Unsupported versions: 9.3
This documentation is for an unsupported version of PostgreSQL.
You may want to view the same page for the current version, or one of the other supported versions listed above instead.

E.26. Release 9.3

Release date: 2013-09-09

E.26.1. Overview

Major enhancements in PostgreSQL 9.3 include:

The above items are explained in more detail in the sections below.

E.26.2. Migration to Version 9.3

A dump/restore using pg_dumpall, or use of pg_upgrade, is required for those wishing to migrate data from any previous release.

Version 9.3 contains a number of changes that may affect compatibility with previous releases. Observe the following incompatibilities:

E.26.2.1. Server Settings

  • Rename replication_timeout to wal_sender_timeout (Amit Kapila)

    This setting controls the WAL sender timeout.

  • Require superuser privileges to set commit_delay because it can now potentially delay other sessions (Simon Riggs)

  • Allow in-memory sorts to use their full memory allocation (Jeff Janes)

    Users who have set work_mem based on the previous behavior may need to revisit that setting.

E.26.2.2. Other

  • Throw an error if a tuple to be updated or deleted has already been updated or deleted by a BEFORE trigger (Kevin Grittner)

    Formerly, the originally-intended update was silently skipped, resulting in logical inconsistency since the trigger might have propagated data to other places based on the intended update. Now an error is thrown to prevent the inconsistent results from being committed. If this change affects your application, the best solution is usually to move the data-propagation actions to an AFTER trigger.

    This error will also be thrown if a query invokes a volatile function that modifies rows that are later modified by the query itself. Such cases likewise previously resulted in silently skipping updates.

  • Change multicolumn ON UPDATE SET NULL/SET DEFAULT foreign key actions to affect all columns of the constraint, not just those changed in the UPDATE (Tom Lane)

    Previously, we would set only those referencing columns that correspond to referenced columns that were changed by the UPDATE. This was what was required by SQL-92, but more recent editions of the SQL standard specify the new behavior.

  • Force cached plans to be replanned if the search_path changes (Tom Lane)

    Previously, cached plans already generated in the current session were not redone if the query was re-executed with a new search_path setting, resulting in surprising behavior.

  • Fix to_number() to properly handle a period used as a thousands separator (Tom Lane)

    Previously, a period was considered to be a decimal point even when the locale says it isn't and the D format code is used to specify use of the locale-specific decimal point. This resulted in wrong answers if FM format was also used.

  • Fix STRICT non-set-returning functions that have set-returning functions in their arguments to properly return null rows (Tom Lane)

    A null value passed to the strict function should result in a null output, but instead, that output row was suppressed entirely.

  • Store WAL in a continuous stream, rather than skipping the last 16MB segment every 4GB (Heikki Linnakangas)

    Previously, WAL files with names ending in FF were not used because of this skipping. If you have WAL backup or restore scripts that took this behavior into account, they will need to be adjusted.

  • In pg_constraint.confmatchtype, store the default foreign key match type (non-FULL, non-PARTIAL) as s for "simple" (Tom Lane)

    Previously this case was represented by u for "unspecified".

E.26.3. Changes

Below you will find a detailed account of the changes between PostgreSQL 9.3 and the previous major release.

E.26.3.1. Server

E.26.3.1.1. Locking

  • Prevent non-key-field row updates from blocking foreign key checks (Álvaro Herrera, Noah Misch, Andres Freund, Alexander Shulgin, Marti Raudsepp, Alexander Shulgin)

    This change improves concurrency and reduces the probability of deadlocks when updating tables involved in a foreign-key constraint. UPDATEs that do not change any columns referenced in a foreign key now take the new NO KEY UPDATE lock mode on the row, while foreign key checks use the new KEY SHARE lock mode, which does not conflict with NO KEY UPDATE. So there is no blocking unless a foreign-key column is changed.

