* Changed podManagementPolicy to parallel in order to allow recovery
from secondary or third master
After rebooting the whole cluster on purpose or after a power failure
a primary node the cluster can synchronize with is required. This is
usually done automatically by selecting the node with the highest
transaction id. The current implementation of the stateful set starts
the nodes in sequence, preventing the start of further nodes if the
process terminates with error state. Because of this, the cluster may
not come up if the first or second node are not in primary state.
* Elects first node started in primary state as bootstrap source.
* Display warnings and runs mysqld with wsrep-recover on crashed nodes
* Introduces FORCE_RECOVERY argument for crash recovery
In case the primary selection failed, the cluster bootstrap process
must be manually initiated from the most advanced node (highest
committed transaction id). This information is available from the
grastate.dat file in case of a clean shutdown. On crashed nodes an
InnoDB recovery is required to get the last committed transaction id.
start.sh will handle both cases and gives instructions on how to
recover the cluster on a hard failure. If FORCE_RECOVERY was set to
the name of a POD (mariadb-0, mariadb-1, ...), the bootstrap process
will be initiated from the specified node.
DocImpact
Closes-Bug: #1716461
Change-Id: I96a8cb52124f64920a7d9cf21a8924ede78ebf7b
d78f8e0901