E.1. Release 9.6

E.1.3.1. Server

E.1.3.2. Replication and Recovery

• Allow synchronous replication to support multiple simultaneous synchronous standby servers, not just one (Masahiko Sawada, Beena Emerson, Michael Paquier, Fujii Masao, Kyotaro Horiguchi)

  The number of standby servers that must acknowledge a commit before it is considered complete is now configurable as part of the synchronous_standby_names parameter.

• Add new setting remote_apply for configuration parameter synchronous_commit (Thomas Munro)

  In this mode, the master waits for the transaction to be applied on the standby server, not just written to disk. That means that you can count on a transaction started on the standby to see all commits previously acknowledged by the master.

• Add a feature to the replication protocol, and a corresponding option to pg_create_physical_replication_slot(), to allow reserving WAL immediately when creating a replication slot (Gurjeet Singh, Michael Paquier)

  This allows the creation of a replication slot to guarantee that all the WAL needed for a base backup will be available.

• Add a --slot option to pg_basebackup (Peter Eisentraut)

  This lets pg_basebackup use a replication slot defined for WAL streaming. After the base backup completes, selecting the same slot for regular streaming replication allows seamless startup of the new standby server.

• Extend pg_start_backup() and pg_stop_backup() to support non-exclusive backups (Magnus Hagander)

E.1.3.3. Queries

• Allow functions that return sets of tuples to return simple NULLs (Andrew Gierth, Tom Lane)

  In the context of SELECT FROM function(...), a function that returned a set of composite values was previously not allowed to return a plain NULL value as part of the set. Now that is allowed and interpreted as a row of NULLs. This avoids corner-case errors with, for example, unnesting an array of composite values.

• Fully support array subscripts and field selections in the target column list of an INSERT with multiple VALUES rows (Tom Lane)

  Previously, such cases failed if the same target column was mentioned more than once, e.g., INSERT INTO tab (x[1], x[2]) VALUES (...).

• When appropriate, postpone evaluation of SELECT output expressions until after an ORDER BY sort (Konstantin Knizhnik)

  This change ensures that volatile or expensive functions in the output list are executed in the order suggested by ORDER BY, and that they are not evaluated more times than required when there is a LIMIT clause. Previously, these properties held if the ordering was performed by an index scan or pre-merge-join sort, but not if it was performed by a top-level sort.

• Widen counters recording the number of tuples processed to 64 bits (Andreas Scherbaum)

  This change allows command tags, e.g. SELECT, to correctly report tuple counts larger than 4 billion. This also applies to PL/pgSQL's GET DIAGNOSTICS ... ROW_COUNT command.

• Avoid doing encoding conversions by converting through the MULE_INTERNAL encoding (Tom Lane)

  Previously, many conversions for Cyrillic and Central European single-byte encodings were done by converting to a related MULE_INTERNAL coding scheme and then to the destination encoding. Aside from being inefficient, this meant that when the conversion encountered an untranslatable character, the error message would confusingly complain about failure to convert to or from MULE_INTERNAL, rather than the user-visible encoding.

• Consider performing joins of foreign tables remotely only when the tables will be accessed under the same role ID (Shigeru Hanada, Ashutosh Bapat, Etsuro Fujita)

  Previously, the foreign join pushdown infrastructure left the question of security entirely up to individual foreign data wrappers, but that made it too easy for an FDW to inadvertently create subtle security holes. So, make it the core code's job to determine which role ID will access each table, and do not attempt join pushdown unless the role is the same for all relevant relations.

E.1.3.4. Utility Commands

• Allow COPY to copy the output of an INSERT/UPDATE/DELETE ... RETURNING query (Marko Tiikkaja)

  Previously, an intermediate CTE had to be written to get this result.

• Introduce ALTER object DEPENDS ON EXTENSION (Abhijit Menon-Sen)

  This command allows a database object to be marked as depending on an extension, so that it will be dropped automatically if the extension is dropped (without needing CASCADE). However, the object is not part of the extension, and thus will be dumped separately by pg_dump.

