A data grid is a cluster of (typically commodity) servers, normally residing on a single local-area network, connected to each other using IP based networking. Data grids behave as a single resource, exposing the aggregate storage capacity of all servers in the cluster. Data stored in the grid is usually partitioned, using a variety of techniques, to balance load across all servers in the cluster as evenly as possible. Data is often redundantly stored in the grid to provide resilience to individual servers in the grid failing i.e. more than one copy is stored in the grid, transparently to the application.

Data grids typically behave in a peer-to-peer fashion. Infinispan, for example, makes use of JGroups as a group communication library and is hence biased towards a peer-to-peer design. Such design allows Infinispan to exhibit self-healing characteristics, providing service even when individual servers fail and new nodes are dynamically added to the grid.

Infinispan also makes use of TCP and optionally UDP network protocols, and can be configured to make use of IP multicast for efficiency if supported by the network.