  • Add configuration variable lock_timeout to allow limiting how long a session will wait to acquire any one lock (Zoltán Böszörményi)

E.26.3.1.2. Indexes

  • Add SP-GiST support for range data types (Alexander Korotkov)

  • Allow GiST indexes to be unlogged (Jeevan Chalke)

  • Improve performance of GiST index insertion by randomizing the choice of which page to descend to when there are multiple equally good alternatives (Heikki Linnakangas)

  • Improve concurrency of hash index operations (Robert Haas)

E.26.3.1.3. Optimizer

  • Collect and use histograms of upper and lower bounds, as well as range lengths, for range types (Alexander Korotkov)

  • Improve optimizer's cost estimation for index access (Tom Lane)

  • Improve optimizer's hash table size estimate for doing DISTINCT via hash aggregation (Tom Lane)

  • Suppress no-op Result and Limit plan nodes (Kyotaro Horiguchi, Amit Kapila, Tom Lane)

  • Reduce optimizer overhead by not keeping plans on the basis of cheap startup cost when the optimizer only cares about total cost overall (Tom Lane)

E.26.3.1.4. General Performance

  • Add COPY FREEZE option to avoid the overhead of marking tuples as frozen later (Simon Riggs, Jeff Davis)

  • Improve performance of NUMERIC calculations (Kyotaro Horiguchi)

  • Improve synchronization of sessions waiting for commit_delay (Peter Geoghegan)

    This greatly improves the usefulness of commit_delay.

  • Improve performance of the CREATE TEMPORARY TABLE ... ON COMMIT DELETE ROWS option by not truncating such temporary tables in transactions that haven't touched any temporary tables (Heikki Linnakangas)

  • Make vacuum recheck visibility after it has removed expired tuples (Pavan Deolasee)

    This increases the chance of a page being marked as all-visible.

  • Add per-resource-owner lock caches (Jeff Janes)

    This speeds up lock bookkeeping at statement completion in multi-statement transactions that hold many locks; it is particularly useful for pg_dump.

  • Avoid scanning the entire relation cache at commit of a transaction that creates a new relation (Jeff Janes)

    This speeds up sessions that create many tables in successive small transactions, such as a pg_restore run.

  • Improve performance of transactions that drop many relations (Tomas Vondra)

E.26.3.1.5. Monitoring

  • Add optional ability to checksum data pages and report corruption (Simon Riggs, Jeff Davis, Greg Smith, Ants Aasma)

    The checksum option can be set during initdb.

  • Split the statistics collector's data file into separate global and per-database files (Tomas Vondra)

    This reduces the I/O required for statistics tracking.

  • Fix the statistics collector to operate properly in cases where the system clock goes backwards (Tom Lane)

    Previously, statistics collection would stop until the time again reached the latest time previously recorded.

  • Emit an informative message to postmaster standard error when we are about to stop logging there (Tom Lane)

    This should help reduce user confusion about where to look for log output in common configurations that log to standard error only during postmaster startup.

E.26.3.1.6. Authentication

  • When an authentication failure occurs, log the relevant pg_hba.conf line, to ease debugging of unintended failures (Magnus Hagander)

  • Improve LDAP error reporting and documentation (Peter Eisentraut)

  • Add support for specifying LDAP authentication parameters in URL format, per RFC 4516 (Peter Eisentraut)

  • Change the ssl_ciphers parameter to start with DEFAULT, rather than ALL, then remove insecure ciphers (Magnus Hagander)

    This should yield a more appropriate SSL cipher set.

  • Parse and load pg_ident.conf once, not during each connection (Amit Kapila)

    This is similar to how pg_hba.conf is processed.

E.26.3.1.7. Server Settings

  • Greatly reduce System V shared memory requirements (Robert Haas)

    On Unix-like systems, mmap() is now used for most of PostgreSQL's shared memory. For most users, this will eliminate any need to adjust kernel parameters for shared memory.

  • Allow the postmaster to listen on multiple Unix-domain sockets (Honza Horák)

    The configuration parameter unix_socket_directory is replaced by unix_socket_directories, which accepts a list of directories.

  • Allow a directory of configuration files to be processed (Magnus Hagander, Greg Smith, Selena Deckelmann)

    Such a directory is specified with include_dir in the server configuration file.

  • Increase the maximum initdb-configured value for shared_buffers to 128MB (Robert Haas)

    This is the maximum value that initdb will attempt to set in postgresql.conf; the previous maximum was 32MB.

  • Remove the external PID file, if any, on postmaster exit (Peter Eisentraut)

E.26.3.2. Replication and Recovery

  • Allow a streaming replication standby to follow a timeline switch (Heikki Linnakangas)

    This allows streaming standby servers to receive WAL data from a slave newly promoted to master status. Previously, other standbys would require a resync to begin following the new master.