• Make ALTER object SET SCHEMA do nothing when the object is already in the requested schema, rather than throwing an error as it historically has for most object types (Marti Raudsepp)

• Add options to ALTER OPERATOR to allow changing the selectivity functions associated with an existing operator (Yury Zhuravlev)

• Add an IF NOT EXISTS option to ALTER TABLE ADD COLUMN (Fabrízio de Royes Mello)

• Reduce the lock strength needed by ALTER TABLE when setting fillfactor and autovacuum-related relation options (Fabrízio de Royes Mello, Simon Riggs)

• Introduce CREATE ACCESS METHOD to allow extensions to create index access methods (Alexander Korotkov, Petr Jelínek)

• Add a CASCADE option to CREATE EXTENSION to automatically create any extensions the requested one depends on (Petr Jelínek)

• Make CREATE TABLE ... LIKE include an OID column if any source table has one (Bruce Momjian)

• If a CHECK constraint is declared NOT VALID in a table creation command, automatically mark it as valid (Amit Langote, Amul Sul)

  This is safe because the table has no existing rows. This matches the longstanding behavior of FOREIGN KEY constraints.

• Fix DROP OPERATOR to clear pg_operator.oprcom and pg_operator.oprnegate links to the dropped operator (Roma Sokolov)

  Formerly such links were left as-is, which could pose a problem in the somewhat unlikely event that the dropped operator's OID was reused for another operator.

• Do not show the same subplan twice in EXPLAIN output (Tom Lane)

  In certain cases, typically involving SubPlan nodes in index conditions, EXPLAIN would print data for the same subplan twice.

• Disallow creation of indexes on system columns, except for OID columns (David Rowley)

  Such indexes were never considered supported, and would very possibly misbehave since the system might change the system-column fields of a tuple without updating indexes. However, previously there were no error checks to prevent them from being created.

E.1.3.5. Permissions Management

• Use the privilege system to manage access to sensitive functions (Stephen Frost)

  Formerly, many security-sensitive functions contained hard-wired checks that would throw an error if they were called by a non-superuser. This forced the use of superuser roles for some relatively pedestrian tasks. The hard-wired error checks are now gone in favor of making initdb revoke the default public EXECUTE privilege on these functions. This allows installations to choose to grant usage of such functions to trusted roles that do not need all superuser privileges.

• Create some built-in roles that can be used to grant access to what were previously superuser-only functions (Stephen Frost)

  Currently the only such role is pg_signal_backend, but more are expected to be added in future.

E.1.3.6. Data Types

• Improve full-text search to support searching for phrases, that is, lexemes appearing adjacent to each other in a specific order, or with a specified distance between them (Teodor Sigaev, Oleg Bartunov, Dmitry Ivanov)

  A phrase-search query can be specified in tsquery input using the new operators <-> and <N>. The former means that the lexemes before and after it must appear adjacent to each other in that order. The latter means they must be exactly N lexemes apart.

• Allow omitting one or both boundaries in an array slice specifier, e.g. array_col[3:] (Yury Zhuravlev)

  Omitted boundaries are taken as the upper or lower limit of the corresponding array subscript. This allows simpler specification for many common use-cases.

• Be more careful about out-of-range dates and timestamps (Vitaly Burovoy)

  This change prevents unexpected out-of-range errors for timestamp with time zone values very close to the implementation limits. Previously, the "same" value might be accepted or not depending on the timezone setting, meaning that a dump and reload could fail on a value that had been accepted when presented. Now the limits are enforced according to the equivalent UTC time, not local time, so as to be independent of timezone.

  Also, PostgreSQL is now more careful to detect overflow in operations that compute new date or timestamp values, such as date + integer.

