Clusters that are actively using cross-cluster replication require a careful approach to upgrades. Otherwise index following may fail during a rolling upgrade, because of the following reasons:
- If a new index setting or mapping type is replicated from an upgraded cluster to a non-upgraded cluster then the non-upgraded cluster will reject that and will fail index following.
- Lucene is not forwards compatible and when index following is falling back to file based recovery then a node in a non-upgraded cluster will reject index files from a newer Lucene version compared to what it is using.
Rolling upgrading clusters with cross-cluster replication is different in case of uni-directional index following and bi-directional index following.
In a uni-directional setup between two clusters, one cluster contains only leader indices, and the other cluster contains only follower indices following indices in the first cluster.
In this setup, the cluster with follower indices should be upgraded first and the cluster with leader indices should be upgraded last. If clusters are upgraded in this order then index following can continue during the upgrade without downtime.
Note that a chain index following setup can also be upgraded in this way. For example if there is a cluster A that contains all leader indices, cluster B that follows indices in cluster A and cluster C that follows indices in cluster B. In this case the cluster C should be upgraded first, then cluster B and finally cluster A.
In a bi-directional setup between two clusters, each cluster contains both leader and follower indices.
When upgrading clusters in this setup, all index following needs to be paused using the pause follower API prior to upgrading both clusters. After both clusters have been upgraded then index following can be resumed using the resume follower API].