PostgreSQL 12.3

What's new in this version:

- Fix possible failure with GENERATED columns. If a GENERATED column's value is an exact copy of another column of the table (and it is a pass-by-reference data type), it was possible to crash or insert corrupted data into the table. While it would be rather pointless for a GENERATED expression to just duplicate another column, an expression using a function that sometimes returns its input unchanged could create the situation.
- Handle inheritance of generated columns better (Peter Eisentraut). When a table column is inherited during CREATE TABLE ... INHERITS, disallow changing any generation properties when the parent column is already marked GENERATED; but allow a child column to be marked GENERATED when its parent is not.
- Fix cross-column references in CREATE TABLE LIKE INCLUDING GENERATED (Peter Eisentraut). CREATE TABLE ... LIKE failed when trying to copy a GENERATED expression that references a physically-later column.
- Propagate ALTER TABLE ... SET STORAGE to indexes (Peter Eisentraut). Non-expression index columns have always copied the attstorage property of their table column at creation. Update them when ALTER TABLE ... SET STORAGE is done, to maintain consistency.
- Preserve the indisclustered setting of indexes rewritten by ALTER TABLE. Previously, ALTER TABLE lost track of which index had been used for CLUSTER.
- Preserve the replica identity properties of indexes rewritten by ALTER TABLE
- Preserve the indisclustered setting of indexes rebuilt by REINDEX CONCURRENTLY (Justin Pryzby)
- Lock objects sooner during DROP OWNED BY (Álvaro Herrera).This avoids failures in race-condition cases where another session is deleting some of the same objects.
- Fix error-case processing for CREATE ROLE ... IN ROLE (Andrew Gierth). Some error cases would be reported as “unexpected node type” or the like, instead of the intended message.
- Ensure that when a partition is detached, any triggers cloned from its formerly-parent table are removed
- Fix crash when COLLATE is applied to a non-collatable type in a partition bound expression (Dmitry Dolgov). Ensure that unique indexes over partitioned tables match the equality semantics of the partitioning key. This would only be an issue with index opclasses that have unusual notions of equality, but it's wrong in theory, so check.
- Ensure that members of the pg_read_all_stats role can read all statistics views, as expected (Magnus Hagander). The functions underlying the pg_stat_progress_* views had not gotten this memo.
- Repair performance regression in information_schema.triggers view (Tom Lane). This patch redefines that view so that an outer WHERE clause constraining the table name can be pushed down into the view, allowing its calculations to be done only for triggers belonging to the table of interest rather than all triggers in the database. In a database with many triggers this would make a significant speed difference for queries of that form. Since things worked that way before v11, this is a potential performance regression. Users who find this to be a problem can fix it by replacing the view definition (or, perhaps, just deleting and reinstalling the whole information_schema schema).
- Repair performance regression in floating point overflow/underflow detection. Previous refactoring had resulted in isinf() being called extra times in some hot code paths.
- Fix full text search to handle NOT above a phrase search correctly. Queries such as !(foobar) failed to find matching rows when implemented as a GiST or GIN index search.
- Fix full text search for cases where a phrase search includes an item with both prefix matching and a weight restriction
- Fix ts_headline() to make better headline selections when working with phrase queries (Tom Lane)
- Fix bugs in gin_fuzzy_search_limit processing (Adé Heyward, Tom Lane). A small value of gin_fuzzy_search_limit could result in unexpected slowness due to unintentionally rescanning the same index page many times. Another code path failed to apply the intended filtering at all, possibly returning too many values.
- Allow input of type circle to accept the format “(x,y),r” as the documentation says it does (David Zhang)
- Make the get_bit() and set_bit() functions cope with bytea strings longer than 256MB (Movead Li). Since the bit number argument is only int4, it's impossible to use these functions to access bits beyond the first 256MB of a long bytea. We'll widen the argument to int8 in v13, but in the meantime, allow these functions to work on the initial substring of a long bytea.
- Ignore file-not-found errors in pg_ls_waldir() and allied functions (Tom Lane). This prevents a race condition failure if a file is removed between when we see its directory entry and when we attempt to stat() it.
- Avoid possibly leaking an open-file descriptor for a directory in pg_ls_dir(), pg_timezone_names(), pg_tablespace_databases(), and allied functions
- Fix polymorphic-function type resolution to correctly infer the actual type of an anyarray output when given only an anyrange input
- Fix server's connection-startup logic for case where a GSSAPI connection is rejected because support is not compiled in, and the client then tries SSL instead. This led to a bogus “unsupported frontend protocol” failure.
- Fix memory leakage during GSSAPI encryption (Tom Lane). Both the backend and libpq would leak memory equivalent to the total amount of data sent during the session, if GSSAPI encryption is in use.
- Fix query-lifespan memory leak for a set-returning function used in a query's FROM clause
- Avoid leakage of a hashed subplan's hash tables across multiple executions (Andreas Karlsson, Tom Lane). This mistake could result in severe memory bloat if a query re-executed a hashed subplan enough times.
- Improve planner's handling of no-op domain coercions. Fix some cases where a domain coercion that does nothing was not completely removed from expressions.
- Avoid unlikely crash when REINDEX is terminated by a session-shutdown signal
- Prevent printout of possibly-incorrect hash join table statistics in EXPLAIN
- Fix reporting of elapsed time for heap truncation steps in VACUUM VERBOSE
- Fix possible undercounting of deleted B-tree index pages in VACUUM VERBOSE output
- Fix wrong bookkeeping for oldest deleted page in a B-tree index. This could cause subtly wrong decisions about when VACUUM can skip an index cleanup scan; although it appears there may be no significant user-visible effects from that.