• For geometric data types, make sure infinity and NaN component values are treated consistently during input and output (Tom Lane)

  Such values will now always print the same as they would in a simple float8 column, and be accepted the same way on input. Previously the behavior was platform-dependent.

• Upgrade the ispell dictionary type to handle modern Hunspell files and support more languages (Artur Zakirov)

• Implement look-behind constraints in regular expressions (Tom Lane)

  A look-behind constraint is like a lookahead constraint in that it consumes no text; but it checks for existence (or nonexistence) of a match ending at the current point in the string, rather than one starting at the current point. Similar features exist in many other regular-expression engines.

• In regular expressions, if an apparent three-digit octal escape \nnn would exceed 377 (255 decimal), assume it is a two-digit octal escape instead (Tom Lane)

  This makes the behavior match current Tcl releases.

• Add transaction ID operators xid <> xid and xid <> int4, for consistency with the corresponding equality operators (Michael Paquier)

E.1.3.7. Functions

• Add jsonb_insert() function to insert a new element into a jsonb array, or a not-previously-existing key into a jsonb object (Dmitry Dolgov)

• Improve the accuracy of the ln(), log(), exp(), and pow() functions for type numeric (Dean Rasheed)

• Add a scale(numeric) function to extract the display scale of a numeric value (Marko Tiikkaja)

• Add trigonometric functions that work in degrees (Dean Rasheed)

  For example, sind() measures its argument in degrees, whereas sin() measures in radians. These functions go to some lengths to deliver exact results for values where an exact result can be expected, for instance sind(30) = 0.5.

• Ensure that trigonometric functions handle infinity and NaN inputs per the POSIX standard (Dean Rasheed)

  The POSIX standard says that these functions should return NaN for NaN input, and should throw an error for out-of-range inputs including infinity. Previously our behavior varied across platforms.

• Make to_timestamp(float8) convert float infinity to timestamp infinity (Vitaly Burovoy)

  Formerly it just failed on an infinite input.

• Add new functions for tsvector data (Stas Kelvich)

  The new functions are ts_delete(), ts_filter(), unnest(), tsvector_to_array(), array_to_tsvector(), and a variant of setweight() that sets the weight only for specified lexeme(s).

• Allow ts_stat() and tsvector_update_trigger() to operate on values that are of types binary-compatible with the expected argument type, not just exactly that type; for example allow citext where text is expected (Teodor Sigaev)

• Add variadic functions num_nulls() and num_nonnulls() that count the number of their arguments that are null or non-null (Marko Tiikkaja)

  An example usage is CHECK(num_nonnulls(a,b,c) = 1) which asserts that exactly one of a,b,c is not NULL. These functions can also be used to count the number of null or nonnull elements in an array.

• Add function parse_ident() to split a qualified, possibly quoted SQL identifier into its parts (Pavel Stehule)

• In to_number(), interpret a V format code as dividing by 10 to the power of the number of digits following V (Bruce Momjian)

  This makes it operate in an inverse fashion to to_char().

• Make the to_reg*() functions accept type text not cstring (Petr Korobeinikov)

  This avoids the need to write an explicit cast in most cases where the argument is not a simple literal constant.

• Add pg_size_bytes() function to convert human-readable size strings to numbers (Pavel Stehule, Vitaly Burovoy, Dean Rasheed)

  This function converts strings like those produced by pg_size_pretty() into bytes. An example usage is SELECT oid::regclass FROM pg_class WHERE pg_total_relation_size(oid) > pg_size_bytes('10 GB').

• In pg_size_pretty(), format negative numbers similarly to positive ones (Adrian Vondendriesch)

  Previously, negative numbers were never abbreviated, just printed in bytes.

• Add an optional missing_ok argument to the current_setting() function (David Christensen)

  This allows avoiding an error for an unrecognized parameter name, instead returning a NULL.

