of the index, however.

• Fix TOAST-related data corruption during CREATE TABLE dest AS SELECT * FROM src or INSERT INTO dest SELECT * FROM src (Tom Lane)

  If a table has been modified by ALTER TABLE ADD COLUMN, attempts to copy its data verbatim to another table could produce corrupt results in certain corner cases. The problem can only manifest in this precise form in 8.4 and later, but we patched earlier versions as well in case there are other code paths that could trigger the same bug.

• Fix possible failures during hot standby startup (Simon Riggs)

• Start hot standby faster when initial snapshot is incomplete (Simon Riggs)

• Fix race condition during toast table access from stale syscache entries (Tom Lane)

  The typical symptom was transient errors like "missing chunk number 0 for toast value NNNNN in pg_toast_2619", where the cited toast table would always belong to a system catalog.

• Track dependencies of functions on items used in parameter default expressions (Tom Lane)

  Previously, a referenced object could be dropped without having dropped or modified the function, leading to misbehavior when the function was used. Note that merely installing this update will not fix the missing dependency entries; to do that, you'd need to CREATE OR REPLACE each such function afterwards. If you have functions whose defaults depend on non-built-in objects, doing so is recommended.

• Fix incorrect management of placeholder variables in nestloop joins (Tom Lane)

  This bug is known to lead to "variable not found in subplan target list" planner errors, and could possibly result in wrong query output when outer joins are involved.

• Fix window functions that sort by expressions involving aggregates (Tom Lane)

  Previously these could fail with "could not find pathkey item to sort" planner errors.

• Fix "MergeAppend child's targetlist doesn't match MergeAppend" planner errors (Tom Lane)

• Fix index matching for operators with both collatable and noncollatable inputs (Tom Lane)

  In 9.1.0, an indexable operator that has a non-collatable left-hand input type and a collatable right-hand input type would not be recognized as matching the left-hand column's index. An example is the hstore ? text operator.

• Allow inlining of set-returning SQL functions with multiple OUT parameters (Tom Lane)

• Don't trust deferred-unique indexes for join removal (Tom Lane and Marti Raudsepp)

  A deferred uniqueness constraint might not hold intra-transaction, so assuming that it does could give incorrect query results.

• Make DatumGetInetP() unpack inet datums that have a 1-byte header, and add a new macro, DatumGetInetPP(), that does not (Heikki Linnakangas)

  This change affects no core code, but might prevent crashes in add-on code that expects DatumGetInetP() to produce an unpacked datum as per usual convention.

• Improve locale support in money type's input and output (Tom Lane)

  Aside from not supporting all standard lc_monetary formatting options, the input and output functions were inconsistent, meaning there were locales in which dumped money values could not be re-read.

• Don't let transform_null_equals affect CASE foo WHEN NULL ... constructs (Heikki Linnakangas)

  transform_null_equals is only supposed to affect foo = NULL expressions written directly by the user, not equality checks generated internally by this form of CASE.

• Change foreign-key trigger creation order to better support self-referential foreign keys (Tom Lane)

  For a cascading foreign key that references its own table, a row update will fire both the ON UPDATE trigger and the CHECK trigger as one event. The ON UPDATE trigger must execute first, else the CHECK will check a non-final state of the row and possibly throw an inappropriate error. However, the firing order of these triggers is determined by their names, which generally sort in creation order since the triggers have auto-generated names following the convention "RI_ConstraintTrigger_NNNN". A proper fix would require modifying that convention, which we will do in 9.2, but it seems risky to change it in existing releases. So this patch just changes the creation order of the triggers. Users encountering this type of error should drop and re-create the foreign key constraint to get its triggers into the right order.

• Fix IF EXISTS to work correctly in DROP OPERATOR FAMILY (Robert Haas)

• Disallow dropping of an extension from within its own script (Tom Lane)

  This prevents odd behavior in case of incorrect management of extension dependencies.

• Don't mark auto-generated types as extension members (Robert Haas)

  Relation rowtypes and automatically-generated array types do not need to have their own extension membership entries in pg_depend, and creating such entries complicates matters for extension upgrades.

• Cope with invalid pre-existing search_path settings during CREATE EXTENSION (Tom Lane)

• Avoid floating-point underflow while tracking buffer allocation rate (Greg Matthews)

  While harmless in itself, on certain platforms this would result in annoying kernel log messages.