  • Add SQL functions pg_is_in_backup() and pg_backup_start_time() (Gilles Darold)

    These functions report the status of base backups.

  • Improve performance of streaming log shipping with synchronous_commit disabled (Andres Freund)

  • Allow much faster promotion of a streaming standby to primary (Simon Riggs, Kyotaro Horiguchi)

  • Add the last checkpoint's redo location to pg_controldata's output (Fujii Masao)

    This information is useful for determining which WAL files are needed for restore.

  • Allow tools like pg_receivexlog to run on computers with different architectures (Heikki Linnakangas)

    WAL files can still only be replayed on servers with the same architecture as the primary; but they can now be transmitted to and stored on machines of any architecture, since the streaming replication protocol is now machine-independent.

  • Make pg_basebackup --write-recovery-conf output a minimal recovery.conf file (Zoltán Böszörményi, Magnus Hagander)

    This simplifies setting up a standby server.

  • Allow pg_receivexlog and pg_basebackup --xlog-method to handle streaming timeline switches (Heikki Linnakangas)

  • Add wal_receiver_timeout parameter to control the WAL receiver's timeout (Amit Kapila)

    This allows more rapid detection of connection failure.

  • Change the WAL record format to allow splitting the record header across pages (Heikki Linnakangas)

    The new format is slightly more compact, and is more efficient to write.

E.26.3.3. Queries

  • Implement SQL-standard LATERAL option for FROM-clause subqueries and function calls (Tom Lane)

    This feature allows subqueries and functions in FROM to reference columns from other tables in the FROM clause. The LATERAL keyword is optional for functions.

  • Add support for piping COPY and psql \copy data to/from an external program (Etsuro Fujita)

  • Allow a multirow VALUES clause in a rule to reference OLD/NEW (Tom Lane)

E.26.3.4. Object Manipulation

  • Add support for event triggers (Dimitri Fontaine, Robert Haas, Álvaro Herrera)

    This allows server-side functions written in event-enabled languages to be called when DDL commands are run.

  • Allow foreign data wrappers to support writes (inserts/updates/deletes) on foreign tables (KaiGai Kohei)

  • Add CREATE SCHEMA ... IF NOT EXISTS clause (Fabrízio de Royes Mello)

  • Make REASSIGN OWNED also change ownership of shared objects (Álvaro Herrera)

  • Make CREATE AGGREGATE complain if the given initial value string is not valid input for the transition datatype (Tom Lane)

  • Suppress CREATE TABLE's messages about implicit index and sequence creation (Robert Haas)

    These messages now appear at DEBUG1 verbosity, so that they will not be shown by default.

  • Allow DROP TABLE IF EXISTS to succeed when a non-existent schema is specified in the table name (Bruce Momjian)

    Previously, it threw an error if the schema did not exist.

  • Provide clients with constraint violation details as separate fields (Pavel Stehule)

    This allows clients to retrieve table, column, data type, or constraint name error details. Previously such information had to be extracted from error strings. Client library support is required to access these fields.

E.26.3.4.1. ALTER

  • Support IF NOT EXISTS option in ALTER TYPE ... ADD VALUE (Andrew Dunstan)

    This is useful for conditionally adding values to enumerated types.

  • Add ALTER ROLE ALL SET to establish settings for all users (Peter Eisentraut)

    This allows settings to apply to all users in all databases. ALTER DATABASE SET already allowed addition of settings for all users in a single database. postgresql.conf has a similar effect.

  • Add support for ALTER RULE ... RENAME (Ali Dar)

E.26.3.4.2. VIEWs

  • Add materialized views (Kevin Grittner)

    Unlike ordinary views, where the base tables are read on every access, materialized views create physical tables at creation or refresh time. Access to the materialized view then reads from its physical table. There is not yet any facility for incrementally refreshing materialized views or auto-accessing them via base table access.

  • Make simple views auto-updatable (Dean Rasheed)

    Simple views that reference some or all columns from a single base table are now updatable by default. More complex views can be made updatable using INSTEAD OF triggers or INSTEAD rules.

  • Add CREATE RECURSIVE VIEW syntax (Peter Eisentraut)

    Internally this is translated into CREATE VIEW ... WITH RECURSIVE ....

