Added the cfs_log_verbose configuration parameter that controls the CFS log level of defragmentation messages. It allows reducing the number of CFS messages in the log file.

Added the cfs_gc_lock_file configuration parameter that sets the path to the lock file to be used to ensure running only one garbage collection worker at a time for several Postgres Pro servers.

Modified the logic of truncating *.cfm files in CFS to improve performance. Previously, files were processed even if the file size didn't change, which could be unexpectedly heavy. The new approach involves checking the file size and only truncating if the size differs from the desired size to reduce unnecessary system calls.

Reduced the overhead of file size determination during query planning, particularly for queries involving numerous partitioned tables, which could be especially costly for CFS. Now file size is determined directly from the header without fully opening the file.

Fixed an issue in CFS, which could cause GC worker to fail with warning “CFS GC failed to read block 0 of file X at position 0 size 0: Success”. It happened due to incorrect handling of the first MB of the data file containing only zero pages.

Implemented the ability to add the referee node in biha, which allows setting up the 2+1 high-availability cluster with two regular nodes and one referee node. The referee is used only during elections and helps to avoid potential split-brain issues if your cluster contains only two nodes, i.e. the leader and one follower. Note that it does not contain any user databases and cannot be used for data querying.

Fixed an issue that could cause a cluster to fail when trying to call the biha.set_leader function to promote each follower manually.

Fixed a segmentation fault that could occur when a node started after a manual rewind using pg_rewind. Now the node can restore automatically after the rewind.

Fixed a bug related to alive node removal using biha.remove_node, which could cause leader re-elections after the removal. Now the node must be stopped before removing.

Added support for the Zabbix 6.4 API: handling of deprecated parameters for authentication requests.

Removed caching of the pgsql.connections[max_connections] metric.

Updated the default log rotation rules.

Prepared for support of Python 3.12.

Changed metric names of the pg_locks plugin. Be aware that the changes could break custom user-defined triggers and processing functions that use the item.name parameter.

Fixed type mismatch for pgpro_stats and pg_wait_sampling.

Fixed privileges for the mamonsu role created by bootstrap.

Added a possibility to limit the rate of disk writes using the option --write-rate-limit=bitrate (Mbps, Gbps).

Decreased the memory consumption when restoring long sequences of increments twice on average.

Added checksum validation for CFS files by checkdb.

Added a possibility to validate only a WAL archive.

Extended the use of the --dry-run option for all pg_probackup commands.

Made creation of a temporary slot during backups in the STREAM mode the default behavior unless it is specified otherwise.

Changed the default compression algorithm to zstd. If zstd is not supported by the system, lz4 has the next priority. The --compress option now sets the default values for --compress-level and --compress-algorithm.

Added a possibility to specify several hosts to connect to an S3 storage.

Implemented a new locking technique, which enables using the locks with S3 and NFS protocols.

Added saving non-normalized plans to pgpro_stats for queries where previously no plans were saved.

Fixed an issue that hindered the monitoring when the pgpro_stats_statements view contained quite a few rows with the same values of plan and queryid, but different values of planid. The issue was caused by an error in parsing the plan tree containing a T_Memoise node.

Added the sr_plan_hintset_update function, which allows changing the generated hint set with the set of custom hints.

Added the sr_plan.max_local_cache_size configuration parameter, which allows setting the maximum size of local cache, in kB. Also, the default value of sr_plan.max_items has been changed to 100.

Restricted query registration process so that only one query can be registered per backend.

Improved the algorithm for identifying a frozen query.

Implemented storage of query plans as separate JSON files.

Implemented type handling to attempt casting constants in the query to match the types of frozen query parameters automatically. If type casting is not possible, the frozen plan is ignored.

Removed the query tree validation for hint-set plans, allowing the use of hint-set plans during table recreations, field additions, etc.