EDB Postgres Distributed 6.3.0 release notes v6.3.0

Connection Manager now initializes automatically on subscriber-only nodes. This release also adds the /connection/is-ready-ro and /connection/is-ready-rw HTTP endpoints to provide granular readiness checks for read-only and read-write traffic.

Introduced the pgd raft set-leader command, allowing users to manually change the Raft leader for a specific group.

Added commands to enable or disable consensus (Raft) across different nodes, groups, or cluster levels.

Enhanced the node group removal logic to automatically drop replication sets of the same name.

Enhanced how the Connection Manager matches database names and usernames against pg_hba.conf fields. Key improvements include support for sameuser matching, refined samerole logic, and expanded +rolename matching to include the role itself and indirect members. Additionally, quotes are now stripped from usernames before matching to ensure consistent authentication behavior.

This enhancement allows the Connection Manager to use pg_ident.conf for user name mapping. Currently, this functionality applies to the cert authentication method, aligning Connection Manager behavior with PostgreSQL’s native authentication identity mapping.

Enhanced bdr.reconcile_2pc_requests to include more details such as request time, origin LSN, and commit timestamp. Additionally, added a user-friendly view, bdr.reconcile_2pc_requests_summary, to simplify monitoring of Two-Phase Commit (2PC) reconciliation.

Enhanced group commit to perform eager conflict detection and resolution when the origin of a transaction changes. Conflict handling now relies on timestamp-based detection rather than replication-progress speculation. To ensure optimal performance and avoid increased commit latency, you must synchronize system clocks across all nodes.

Introduced a new system view, bdr.stat_reconcile_commit_scope, which provides real-time statistics for commit scope reconciliation.

Introduced new options for bdr.node_group_config, which allow the cluster to select the most advanced candidate replica during leader transitions:

• leader_progress_tracking_enable: Periodically reports replication progress to the Raft leader.
• leader_progress_tracking_period: Configures the reporting interval.

Renaming galloc sequences or moving them between schemas is now fully supported. This is enabled by a global UUID identifier (sequuid) in the bdr.sequence_kind catalog, which allows PGD to track these sequences consistently and concurrently. This functionality works even for sequences upgraded from versions prior to PGD 6.3, as PGD automatically generates the required UUIDs for existing sequences to ensure compatibility.

Introduced bdr.prepared_xacts_summary, a local diagnostic view for monitoring prepared transactions implicitly created by group commit scopes.

The standalone PGD CLI for macOS is now available exclusively for the Apple Silicon (ARM64) architecture. This change applies only to the standalone macOS binary; the availability and support for standalone PGD CLI binaries on GNU/Linux remain unchanged.

Resolved an issue where logical replication would enter an infinite crash loop if a DETACH PARTITION CONCURRENTLY command was interrupted. Previously, the walsender would fail with the error relation has no parent because it's being detached when decoding WAL for a partition in a detach-pending state.

Fixed an issue where the walsender process could enter a crash loop when encountering partitions with inhdetachpending = true, which reported ERROR: relation %u has no parent failures.

Fixed an issue where replication would break because attributes like NOSUPERUSER or NOREPLICATION were automatically added to the replicated command. Since these options can only be altered by a superuser, the apply process failed on downstream nodes. This change ensures the replicated command only includes options that were explicitly specified and for which the user has sufficient privileges.

The command now attempts to fetch the encoding value directly from the remote node, defaulting to UTF8 if unavailable.

Corrected the help text for pgd group set-option to accurately reflect group-level options.

Inactive/disconnected slots now display ! instead of * when lag is below 10 MB, restoring consistency.

Prevented shared memory queues used by consensus from growing while a node is unreachable, ensuring that memory remains stable during node outages.

Resolved a malloc-based memory leak related to the improper freeing of connection strings.

Resolved a bug where Connection Manager incorrectly ignored the use_https setting. When upgrading to PGD 6.3.0 or 5.9.2, you must explicitly enable HTTPS by setting the use_https group option to true (in addition to having ssl_cert_file and ssl_key_file configured in postgresql.conf or via the command line) before the upgrade to maintain previous behavior.

Resolved a memory leak in the Connection Manager caused by incorrect handling of Postgres memory contexts during configuration reloads.

Resolved an issue where Postgres would crash with a segmentation fault when executing a COPY command if the pgaudit.role configuration was set to a non-default value.

Improved the coordination between data retention and analytics offloading. Partitions are now offloaded using Postgres Analytics Accelerator (PGAA) before the data retention policy is applied.

In cluster topologies with subgroups and analytics enabled in the parent group, a bug caused each subgroup's write leader to start its own analytics replicator instance, resulting in duplicate rows. This fix ensures that only one analytics replicator per PGD group runs across the cluster.

Resolved a race condition in the Connection Manager where an unexpected server disconnection during the retrieval process from a connection pool would trigger an infinite loop. This issue was specifically limited to clusters operating in session pooling mode.

Table and index locks are now held until transaction commit when applying changes. This ensures correct serialization and prevents data consistency issues during concurrent operations, specifically addressing risks introduced by parallel apply.

Resolved an issue where enabling Parallel Apply on community PostgreSQL caused writer hangs and stuck subscriptions due to frequent lock timeouts. This fix allows Parallel Apply to be used reliably without requiring it to be disabled as a workaround.

Connection Manager now automatically recovers from file descriptor exhaustion. If the system limit is reached, the process will close listening sockets and automatically restart after 30 seconds to restore connectivity.

Resolved a race condition where bdr.create_node_group could return an error even if the group was successfully created. The function now correctly waits for node creation to fully complete before returning.

Fixed an issue where proxy configurations were not preserved after running bdr_init_physical due to Raft snapshot restores failing to persist proxy settings.

Fixed a bug where sequence_alloc entries persisted after a galloc sequence was dropped and recreated. The fix ensures proper cleanup propagation across the cluster, including witness nodes.

Fixed an issue where changing a sequence kind to galloc would silently discard the provided start_value. Additionally, resolved a bug where a startval of 0 was lost during WAL replication.

Fixed an issue where the analytics replicator on a previous write leader could continue running simultaneously with the replicator on the new leader. This synchronization improvement ensures the old process stops completely before the new one initializes, preventing duplicate data inserts. This fix includes the addition of the bdr.group_lease catalog table and the bdr.group_lease_override function.