• Change various catalog-inspection functions to return NULL for invalid input (Michael Paquier)

  pg_get_viewdef() now returns NULL if given an invalid view OID, and several similar functions likewise return NULL for bad input. Previously, such cases usually led to "cache lookup failed" errors, which are not meant to occur in user-facing cases.

• Fix pg_replication_origin_xact_reset() to not have any arguments (Fujii Masao)

  The documentation said that it has no arguments, and the C code did not expect any arguments, but the entry in pg_proc mistakenly specified two arguments.

E.1.3.8. Server-Side Languages

• In PL/pgSQL, detect mismatched CONTINUE and EXIT statements while compiling a function, rather than at execution time (Jim Nasby)

• Extend PL/Python's error-reporting and message-reporting functions to allow specifying additional message fields besides the primary error message (Pavel Stehule)

• Allow PL/Python functions to call themselves recursively via SPI, and fix the behavior when multiple set-returning PL/Python functions are called within one query (Alexey Grishchenko, Tom Lane)

• Fix session-lifespan memory leaks in PL/Python (Heikki Linnakangas, Haribabu Kommi, Tom Lane)

• Modernize PL/Tcl to use Tcl's "object" APIs instead of simple strings (Jim Nasby, Karl Lehenbauer)

  This can improve performance substantially in some cases. Note that PL/Tcl now requires Tcl 8.4 or later.

• In PL/Tcl, make database-reported errors return additional information in Tcl's errorCode global variable (Jim Nasby, Tom Lane)

  This feature follows the Tcl convention for returning auxiliary data about an error.

• Fix PL/Tcl to perform encoding conversion between the database encoding and UTF-8, which is what Tcl expects (Tom Lane)

  Previously, strings were passed through without conversion, leading to misbehavior with non-ASCII characters when the database encoding was not UTF-8.

E.1.3.9. Client Interfaces

• Add a nonlocalized version of the severity field in error and notice messages (Tom Lane)

  This change allows client code to determine severity of an error or notice without having to worry about localized variants of the severity strings.

• Introduce a feature in libpq whereby the CONTEXT field of messages can be suppressed, either always or only for non-error messages (Pavel Stehule)

  The default behavior of PQerrorMessage() is now to print CONTEXT only for errors. The new function PQsetErrorContextVisibility() can be used to adjust this.

• Add support in libpq for regenerating an error message with a different verbosity level (Alex Shulgin)

  This is done with the new function PQresultVerboseErrorMessage(). This supports psql's new \errverbose feature, and may be useful for other clients as well.

• Improve libpq's PQhost() function to return useful data for default Unix-socket connections (Tom Lane)

  Previously it would return NULL if no explicit host specification had been given; now it returns the default socket directory path.

• Fix ecpg's lexer to handle line breaks within comments starting on preprocessor directive lines (Michael Meskes)

E.1.3.10. Client Applications

• Add a --strict-names option to pg_dump and pg_restore (Pavel Stehule)

  This option causes the program to complain if there is no match for a -t or -n option, rather than silently doing nothing.

• In pg_dump, dump locally-made changes of privilege assignments for system objects (Stephen Frost)

  While it has always been possible for a superuser to change the privilege assignments for built-in or extension-created objects, such changes were formerly lost in a dump and reload. Now, pg_dump recognizes and dumps such changes. (This works only when dumping from a 9.6 or later server, however.)

• Allow pg_dump to dump non-extension-owned objects that are within an extension-owned schema (Martín Marqués)

  Previously such objects were ignored because they were mistakenly assumed to belong to the extension owning their schema.

• In pg_dump output, include the table name in object tags for object types that are only uniquely named per-table (for example, triggers) (Peter Eisentraut)

E.1.3.11. Server Applications

• Improve error reporting during initdb's post-bootstrap phase (Tom Lane)

  Previously, an error here led to reporting the entire input file as the "failing query"; now just the current query is reported. To get the desired behavior, queries in initdb's input files must be separated by blank lines.

