For the complete documentation index, see llms.txt. This page is also available as Markdown.

Architecture

How GridGain 9 is built: cluster architecture, consensus, data distribution, consistency, availability, and storage.

This section explains how GridGain 9 works internally — the concepts you need to reason about performance, availability, and data placement. It is background reading rather than task instructions; for operational procedures, see Cluster Management.

Start with Cluster Architecture for an end-to-end overview of the node components and how a request flows through the system, then explore the individual topics below.

The Big Picture

GridGain 8 and GridGain 9 share the same architecture at the level of ideas. What differs between the two versions is how each idea is implemented.

  • Distributed cluster. GridGain runs as a cluster of server nodes that together store and process data. Applications work with the cluster through clients and drivers.

  • In memory and on disk. Data can be kept entirely in memory or persisted to disk, combining in-memory speed with disk durability.

  • B+ tree storage. GridGain stores data in memory pages and organizes its indexes as B+ trees.

  • Hash partitioning. Each table or cache is split into partitions. A hash of a record's key determines its partition, and the partitions are spread across the nodes of the cluster.

  • Distributed copies. Each partition can be stored on several nodes, so data stays available when a node fails.

  • Data colocation. Related records that share a colocation key are stored on the same node, so queries and computations over them run locally.

  • SQL, key-value, and compute. You can work with the same data through SQL or a key-value API, and send computations to the nodes that hold the data.

How GridGain 9 Implements It

Concept
GridGain 9
Read more

Cluster membership and lifecycle

A new cluster is initialized once with cluster init. The cluster management group (CMG) and the metastorage are RAFT groups that manage topology and cluster metadata.

Memory and persistence

Storage engines, selected through storage profiles: aimem (in-memory page memory), aipersist (persistent page memory), RocksDB, and columnar storage.

Partitioning

Each table belongs to a distribution zone that sets its number of partitions. A hash of the colocation key selects the partition, and the Fair distribution algorithm assigns partitions to nodes.

Copies and consistency

Each partition is replicated REPLICAS times through RAFT consensus, and a primary replica holds a lease to serve operations.

Colocation

The colocation key, set with COLOCATE BY.

SQL engine

Apache Calcite.

Architecture Topics

How a GridGain 9 cluster is built and run — node components and request flow, RAFT consensus, cluster initialization and lifecycle, and fault tolerance.

How GridGain 9 distributes data across a cluster — distribution zones, data partitioning, and data colocation.

How GridGain 9 replicates partitions and keeps data consistent — consistency guarantees, RAFT replication, and high availability mode.

How ACID transactions work in GridGain 9 — transaction types, guarantees, isolation, MVCC, locking, timeouts, coordination, and monitoring.

Overview of the GridGain 9 storage system: how tables, distribution zones, storage profiles, and storage engines work together to store your data.

This page is: Copyright © 2026 MariaDB. All rights reserved.

spinner

Last updated

Was this helpful?