- Update to 25.3.29
  - A bug in GCS where JOIN message was delivered even if the node
    was in DONOR state was fixed.
  - An issue where GCache could contain mixed histories from
    different clusters was fixed.
  - Setting socket.recv_buf_size was not effective because it was
    done after the socket was connected or accepted. Also the
    default value caused TCP receive buffer autotuning to be
    disabled. This lead to suboptimal performance in high bandwidth
    WAN clusters.
    The default value for socket.recv_buf_size was changed to
    'auto' which lets the kernel to tune TCP receive buffer. A new
    variable socket.send_buf_size with default value 'auto' was
    added to allow also send buffer tuning.  (codership/galera#552)
  - GComm socket timestamping/liveness checking caused false
    positives during replication of large transactions, which
    caused excessive amounts of broken connections.
    (codership/galera#553)
  - Large transactions were able to monopolize bandwidth when
    segmentation was configured. This caused messages which were
    relayed by segment representative to be delayed. As a fix
    implemented fair queuing of messages. (codership/galera#553)
  - GComm EVS layer did excessive broadcasting to retransmit
    missing messages when packets were lost or delayed. The
    communication protocol was optimized to use point-to-point
    messaging to deliver missing messages and to rate limit
    retransmission requests. (codership/galera#554)
  - Because of a bug in quorum computation, two primary conflicting
    primary components were formed when the group merged and
    partitioned again while the new primary view was forming.

• Files Changed
• Browse Source