  • Improve view/rule printing code to handle cases where referenced tables are renamed, or columns are renamed, added, or dropped (Tom Lane)

    Table and column renamings can produce cases where, if we merely substitute the new name into the original text of a rule or view, the result is ambiguous. This change fixes the rule-dumping code to insert manufactured table and column aliases when needed to preserve the original semantics.

E.26.3.5. Data Types

  • Increase the maximum size of large objects from 2GB to 4TB (Nozomi Anzai, Yugo Nagata)

    This change includes adding 64-bit-capable large object access functions, both in the server and in libpq.

  • Allow text timezone designations, e.g. "America/Chicago", in the "T" field of ISO-format timestamptz input (Bruce Momjian)

E.26.3.5.1. JSON

E.26.3.6. Functions

  • Add array_remove() and array_replace() functions (Marco Nenciarini, Gabriele Bartolini)

  • Allow concat() and format() to properly expand VARIADIC-labeled arguments (Pavel Stehule)

  • Improve format() to provide field width and left/right alignment options (Pavel Stehule)

  • Make to_char(), to_date(), and to_timestamp() handle negative (BC) century values properly (Bruce Momjian)

    Previously the behavior was either wrong or inconsistent with positive/AD handling, e.g. with the format mask "IYYY-IW-DY".

  • Make to_date() and to_timestamp() return proper results when mixing ISO and Gregorian week/day designations (Bruce Momjian)

  • Cause pg_get_viewdef() to start a new line by default after each SELECT target list entry and FROM entry (Marko Tiikkaja)

    This reduces line length in view printing, for instance in pg_dump output.

  • Fix map_sql_value_to_xml_value() to print values of domain types the same way their base type would be printed (Pavel Stehule)

    There are special formatting rules for certain built-in types such as boolean; these rules now also apply to domains over these types.

E.26.3.7. Server-Side Languages

E.26.3.7.1. PL/pgSQL Server-Side Language

  • Allow PL/pgSQL to use RETURN with a composite-type expression (Asif Rehman)

    Previously, in a function returning a composite type, RETURN could only reference a variable of that type.

  • Allow PL/pgSQL to access constraint violation details as separate fields (Pavel Stehule)

  • Allow PL/pgSQL to access the number of rows processed by COPY (Pavel Stehule)

    A COPY executed in a PL/pgSQL function now updates the value retrieved by GET DIAGNOSTICS x = ROW_COUNT.

  • Allow unreserved keywords to be used as identifiers everywhere in PL/pgSQL (Tom Lane)

    In certain places in the PL/pgSQL grammar, keywords had to be quoted to be used as identifiers, even if they were nominally unreserved.

E.26.3.7.2. PL/Python Server-Side Language

  • Add PL/Python result object string handler (Peter Eisentraut)

    This allows plpy.debug(rv) to output something reasonable.

  • Make PL/Python convert OID values to a proper Python numeric type (Peter Eisentraut)

  • Handle SPI errors raised explicitly (with PL/Python's RAISE) the same as internal SPI errors (Oskari Saarenmaa and Jan Urbanski)

E.26.3.8. Server Programming Interface (SPI)

  • Prevent leakage of SPI tuple tables during subtransaction abort (Tom Lane)

    At the end of any failed subtransaction, the core SPI code now releases any SPI tuple tables that were created during that subtransaction. This avoids the need for SPI-using code to keep track of such tuple tables and release them manually in error-recovery code. Failure to do so caused a number of transaction-lifespan memory leakage issues in PL/pgSQL and perhaps other SPI clients. SPI_freetuptable() now protects itself against multiple freeing requests, so any existing code that did take care to clean up shouldn't be broken by this change.

  • Allow SPI functions to access the number of rows processed by COPY (Pavel Stehule)

E.26.3.9. Client Applications

E.26.3.9.1. psql

  • Adjust function cost settings so psql tab completion and pattern searching are more efficient (Tom Lane)

  • Improve psql's tab completion coverage (Jeff Janes, Dean Rasheed, Peter Eisentraut, Magnus Hagander)

  • Allow the psql --single-transaction mode to work when reading from standard input (Fabien Coelho, Robert Haas)

    Previously this option only worked when reading from a file.

  • Remove psql warning when connecting to an older server (Peter Eisentraut)

    A warning is still issued when connecting to a server of a newer major version than psql's.

