Added the bgfreezer_log parameter for configuring logging of the background freezer mechanism.

Now background freezer is recommended when dealing with table bloat where many outdated row versions appear in tables in case of multiple UPDATE operations. Specifically, it helps with bloating that may occur if a row is inserted with NULL values using INSERT and is then updated multiple times with UPDATE, each time modifying different columns of that row.

Fixed a memory leak in the background freezer mechanism that occurred during its restart after an error.

Fixed an issue with the background freezer failing to log the backend type.

Changed the name format of crash_info output files to specify date and time with microseconds.

Fixed a bug that could occur when trying to write a query text to crash_info output files.

Improved CFS performance that now uses compression contexts with the default compression library of zstd.

Changed the behavior of the cfs_compression_ratio function. Now the function returns NaN for uncompressed relations.

Fixed an issue where pg_rewind could fail when used with unlogged tables containing TOAST data in a compressed tablespace. The issue was caused by missing cleanup of the *.cfm files when copying the “init fork” for the unlogged table.

Fixed an issue with stopping CFS garbage collection when it would continue running after being disabled by setting cfs_gc to false.

Upgraded biha to version 1.7.

Implemented cascading replication for BiHA clusters that helps to decrease workloads on the leader and allows deploying BiHA clusters in different data centers. You can configure cascading replication when initializing a new BiHA cluster or GDBiHA cluster by means of the updated bihactl utility and then modify cascading replication configuration using biha.max_replicas, biha.preferred_roles, biha.priority parameters and corresponding functions. If any replication source fails, BiHA re-configures cascading replication automatically.

Implemented experimental functionality that allows to unite nodes of the BiHA cluster into segments (logical nodes) that can be located in geographically distributed data centers. Data is transferred between segments via a single thread between the leader and the front follower — physical nodes located in the leader segment and the follower segment respectively. If the leader or front follower fails, data replication is re-configured automatically. If required, you can set the leader segment manually. For more information, refer to Multi-Level Geographical Distribution and Disaster Resilience.

Upgraded the bihactl utility that keeps former functionality and provides support for managing cascading replication and segments for multi-level geographical distribution and disaster resilience.

Added the biha.super_status_v view — the extended version of the biha.status_v view that additionally displays data for GDBiHA cluster.

Enhanced the logging functionality to provide better usability and log message clarity. Added separate logging levels for Postgres Controller and Config Controller components.

Implemented the service mode functionality managed by the biha.service_mode function. The service mode allows to temporarily pause BiHA operation. This may be required when you modify configuration parameters and want to ensure changes are applied on all cluster nodes.

Added the biha.monitoring function that allows monitoring the current status of BiHA cluster nodes. You can call this function from any node.

Added a special protection mechanism that helps to prevent node hanging and split-brain issues. You can manage the hanging prevention mechanism using the biha.watchdog_timeout configuration parameter.

Implemented compatibility with the replication slot synchronization feature.

Fixed an issue of inconsistent cluster configuration parameters of a node in the NODE_ERROR state and the leader. Nodes in NODE_ERROR state are now allowed to update cluster configuration parameter values.

Fixed a segmentation fault that appeared when using biha.add_to_ssn and biha.remove_from_ssn.

Fixed an issue where the referee size grew excessively during initialization. The reason was that the referee in the in the referee mode did not filter WAL application. Currently, the referee in the referee mode filters WAL application just like the referee in the referee_with_wal mode.

Fixed an issue that caused the “invalid xl_info in checkpoint record” log message followed by the leader failure. The issue occurred due to lack of the standby.signal file in case of BiHA worker unexpected restart. For example, this could be caused by a segmentation fault of another background worker that forces Postgres Pro to restart all processes. Missing standby.signal led to incorrect timeline selection on Postgres Pro startup.

Fixed a bug where a node in the CSTATE_FORMING state could vote for a candidate with a lower LSN when electing the new leader.

Fixed a bug where the pg_hba.biha.conf include directive was written at the end of pb_hba.conf file.

Removed the following redundant options of the bihactl utility: --user-postgres-cert, --user-postgres-key, and --postgres-ssl-mode.

Ended support for OpenSSL version lower than 1.1.1. If you want to use SSL for service connections, ensure that your installed OpenSSL version is 1.1.1 or higher.

Ended support for Elbrus CPU architecture.

Optimized the mechanism of aqo learning. Now aqo ignores plan nodes that do not affect cardinality estimation, for instance Materialized and Sort nodes. These ignored nodes no longer write aqo details in the EXPLAIN output and are not added to aqo statistics.

Changed the implementation of the least recently used (LRU) cache mechanism for the auto mode. Now aqo uses two distinct caches for aqo_data and aqo_query_texts. When the aqo.dsm_size_max limit is reached, the LRU algorithm first removes data from aqo_query_texts, if any, and only then from aqo_data.

Fixed an issue where aqo did not learn on Group plan nodes.

Fixed an issue where aqo did not distinguish subqueries from each other, which resulted in incorrect calculation of learning statistics for parent plan nodes.

Fixed a memory leak in aqo during data recovery on physical replicas.

Enhanced the mechanism of synchronization point creation when adding a new node to the cluster that fixes the potential issue of data divergence on nodes.

Enhanced filtering of replicated transactions that fixes the potential issue of data divergence on nodes.

Implemented filtering of the ALTER DATABASE command type to execute these commands locally on the current node. Previously, executing these commands could cause an error due to an attempt to replicate them to all cluster nodes. For more information about the commands that can only be executed locally, refer to Section F.36.6.1.