• Speed up initdb by using just one standalone-backend session for all the post-bootstrap steps (Tom Lane)

• Improve pg_rewind so that it can work when the target timeline changes (Alexander Korotkov)

  This allows, for example, rewinding a promoted standby back to some state of the old master's timeline.

E.1.3.12. Source Code

• Remove obsolete heap_formtuple/heap_modifytuple/heap_deformtuple functions (Peter Geoghegan)

• Add macros to make AllocSetContextCreate() calls simpler and safer (Tom Lane)

  Writing out the individual sizing parameters for a memory context is now deprecated in favor of using one of the new macros ALLOCSET_DEFAULT_SIZES, ALLOCSET_SMALL_SIZES, or ALLOCSET_START_SMALL_SIZES. Existing code continues to work, however.

• Unconditionally use static inline functions in header files (Andres Freund)

  This may result in warnings and/or wasted code space with very old compilers, but the notational improvement seems worth it.

• Improve TAP testing infrastructure (Michael Paquier, Craig Ringer, Álvaro Herrera, Stephen Frost)

  Notably, it is now possible to test recovery scenarios using this infrastructure.

• Make trace_lwlocks identify individual locks by name (Robert Haas)

• Improve psql's tab-completion code infrastructure (Thomas Munro, Michael Paquier)

  Tab-completion rules are now considerably easier to write, and more compact.

• Nail the pg_shseclabel system catalog into cache, so that it is available for access during connection authentication (Adam Brightwell)

  The core code does not use this catalog for authentication, but extensions might wish to consult it.

• Restructure index access method API to hide most of it at the C level (Alexander Korotkov, Andrew Gierth)

  This change modernizes the index AM API to look more like the designs we have adopted for foreign data wrappers and tablesample handlers. This simplifies the C code and makes it much more practical to define index access methods in installable extensions. A consequence is that most of the columns of the pg_am system catalog have disappeared. New inspection functions have been added to allow SQL queries to determine index AM properties that used to be discoverable from pg_am.

• Add pg_init_privs system catalog to hold original privileges of initdb-created and extension-created objects (Stephen Frost)

  This infrastructure allows pg_dump to dump changes that an installation may have made in privileges attached to system objects. Formerly, such changes would be lost in a dump and reload, but now they are preserved.

• Change the way that extensions allocate custom LWLocks (Amit Kapila, Robert Haas)

  The RequestAddinLWLocks() function is removed, and replaced by RequestNamedLWLockTranche(). This allows better identification of custom LWLocks, and is less error-prone.

• Improve the isolation tester to allow multiple sessions to wait concurrently, allowing testing of deadlock scenarios (Robert Haas)

• Introduce extensible node types (KaiGai Kohei)

  This change allows FDWs or custom scan providers to store data in a plan tree in a more convenient format than was previously possible.

• Make the planner deal with post-scan/join query steps by generating and comparing Paths, replacing a lot of ad-hoc logic (Tom Lane)

  This change provides only marginal user-visible improvements today, but it enables future work on a lot of upper-planner improvements that were impractical to tackle using the old code structure.

• Support partial aggregation (David Rowley, Simon Riggs)

  This change allows the computation of an aggregate function to be split into separate parts, for example so that parallel worker processes can cooperate on computing an aggregate. In future it might allow aggregation across local and remote data to occur partially on the remote end.

• Add a generic command progress reporting facility (Vinayak Pokale, Rahila Syed, Amit Langote, Robert Haas)

• Separate out psql's flex lexer to make it usable by other client programs (Tom Lane, Kyotaro Horiguchi)

  This eliminates code duplication for programs that need to be able to parse SQL commands well enough to identify command boundaries. Doing that in full generality is more painful than one could wish, and up to now only psql has really gotten it right among our supported client programs.

  A new source-code subdirectory src/fe_utils/ has been created to hold this and other code that is shared across our client programs. Formerly such sharing was accomplished by symbolic linking or copying source files at build time, which was ugly and required duplicate compilation.