E.26.3.9.1.1. Backslash Commands
  • Add psql command \watch to repeatedly execute a SQL command (Will Leinweber)

  • Add psql command \gset to store query results in psql variables (Pavel Stehule)

  • Add SSL information to psql's \conninfo command (Alastair Turner)

  • Add "Security" column to psql's \df+ output (Jon Erdman)

  • Allow psql command \l to accept a database name pattern (Peter Eisentraut)

  • In psql, do not allow \connect to use defaults if there is no active connection (Bruce Momjian)

    This might be the case if the server had crashed.

  • Properly reset state after failure of a SQL command executed with psql's \g file (Tom Lane)

    Previously, the output from subsequent SQL commands would unexpectedly continue to go to the same file.

E.26.3.9.1.2. Output
  • Add a latex-longtable output format to psql (Bruce Momjian)

    This format allows tables to span multiple pages.

  • Add a border=3 output mode to the psql latex format (Bruce Momjian)

  • In psql's tuples-only and expanded output modes, no longer emit "(No rows)" for zero rows (Peter Eisentraut)

  • In psql's unaligned, expanded output mode, no longer print an empty line for zero rows (Peter Eisentraut)

E.26.3.9.2. pg_dump

  • Add pg_dump --jobs option to dump tables in parallel (Joachim Wieland)

  • Make pg_dump output functions in a more predictable order (Joel Jacobson)

  • Fix tar files emitted by pg_dump to be POSIX conformant (Brian Weaver, Tom Lane)

  • Add --dbname option to pg_dump, for consistency with other client commands (Heikki Linnakangas)

    The database name could already be supplied last without a flag.

E.26.3.9.3. initdb

  • Make initdb fsync the newly created data directory (Jeff Davis)

    This insures data integrity in event of a system crash shortly after initdb. This can be disabled by using --nosync.

  • Add initdb --sync-only option to sync the data directory to durable storage (Bruce Momjian)

    This is used by pg_upgrade.

  • Make initdb issue a warning about placing the data directory at the top of a file system mount point (Bruce Momjian)

E.26.3.10. Source Code

  • Add infrastructure to allow plug-in background worker processes (Álvaro Herrera)

  • Create a centralized timeout API (Zoltán Böszörményi)

  • Create libpgcommon and move pg_malloc() and other functions there (Álvaro Herrera, Andres Freund)

    This allows libpgport to be used solely for portability-related code.

  • Add support for list links embedded in larger structs (Andres Freund)

  • Use SA_RESTART for all signals, including SIGALRM (Tom Lane)

  • Ensure that the correct text domain is used when translating errcontext() messages (Heikki Linnakangas)

  • Standardize naming of client-side memory allocation functions (Tom Lane)

  • Provide support for "static assertions" that will fail at compile time if some compile-time-constant condition is not met (Andres Freund, Tom Lane)

  • Support Assert() in client-side code (Andrew Dunstan)

  • Add decoration to inform the C compiler that some ereport() and elog() calls do not return (Peter Eisentraut, Andres Freund, Tom Lane, Heikki Linnakangas)

  • Allow options to be passed to the regression test output comparison utility via PG_REGRESS_DIFF_OPTS (Peter Eisentraut)

  • Add isolation tests for CREATE INDEX CONCURRENTLY (Abhijit Menon-Sen)

  • Remove typedefs for int2/int4 as they are better represented as int16/int32 (Peter Eisentraut)

  • Fix install-strip on Mac OS X (Peter Eisentraut)

  • Remove configure flag --disable-shared, as it is no longer supported (Bruce Momjian)

  • Rewrite pgindent in Perl (Andrew Dunstan)

  • Provide Emacs macro to set Perl formatting to match PostgreSQL's perltidy settings (Peter Eisentraut)

  • Run tool to check the keyword list whenever the backend grammar is changed (Tom Lane)

  • Change the way UESCAPE is lexed, to significantly reduce the size of the lexer tables (Heikki Linnakangas)

  • Centralize flex and bison make rules (Peter Eisentraut)

    This is useful for pgxs authors.

  • Change many internal backend functions to return object OIDs rather than void (Dimitri Fontaine)

    This is useful for event triggers.

  • Invent pre-commit/pre-prepare/pre-subcommit events for transaction callbacks (Tom Lane)

    Loadable modules that use transaction callbacks might need modification to handle these new event types.

