EDB Postgres Distributed 5.6.1 release notes v5.6

Support for Postgres 17 has been added for all flavors (PostgreSQL, EDB Postgres Extended, and EDB Postgres Advanced Server) starting with version 17.2.

Support ARM architecture for EDB Postgres Distributed on Debian 12 and RHEL 9.

The function bdr.wait_node_confirm_lsn() has been introduced to wait until a specific node reaches a designated Log Sequence Number (LSN). It first checks the confirmed_flush_lsn of the replication slot for the specified node. If that information is not available, the function connects to the node and queries pg_replication_origin_progress(), using the invoking node as the origin. If the nodename parameter is NULL, the function will wait for all nodes to reach the specified LSN. If the target LSN is NULL, it will wait for the current wal_flush_lsn.

An update addresses the movement of group slots in SO nodes, ensuring they don't appear as peers in progress updates. Improvements include enhanced watermark management for SO leaders in the Optimized Topology configuration, where write leaders now include watermarks in their updates. Watermarks are broadcasted to simplify progress tracking on idle clusters. The peer progress mapping for SO nodes has been corrected, and the tap test for group slot movement has been revised. Additionally, the bdr_get_all_origins function now considers SO node origins.

While there are no connections from non-leader data nodes to subscriber-only nodes in an optimized topology configuration, the LSN progress of all data nodes is periodically communicated to these subscriber-only nodes through logical replication.

The bdr.permit_unsafe_commands parameter now allows some DDL commands that were previously disallowed. Specifically ALTER COLUMN...TYPE...USING can now be permitted if the user knows the operation is safe.

Ensure that pglogical GUCs are overridden only when operating within the pglogical worker. If this is not the case, MyPGLogicalWorker will be NULL, resulting in a segmentation fault when the walsender attempts a configuration reload from the pgl_wait_for_standby_confirmation() function.

The msgbroker module used to establish consensus connections lazily, meaning that connections were created only when the first message was sent to a specific destination. This method negatively affected the latency of Raft leader elections. The behavior was modified to create connections to consensus peers eagerly. However, this change resulted in an unintended consequence: a fully meshed consensus network among subscriber-only nodes, which may conflict with customer network designs. This patch keeps the eager connection setup but limits it to voting nodes only, reverting to a lazy connection setup for non-voting nodes.

To improve reliability, shuffle the scheduling of autopartition tasks. This way, tasks that are prone to failure won't consistently impact the success of other tasks.

The parting subscription used to hang, failing to wait for standbys when the bdr.standby_slot_names parameter was defined.

All relevant origins must be removed when parting SO node. With Optimized Topology, parting an SO node should result in removing all origins it has, not just the one related to its SO group leader. When parting a data node, even though there is no subscription to it from SO node, the origin should be removed. DO not make SO node target of a part catchup subscription when Optimized Topology is enabled.

Subscriber only nodes should not have slots on witness nodes.

When using commit scope with DEGRADE clause, if system detects that it's in degraded state, transactions should start in the DEGRADE mode. This ensures that the timeout is not applied on every commit.

Corrected routing strategy for read nodes after a network partition.