Implemented a set of optimizations to improve query planning and execution time. In particular, time spent executing each plan node is no longer measured when the query execution time trigger is disabled, which has a significant impact on query performance.

Fixed an issue with logging the duration of completed statements with the extended query protocol when the log_duration parameter was enabled.

Added the crash_info_timer, crash_info_timer_interval, and crash_info_query_threshold configuration parameters that allow profiling long queries in crash_info output files.

Extended the diagnostic information that the crash_info_dump configuration parameter can specify to write in case of a backend crash. By default, now the general information on the signal is dumped and for Linux x64, also the information on the processor state and bytes near RIP (return instruction pointer). All this ensures better diagnostics, in particular, in case of SIGILL (illegal instruction) signals. The way of dumping stack files in case of crashes has also been changed for Linux x64. Besides, memory_context has been removed from the default list of sources to be dumped in order to avoid failures that could occur when writing the diagnostic information.

Improved compressed tablespace operations on s390/s390x architectures.

Improved the compression ratio estimation for tables using the cfs_estimate(relation) function. Note that cfs_estimate(relation) should be used only after a checkpoint to avoid incorrect estimations.

Disabled truncation for compressed relations during vacuum operations.

Removed encryption support for CFS, including the cfs_encryption configuration parameter.

Added the tmpoids field to the aqo_data view and the tmpoids parameter to the aqo_data_update function to support queries that involve temporary tables along with persistent tables.

Fixed the race condition when calling the aqo_cleanup function from different sessions.

Fixed an issue where aqo learning data was not saved for queries that use materialization.

Fixed an issue with learning on early terminated nodes. Now for these nodes, if the number of actual rows exceeds the number of predicted rows, aqo learns but with the minimum reliability value. Otherwise, learning on these nodes is disabled.

Implemented the ability to use SSL certificates for user authentication. For more information about SSL authentication configuration and other available authentication methods, see User Authentication Configuration.

Implemented the STATUS_CHANGED callback that is called on every node whenever values in the fields of the biha.status_v view change (except since_last_hb).

Implemented a set of functions to manage quorum-based synchronous replication in a BiHA cluster: biha.add_to_ssn, biha.get_ssn, biha.remove_from_ssn, and biha.set_ssn. For more information, see Managing Replication.

Implemented the ability to force the pg_rewind process in a BiHA cluster when required, for example, for partial database recovery, emergency repairs when standard rewind fails, or for recovery of specific WAL ranges. You can execute force run of pg_rewind by specifying both --force and --start options. The --force option overrides failed safety check to allow pg_rewind execution. The --start=LSN[,TLI] option specifies the exact WAL location (LSN) and an optional timeline ID (TLI) that defaults to the current timeline of the target node if omitted. If these options are specified, pg_rewind replays WAL using the checkpoint preceding the specified WAL location instead of the last common checkpoint.

Implemented the WAL trimming functionality that helps to avoid running pg_rewind to synchronize cluster nodes after failover when WAL divergence is caused only by redundant heartbeat records. The WAL trimming algorithm finds the last common point in WALs of the diverged node and the new leader, verifies that all subsequent records are only heartbeats, and automatically deletes redundant heartbeat records from the diverged node so it can resume replication. You can manage the WAL trimming functionality by means of the new biha.autowaltrim configuration parameter. For more information, see Automatic Cluster Synchronization after Failover.

Improved the LEADER_CHANGE_STARTED callback functionality. Previously, the callback could fail on the old leader after you set the new leader by biha.set_leader, as the old leader immediately restarts. Currently, the callback is delayed and then called on the old leader right after its restart. For more information, see the LEADER_CHANGE_STARTED description in the Callback Types table.

Improved the mechanism of managing relaxed synchronous replication in a BiHA cluster. Previously, it was required that you set the --sync-standbys-min option of bihactl init when initializing the cluster to be able to relax synchronous replication restrictions later. Currently, you can enable, modify, or disable relaxed synchronous replication at any time using the biha.set_sync_standbys_min function. The biha.sync_standbys_min parameter was deleted as redundant.

Fixed a bug where the biha.port value greater than or equal to 32768 caused the server to loop.

Improved the ability of the node to return to normal operation after catchup by accelerating the transaction filter building process at the start of walreceiver.

Fixed a bug that caused recovery hanging and potential data desynchronization when adding a new node.

Fixed unexpected errors in execution of DDL commands that were caused by incorrect processing of catalog cache invalidation messages.

Added support for CEF format for event log files.

Adjusted pg_proaudit output. xid and vxid fields are no longer displayed if they are zero.

Added the --wal-tree option for the add-instance command to store archived WAL files in a tree-like structure. This improves WAL operation performance with large volumes of logs.

Added the postgres-edition field to the show command output in the JSON format to display DBMS edition information.

Improved the --dry-run option operation during tablespace recovery. The operation estimation and plan output are now more precise.

Fixed an issue related to the retention window calculation. The time value is now taken from the host where pg_probackup is running. The retention window is now calculated correctly even with the timezone differences between the servers.

Fixed a problem with the archive-get command operation. The pbk_prefetch directory is now created with correct access rights, inheriting permissions from PGDATA.

Fixed false failure of the checkdb command in the remote mode during concurrent TRUNCATE operations or table drops on standby nodes.

Fixed potential pg_probackup3 failure caused by temporary tables during backup validation using the pgpro_bindump module in the PRO mode.