  • Add function pg_identify_object() to produce a machine-readable description of a database object (Álvaro Herrera)

  • Add post-ALTER-object server hooks (KaiGai Kohei)

  • Implement a generic binary heap and use it for Merge-Append operations (Abhijit Menon-Sen)

  • Provide a tool to help detect timezone abbreviation changes when updating the src/timezone/data files (Tom Lane)

  • Add pkg-config support for libpq and ecpg libraries (Peter Eisentraut)

  • Remove src/tools/backend, now that the content is on the PostgreSQL wiki (Bruce Momjian)

  • Split out WAL reading as an independent facility (Heikki Linnakangas, Andres Freund)

  • Use a 64-bit integer to represent WAL positions (XLogRecPtr) instead of two 32-bit integers (Heikki Linnakangas)

    Generally, tools that need to read the WAL format will need to be adjusted.

  • Allow PL/Python to support platform-specific include directories (Peter Eisentraut)

  • Allow PL/Python on OS X to build against custom versions of Python (Peter Eisentraut)

E.26.3.11. Additional Modules

  • Add a Postgres foreign data wrapper contrib module to allow access to other Postgres servers (Shigeru Hanada)

    This foreign data wrapper supports writes.

  • Add pg_xlogdump contrib program (Andres Freund)

  • Add support for indexing of regular-expression searches in pg_trgm (Alexander Korotkov)

  • Improve pg_trgm's handling of multibyte characters (Tom Lane)

    On a platform that does not have the wcstombs() or towlower() library functions, this could result in an incompatible change in the contents of pg_trgm indexes for non-ASCII data. In such cases, REINDEX those indexes to ensure correct search results.

  • Add a pgstattuple function to report the size of the pending-insertions list of a GIN index (Fujii Masao)

  • Make oid2name, pgbench, and vacuumlo set fallback_application_name (Amit Kapila)

  • Improve output of pg_test_timing (Bruce Momjian)

  • Improve output of pg_test_fsync (Peter Geoghegan)

  • Create a dedicated foreign data wrapper, with its own option validator function, for dblink (Shigeru Hanada)

    When using this FDW to define the target of a dblink connection, instead of using a hard-wired list of connection options, the underlying libpq library is consulted to see what connection options it supports.

E.26.3.11.1. pg_upgrade

  • Allow pg_upgrade to do dumps and restores in parallel (Bruce Momjian, Andrew Dunstan)

    This allows parallel schema dump/restore of databases, as well as parallel copy/link of data files per tablespace. Use the --jobs option to specify the level of parallelism.

  • Make pg_upgrade create Unix-domain sockets in the current directory (Bruce Momjian, Tom Lane)

    This reduces the possibility that someone will accidentally connect during the upgrade.

  • Make pg_upgrade --check mode properly detect the location of non-default socket directories (Bruce Momjian, Tom Lane)

  • Improve performance of pg_upgrade for databases with many tables (Bruce Momjian)

  • Improve pg_upgrade's logs by showing executed commands (Álvaro Herrera)

  • Improve pg_upgrade's status display during copy/link (Bruce Momjian)

E.26.3.11.2. pgbench

  • Add --foreign-keys option to pgbench (Jeff Janes)

    This adds foreign key constraints to the standard tables created by pgbench, for use in foreign key performance testing.

  • Allow pgbench to aggregate performance statistics and produce output every --aggregate-interval seconds (Tomas Vondra)

  • Add pgbench --sampling-rate option to control the percentage of transactions logged (Tomas Vondra)

  • Reduce and improve the status message output of pgbench's initialization mode (Robert Haas, Peter Eisentraut)

  • Add pgbench -q mode to print one output line every five seconds (Tomas Vondra)

  • Output pgbench elapsed and estimated remaining time during initialization (Tomas Vondra)

  • Allow pgbench to use much larger scale factors, by changing relevant columns from integer to bigint when the requested scale factor exceeds 20000 (Greg Smith)

E.26.3.12. Documentation

  • Allow EPUB-format documentation to be created (Peter Eisentraut)

  • Update FreeBSD kernel configuration documentation (Brad Davis)

  • Improve WINDOW function documentation (Bruce Momjian, Florian Pflug)

  • Add instructions for setting up the documentation tool chain on macOS (Peter Eisentraut)

  • Improve commit_delay documentation (Peter Geoghegan)