• Introduce WaitEventSet API to allow efficient waiting for event sets that usually do not change from one wait to the next (Andres Freund, Amit Kapila)

• Add a generic interface for writing WAL records (Alexander Korotkov, Petr Jelínek, Markus Nullmeier)

  This change allows extensions to write WAL records for changes to pages using a standard layout. The problem of needing to replay WAL without access to the extension is solved by having generic replay code. This allows extensions to implement, for example, index access methods and have WAL support for them.

• Support generic WAL messages for logical decoding (Petr Jelínek, Andres Freund)

  This feature allows extensions to insert data into the WAL stream that can be read by logical-decoding plugins, but is not connected to physical data restoration.

• Allow SP-GiST operator classes to store an arbitrary "traversal value" while descending the index (Alexander Lebedev, Teodor Sigaev)

  This is somewhat like the "reconstructed value", but it could be any arbitrary chunk of data, not necessarily of the same data type as the indexed column.

• Introduce a LOG_SERVER_ONLY message level for ereport() (David Steele)

  This level acts like LOG except that the message is never sent to the client. It is meant for use in auditing and similar applications.

• Provide a Makefile target to build all generated headers (Michael Paquier, Tom Lane)

  submake-generated-headers can now be invoked to ensure that generated backend header files are up-to-date. This is useful in subdirectories that might be built "standalone".

• Support OpenSSL 1.1.0 (Andreas Karlsson, Heikki Linnakangas)

E.1.3.13. Additional Modules

• Add configuration parameter auto_explain.sample_rate to allow contrib/auto_explain to capture just a configurable fraction of all queries (Craig Ringer, Julien Rouhaud)

  This allows reduction of overhead for heavy query traffic, while still getting useful information on average.

• Add contrib/bloom module that implements an index access method based on Bloom filtering (Teodor Sigaev, Alexander Korotkov)

  This is primarily a proof-of-concept for non-core index access methods, but it could be useful in its own right for queries that search many columns.

• In contrib/cube, introduce distance operators for cubes, and support kNN-style searches in GiST indexes on cube columns (Stas Kelvich)

• Make contrib/hstore's hstore_to_jsonb_loose() and hstore_to_json_loose() functions agree on what is a number (Tom Lane)

  Previously, hstore_to_jsonb_loose() would convert numeric-looking strings to JSON numbers, rather than strings, even if they did not exactly match the JSON syntax specification for numbers. This was inconsistent with hstore_to_json_loose(), so tighten the test to match the JSON syntax.

• Add selectivity estimation functions for contrib/intarray operators to improve plans for queries using those operators (Yury Zhuravlev, Alexander Korotkov)

• Make contrib/pageinspect's heap_page_items() function show the raw data in each tuple, and add new functions tuple_data_split() and heap_page_item_attrs() for inspection of individual tuple fields (Nikolay Shaplov)

• Add an optional S2K iteration count parameter to contrib/pgcrypto's pgp_sym_encrypt() function (Jeff Janes)

• Add support for "word similarity" to contrib/pg_trgm (Alexander Korotkov, Artur Zakirov)

  These functions and operators measure the similarity between one string and the most similar single word of another string.

• Add configuration parameter pg_trgm.similarity_threshold for contrib/pg_trgm's similarity threshold (Artur Zakirov)

  This threshold has always been configurable, but formerly it was controlled by special-purpose functions set_limit() and show_limit(). Those are now deprecated.

• Improve contrib/pg_trgm's GIN operator class to speed up index searches in which both common and rare keys appear (Jeff Janes)

• Improve performance of similarity searches in contrib/pg_trgm GIN indexes (Christophe Fornaroli)

• Add contrib/pg_visibility module to allow examining table visibility maps (Robert Haas)

• Add ssl_extension_info() function to contrib/sslinfo, to print information about SSL extensions present in the X509 certificate used for the current connection (Dmitry Voronin)