Fixed a memory leak in pg_probackup3 using the pgpro_bindump module for PTRACK operations.

Fixed an issue that could cause infinite looping in pg_probackup3 when using the pgpro_bindump module.

Fixed an issue with multi-timeline backup in the PTRACK mode in pg_probackup3 using the pgpro_bindump module in the PRO mode.

Added the repack command for changing the compression algorithm and/or compression level of CFS files.

Added the ground command for replacing corrupted blocks with zeros to prevent potential operation failures.

Added the stat command to monitor tablespace statistics.

Added the new pgpro_multiplan.mode parameter that allows you to enable or disable the extension, the use of available plan types, as well as statistics collection.

Added the new pgpro_multiplan.aqe_mode parameter that allows you to enable or disable AQE-related features.

Added the baseline value for the plan_type field of the pgpro_multiplan_storage view to identify allowed plans.

Renamed the following entities:

Updated the set_aqe_trigger function, which now accepts the query text string when setting individual AQE trigger values for a query. The parameterized version of the query text is displayed in the aqe_triggers view.

Changed the default behavior of statistics collection. Now statistics are not gathered unless you enable this feature using the pgpro_multiplan.mode parameter.

Changed the default behavior of the set_aqe_trigger() function. Now this function does not work unless you enable individual AQE trigger values using the pgpro_multiplan.aqe_mode parameter.

Removed the pgpro_multiplan.enable parameter that is no longer required. Instead, use the pgpro_multiplan.mode parameter.

Removed the pgpro_multiplan.aqe_plans_auto_approve and pgpro_multiplan.aqe_collect_stats parameters. Instead, use the pgpro_multiplan.aqe_mode parameter.

Added the sample_string field to the pgpro_multiplan_storage view to store a non-parameterized query for which a plan was frozen or added to a set of allowed plans.

Added the planid field to the pgpro_multiplan_captured_queries and pgpro_multiplan_stats views. This field specifies the plan ID compatible with the pgpro_stats extension and allows you to get extended statistics by joining information from the pgpro_multiplan_stats and pgpro_stats_statements views.

Added the pgpro_multiplan_remove_baseline function that removes the specified plan from a set of allowed plans.

Fixed an issue with pgpro_multiplan failing to register queries involving the INTERVAL 'const' notation.

Fixed an issue where plan_hash was not calculated when adding plans for queries with custom hints to a set of allowed plans (baseline). Now such plans can be added to a baseline with correct plan_hash values.

Fixed a segmentation fault, which occurred on attempts to access the aqe_triggers view or call the aqe_triggers_reset function with disabled pgpro_multiplan. Now all functionality related to AQE triggers works regardless of the whether pgpro_multiplan is enabled.

Fixed an issue with processing too long queue names when using CREATE_QUEUE.

Corrected returned types for message reading functions.

Fixed an issue, which could cause data modification attempts on replicas after restart.

Implemented automatic invalidation of cached result sets that can be enabled using the new pgpro_result_cache.consistent parameter.

Added two new configuration parameters that will be useful for debugging purposes.

Added the schedule.clean_at_jobs_done and schedule.clean_at_jobs_done functions that can clean schedule.at_jobs_done by removing old records.

Provided the ability to change the next start time for a cron job that has next_time_statement by updating cron, date, dates, or rule using the set_job_attributes() function.

Implemented the ability to track utility commands by their query_id, as it is done in pg_stat_statements. Previously, utility commands were tracked based on strict comparison of their textual query strings, and therefore could not be properly normalized.

Fixed a segmentation fault that could occur when defining custom metrics. The error originated from a memory allocation issue, which was fixed.

Implemented the monitoring functionality that allows tracking the proxima state in a BiHA cluster. For more information, see Monitoring.

Improved the distribution strategy for read-only workloads on the Proxy-to-Follower port. Previously, read-only transactions within one session were distributed across cluster replicas according to the selected algorithm. Currently, proxima attaches a client session to a certain cluster replica to ensure that all transactions within this particular session are executed on the same replica.

Improved proxima troubleshooting capabilities by adding code location to the log file for messages of the error log level.

Fixed a bug where proxima was unable to process or break a client connection query in case of unexpected memory shortage that could cause hanging on the client side.

Fixed a bug where proxima could not release some Postgres Pro backend processes when disabled or enabled using the proxima.enabled configuration parameter that led to a growing number of idle backend processes and memory leaks.

Fixed a bug that caused a referee node to fail when added to a BiHA cluster with proxima enabled. The main symptom was the connection loss in pg_basebackup followed by the error message: “error: could not initiate base backup: server closed the connection unexpectedly”.

Fixed an issue with incorrect termination of the system thread in proxima.

Fixed a bug that caused termination of the proxima process after restart of the BiHA cluster.

Fixed an issue with incorrect proxima startup order after restarting a node in a BiHA cluster.

Fixed memory leak in libev caused by unoptimized resource allocation.

Fixed a bug where log messages about modification of proxima configuration parameters were duplicated by other processes.

Fixed a segmentation fault caused by invalid memory access when a buffer pointer was used after the underlying buffer was modified by another thread.

Fixed an index scan infinite loop when using multiple ORDER BY clauses.

Added the support for TLS and MD5.

Improved the error message for setting invalid OPT_SSL_VERIFYHOST.