Data Center Replication (DR) replicates cache data between clusters.
When working with multiple data centers, it is important to make sure that if one data center goes down, another data center is fully capable of picking its load and data. When data center replication is turned on, GridGain will automatically make sure that each cluster is consistently synchronizing its data to other data centers (there can be one or more).
Data Replication Modes
GridGain supports active-passive and active-active modes for replication.
In the active-passive mode, only one cluster (the master) interacts with the user application, and the other cluster (the replica) stores data purely for redundancy and failover purposes.
In the active-active mode, both clusters handle user requests independently and propagate changes to each other. In this mode, care must be taken to resolve conflicts between updates that happen on the same cache entry in both clusters. For more information on how to resolve conflicts, refer to the Conflict Resolution section.
How Data Replication Works
Let’s consider a simple scenario where data from one cluster (the master) is replicated to another cluster (the replica) in active-passive mode.
In the master cluster, a subset of nodes, called a sender group, is configured to connect to the remote cluster, and all data is transferred through this connection. You can have as many sender groups as you need. For example, you can configure different groups to replicate data to different clusters, or you can replicate different caches with different data transfer properties.
In the replica cluster, a subset of nodes, called receivers, is configured to receive data from the master cluster.
Now, when you update a cache entry in the master cluster, the update happens on the node that hosts the entry (the primary node), but the update is not immediately sent to the replica. Instead, this update and other cache entry updates are accumulated into batches on the primary nodes. When the batch reaches its maximum size, or if the batch has been waiting for too long, it is sent to one of the sender nodes in the master cluster. The sender node then transfers the updates to the replica.
Note that DR replicates only the content of caches. It does not copy cluster or cache configuration. This means that the replica cluster has to be configured manually and can have a different topology and settings.
By default, GridGain waits for a receiving node to send a confirmation that data has been transferred. You can configure this behavior with the
awaitAcknowledge property. It is highly recommended to keep this property at a default value for production environments.
You can configure DR to replicate clusters in many different combinations. For example, you can have cluster A replicating updates to several other clusters; or cluster A replicating to cluster B, and cluster B replicating to cluster C; or clusters A, B, and C replicating updates to each other.
Below are the recommended minimum for
Two sender and two receiver nodes to provide redundancy in case of connection issues.
2 cores and 8Gb of memory available for each node.
Sender nodes need less hardware in general.
Receiver nodes need better hardware in general.
IgniteCache.clear()operation does not affect remote clusters when using DR.
Data replication may not work properly for caches that load data using Data Streamers that disallow overwriting existing keys. If you are going to use data replication with data streamers, make sure the
allowOverwriteproperty of the data streamer is set to
Data Center Replication should not be used with caches that have access-based expiry policy.
Cache Interceptors are not invoked in the remote cluster when it receives updates from the master cluster.
Deactivation of the cluster can throw an exception on the sender nodes, which will cause the sender nodes to hang. If this happens, you will need to restart the sender nodes. To avoid this, stop the sender nodes first, if possible, before deactivating the cluster. We recommend using client nodes as senders to avoid this issue.