- Ensure that TimelineHistoryRead and TimelineHistoryWrite wait states are reported in all code paths that read or write timeline history files
- Avoid possibly showing “waiting” twice in a process's PS status
- Avoid race condition when ANALYZE replaces the catalog tuple for extended statistics data
- Remove ill-considered skip of “redundant” anti-wraparound vacuums (Michael Paquier). This avoids a corner case where autovacuum could get into a loop of repeatedly trying and then skipping the same vacuum job.
- Ensure INCLUDE'd columns are always removed from B-tree pivot tuples. This mistake wasted space in some rare cases, but was otherwise harmless.
- Cope with invalid TOAST indexes that could be left over after a failed REINDEX CONCURRENTLY
- Ensure that valid index dependencies are left behind after a failed REINDEX CONCURRENTLY (Michael Paquier). Previously the old index could be left with no pg_depend links at all, so that for example it would not get dropped if the parent table is dropped.
- Avoid failure if autovacuum tries to access a just-dropped temporary schema. This hazard only arises if a superuser manually drops a temporary schema; which isn't normal practice, but should work.
- Avoid premature recycling of WAL segments during crash recovery. WAL segments that become ready to be archived during crash recovery were potentially recycled without being archived.
- Avoid scanning irrelevant timelines during archive recovery (Kyotaro Horiguchi). This can eliminate many attempts to fetch non-existent WAL files from archive storage, which is helpful if archive access is slow.
- Remove bogus “subtransaction logged without previous top-level txn record” error check in logical decoding. This condition is legitimately reachable in various scenarios, so remove the check.
- Avoid possible failure after a replication slot copy, due to premature removal of WAL data
- Ensure that a replication slot's io_in_progress_lock is released in failure code paths. This could result in a walsender later becoming stuck waiting for the lock.
- Ensure that generated columns are correctly handled during updates issued by logical replication
- Fix race conditions in synchronous standby management. During a change in the synchronous_standby_names setting, there was a window in which wrong decisions could be made about whether it is OK to release transactions that are waiting for synchronous commit. Another hazard for similarly wrong decisions existed if a walsender process exited and was immediately replaced by another.
- Add missing SQLSTATE values to a few error reports
- Fix PL/pgSQL to reliably refuse to execute an event trigger function as a plain function
- Fix memory leak in libpq when using sslmode=verify-full. Certificate verification during connection startup could leak some memory. This would become an issue if a client process opened many database connections during its lifetime.
- Fix ecpg to treat an argument of just “-” as meaning “read from stdin” on all platforms
- Fix crash in psql when attempting to re-establish a failed connection
- Allow tab-completion of the filename argument to psql's gx command
- Add pg_dump support for ALTER ... DEPENDS ON EXTENSION. pg_dump previously ignored dependencies added this way, causing them to be forgotten during dump/restore or pg_upgrade.
- Fix pg_dump to dump comments on RLS policy objects
- In pg_dump, postpone restore of event triggers till the end. This minimizes the risk that an event trigger could interfere with the restoration of other objects.
- Ensure that pg_basebackup generates valid tar files. In some cases a partial block of zeroes would be added to the end of the file. While this seems to be harmless with common versions of tar, it's not OK per the POSIX file format spec.
- Make pg_checksums skip tablespace subdirectories that belong to a different PostgreSQL major version. Such subdirectories don't really belong to our database cluster, and so must not be processed.
- Ignore temporary copies of pg_internal.init in pg_checksums and related programs (Michael Paquier)
- Fix quoting of --encoding, --lc-ctype and --lc-collate values in createdb utility (Michael Paquier)
- contrib/lo's lo_manage() function crashed if called directly rather than as a trigger (Tom Lane)
- In contrib/ltree, protect against overflow of ltree and lquery length fields (Nikita Glukhov)
- Work around failure in contrib/pageinspect's bt_metap() function when an oldest_xact value exceeds 2^31-1 (Peter Geoghegan). Such XIDs will now be reported as negative integers, which isn't great but it beats throwing an error. v13 will widen the output argument to int8 to provide saner reporting.
- Fix cache reference leak in contrib/sepgsql
- On Windows, avoid premature creation of postmaster's log file during pg_ctl start. The previous coding could allow the file to be created with permissions that wouldn't allow the postmaster to write on it.
- Avoid failures when dealing with Unix-style locale names on Windows
- On Windows, set console VT100 compatibility mode in programs that support PG_COLOR colorization. Without this, the colorization option doesn't actually work.
- Stop requiring extra parentheses in ereport() calls
- Use pkg-config, if available, to locate libxml2 during configure. If pkg-config is not present or lacks knowledge of libxml2, we still query xml2-config as before. This change could break build processes that try to make PostgreSQL use a non-default version of libxml2 by putting that version's xml2-config into the PATH. Instead, set XML2_CONFIG to point to the non-default xml2-config. That method will work with either older or newer PostgreSQL releases.
- Fix Makefile dependencies for libpq and ecpg
- In MSVC builds, cope with spaces in the path name for Python
- In MSVC builds, fix detection of Visual Studio version to work with more language settings
- In MSVC builds, use -Wno-deprecated with bison versions newer than 3.0, as non-Windows builds already do. Update time zone data files to tzdata release 2020a for DST law changes in Morocco and the Canadian Yukon, plus historical corrections for Shanghai. The America/Godthab zone has been renamed to America/Nuuk to reflect current English usage; however, the old name remains available as a compatibility link. Also, update initdb's list of known Windows time zone names to include recent additions, improving the odds that it will correctly translate the system time zone setting on that platform.