Fixed backup failures due to erroneous inclusion of unlogged tables and temporary tablespaces. Previously, if such files were deleted during the backup process, pg_probackup3 would treat them as missing and abort the operation.

Fixed incorrect Postgres Pro edition detection in the PRO data source mode for Shardman.

Added the pg_proaudit.log_rows configuration parameter that enables or disables logging of the number of rows affected by an SQL command.

Added a new event log field: connection duration. DISCONNECT events now contain the total connection duration alongside other connection parameters.

Added verification and validation of recovery-target-name in the WAL archive during recovery. This prevents situations where the Postgres Pro cluster fails to start due to an unreachable recovery target. If a named recovery target was created before the last backup, the recovery process now correctly checks its existence.

Added the ability to pass parameters to a remote agent in a configuration file format to improve flexibility when working with backup agents in distributed environments.

Implemented compatibility with the pgpro_cpumeter extension. The pgpro_cpumeter.stat file is now added to the exclusion list during backup.

Added support for virtual-hosted-style requests for S3 storages, including automatic detection of virtual-hosted-style based on the hostname format (when the hostname starts with the bucket name) and improved compatibility with various S3 API implementations (AWS S3, SberCloud, and others). In virtual-hosted-style URLs, the bucket name is included in the hostname, eliminating the need to add it to the file path.

Fixed an issue where the --skip-block-validation flag was being ignored during backup. Previously, this resulted in an error if the cluster contained pages with invalid checksums in their headers. This is critically important when enabling checksums in older clusters that may have pre-existing checksum issues. Also, a warning is now displayed when --skip-block-validation is used.

Improved backup status reporting. The backup status is now set to DONE before the validation phase. Backup metadata is written twice: before validation and after backup completion. The show command now displays a more accurate backup status during long-running validation checks.

Fixed an issue that could result in a crash if the backup_content.control file was deleted before the DELTA backup.

Added the estimate command to evaluate compression ratios across different compression algorithms, analogous to the cfs_estimate(relation) function.

Added the --per-file option for the stat command to collect detailed statistics for each file.

Added support for Russian localization.

Fixed compression level ranges for compression algorithms.

Fixed the unpack command behavior. Previously, compressed files were overwritten when unpacked to the same directory and incorrectly deleted when unpacked to a different location.

Added support of references from tables to the global index on a partitioned table using the foreign key constraint.

Added support for row-level locks in queries with global indexes.

Added the new pgpro_multiplan.show_hint_string parameter that allows displaying sets of hints, which are formed based on plans, in the EXPLAIN output.

Added the optional dbid argument to functions that remove frozen queries and update hint sets. For other remove functions, this argument also becomes optional. It allows you to specify the ID of a database for which these functions should be used. If omitted, functions are used for the current database.

Added the new pgpro_multiplan_baseline_cost_update function to update the cost of the specified allowed plan.

Fixed an issue in pgpro_multiplan that could occur when working together with pgpro_stats for collecting statistics about usage of frozen plans.

Added new procedures for queue access management: reset_queue_access, grant_queue_access, and revoke_queue_access. Queue management procedures can now be used only by a superuser who created the pgpro_queue extension, while the grant_queue_access and revoke_queue_access procedures are also available to the queue owner.

Added support for callback functions, enabling asynchronous and parallel message processing to optimize the message reading mechanism. To manage this functionality, use new configuration parameters and procedures.

Added new functions for exchange management and procedures for message publishing to support efficient message distribution based on the publish-subscribe model.

Restored the possibility to track aggregate statistics for sessions, which had previously been removed as unused. Whether aggregate statistics for the sessions are tracked is controlled by the pgpro_stats.track_sessions configuration parameter.

Improved the behavior of session-tracing functions. The pgpro_stats_trace_insert and pgpro_stats_trace_update functions now fail with an error if some filter attribute is listed in the passed sequence of key-value pairs more than once. If a wrong argument is passed to the pgpro_stats_trace_update function, the filter no longer gets updated. The pgpro_stats_trace_show function now returns exactly the values of explain* attributes that were specified for the filter.

Fixed an issue to avoid crash when two or more constants are marked as having the same location in the SQL statement text.

Added the autovacuum_vacuum_scale_factor and autovacuum_vacuum_insert_scale_factor parameters to the pgpro_tune default.tune preset configuration.

Fixed an issue with pgpro_tune not being able to process preset names with whitespaces when specifying a preset in the initdb PGPRO_TUNE environment variable. Now if pgpro_tune fails, initdb also fails.

Fixed the default.tune preset, which now correctly displays the order of effective_io_concurrency and random_page_cost parameter values depending on the drive type.

Implemented new counters for client connection metrics: client.lfe.failed, client.lfe.auth_ldap, and client.lfe.auth_reject. For the full list of available counters to monitor client connections, refer to the Client Connection Counters table.

Implemented a range of internal fixes and optimizations to improve performance and reliability of the proxima extension.

Extended the list of objects and SQL functions that enable proxima to automatically establish the dynamic dedicated session when they are present in a query. A session now turns to dedicated when executing CREATE STATISTICS in a temporary schema. For more information, refer to Dynamic Dedicated Session.

Implemented the adaptive-cpu balancing algorithm.

Implemented system.cpu_load — a system metrics counter that allows to monitor CPU utilization on nodes. In a BiHA cluster, you can query the new proxima.proxima_metrics_system view to check the current counter value. For more information, refer to System Metrics.

Added functions that allow inspecting all types of RUM index pages at a low level.

Fixed a bug that caused scanning of multi-column RUM indexes to return incorrect results. This happened when the index was created with order_by_attach set to true for one of the columns, and a query specified conditions for several index columns, including the one with order_by_attach.