• Prevent autovacuum transactions from running in serializable mode (Tom Lane)

  Autovacuum formerly used the cluster-wide default transaction isolation level, but there is no need for it to use anything higher than READ COMMITTED, and using SERIALIZABLE could result in unnecessary delays for other processes.

• Ensure walsender processes respond promptly to SIGTERM (Magnus Hagander)

• Exclude postmaster.opts from base backups (Magnus Hagander)

• Preserve configuration file name and line number values when starting child processes under Windows (Tom Lane)

  Formerly, these would not be displayed correctly in the pg_settings view.

• Fix incorrect field alignment in ecpg's SQLDA area (Zoltan Boszormenyi)

• Preserve blank lines within commands in psql's command history (Robert Haas)

  The former behavior could cause problems if an empty line was removed from within a string literal, for example.

• Avoid platform-specific infinite loop in pg_dump (Steve Singer)

• Fix compression of plain-text output format in pg_dump (Adrian Klaver and Tom Lane)

  pg_dump has historically understood -Z with no -F switch to mean that it should emit a gzip-compressed version of its plain text output. Restore that behavior.

• Fix pg_dump to dump user-defined casts between auto-generated types, such as table rowtypes (Tom Lane)

• Fix missed quoting of foreign server names in pg_dump (Tom Lane)

• Assorted fixes for pg_upgrade (Bruce Momjian)

  Handle exclusion constraints correctly, avoid failures on Windows, don't complain about mismatched toast table names in 8.4 databases.

• In PL/pgSQL, allow foreign tables to define row types (Alexander Soudakov)

• Fix up conversions of PL/Perl functions' results (Alex Hunsaker and Tom Lane)

  Restore the pre-9.1 behavior that PL/Perl functions returning void ignore the result value of their last Perl statement; 9.1.0 would throw an error if that statement returned a reference. Also, make sure it works to return a string value for a composite type, so long as the string meets the type's input format. In addition, throw errors for attempts to return Perl arrays or hashes when the function's declared result type is not an array or composite type, respectively. (Pre-9.1 versions rather uselessly returned strings like ARRAY(0x221a9a0) or HASH(0x221aa90) in such cases.)

• Ensure PL/Perl strings are always correctly UTF8-encoded (Amit Khandekar and Alex Hunsaker)

• Use the preferred version of xsubpp to build PL/Perl, not necessarily the operating system's main copy (David Wheeler and Alex Hunsaker)

• Correctly propagate SQLSTATE in PL/Python exceptions (Mika Eloranta and Jan Urbanski)

• Do not install PL/Python extension files for Python major versions other than the one built against (Peter Eisentraut)

• Change all the contrib extension script files to report a useful error message if they are fed to psql (Andrew Dunstan and Tom Lane)

  This should help teach people about the new method of using CREATE EXTENSION to load these files. In most cases, sourcing the scripts directly would fail anyway, but with harder-to-interpret messages.

• Fix incorrect coding in contrib/dict_int and contrib/dict_xsyn (Tom Lane)

  Some functions incorrectly assumed that memory returned by palloc() is guaranteed zeroed.

• Remove contrib/sepgsql tests from the regular regression test mechanism (Tom Lane)

  Since these tests require root privileges for setup, they're impractical to run automatically. Switch over to a manual approach instead, and provide a testing script to help with that.

• Fix assorted errors in contrib/unaccent's configuration file parsing (Tom Lane)

• Honor query cancel interrupts promptly in pgstatindex() (Robert Haas)

• Fix incorrect quoting of log file name in Mac OS X start script (Sidar Lopez)

• Revert unintentional enabling of WAL_DEBUG (Robert Haas)

  Fortunately, as debugging tools go, this one is pretty cheap; but it's not intended to be enabled by default, so revert.

• Ensure VPATH builds properly install all server header files (Peter Eisentraut)

• Shorten file names reported in verbose error messages (Peter Eisentraut)

  Regular builds have always reported just the name of the C file containing the error message call, but VPATH builds formerly reported an absolute path name.

• Fix interpretation of Windows timezone names for Central America (Tom Lane)

  Map "Central America Standard Time" to CST6, not CST6CDT, because DST is generally not observed anywhere in Central America.

• Update time zone data files to tzdata release 2011n for DST law changes in Brazil, Cuba, Fiji, Palestine, Russia, and Samoa; also historical corrections for Alaska and British East Africa.