Apache Ignite

Apache Ignite clustering component is based on the shared nothing architecture. The nodes are divided into two main categories - server and client. Server nodes are storage and computational units of the cluster that hold both data and indexes and process incoming requests along with computations. Server nodes are also known as data nodes.[8]

Client nodes are connection points from applications and services to the distributed database represented as a cluster of server nodes. Client nodes are usually embedded in the application code written in Java, C# or C++ that have special libraries developed.

Furthermore, Apache Ignite provides ODBC,[9] JDBC[10] and REST drivers as a way to work with the database from other programming languages or tools. The drivers utilize either client nodes or low-level socket connections internally in order to communicate to the cluster.

Ignite database organizes data in the form of key-value pairs in distributed "caches" (the cache notion is used for historical reasons because initially, the database supported the memory tier). Generally, each cache represents one entity type such as an employee or organization.

Every cache is split into a fixed set of "partitions" that are evenly distributed among cluster nodes using the rendezvous hashing algorithm. There is always one primary and zero or more backup copies of a partition. The number of copies is configured with a replication factor parameter.[11] If the full replication mode is configured, then every cluster node will store a partition's copy. The partitions are rebalanced[12] automatically if a node is added to or removed from the cluster in order to achieve an even data distribution and spread the workload.

The key-value pairs are kept in the partitions. Apache Ignite maps a pair to a partition by taking the key's value and passing it to a special hash function.

The memory architecture in Apache Ignite consists of two storage tiers and is called "durable memory". Internally, it uses paging for memory space management and data reference,[13] similar to the virtual memory of systems like Unix. However, one significant difference between the durable and virtual memory architectures is that the former always keeps the whole data set with indexes on disk (assuming that the disk tier is enabled), while the virtual memory uses the disk when it runs out of RAM, for swapping purposes only.

The first tier of the memory architecture, memory tier, keeps data and indexes in RAM out of Java heap in so-called "off-heap regions". The regions are preallocated and managed by the database on its own which prevents Java heap utilization for storage needs, as a result, helping to avoid long garbage collection pauses. The regions are split into pages of fixed size that store data, indexes, and system metadata.[14]

Apache Ignite is fully operational from the memory tier but it is always possible to use the second tier, disk tier, for the sake of durability. The database comes with its own native persistence and, plus, can use RDBMS, NoSQL or Hadoop databases as its disk tier.

Apache Ignite is a strongly consistent platform that implements two-phase commit protocol.[19] The consistency guarantees are met for both memory and disk tiers. Transactions in Apache Ignite are ACID-compliant and can span multiple cluster nodes and caches. The database supports pessimistic and optimistic concurrency modes, deadlock-free transactions and deadlock detection techniques.

In the scenarios where transactional guarantees are optional, Apache Ignite allows executing queries in the atomic mode that provides better performance.

Apache Ignite can be accessed using SQL APIs exposed via JDBC and ODBC drivers, and native libraries developed for Java, C#, C++ programming languages. Both data manipulation and data definition languages' syntax complies with ANSI-99 specification.

Being a distributed database, Apache Ignite supports both distributed collocated and non-collocated joins.[20] When the data is collocated, joins are executed on the local data of cluster nodes avoiding data movement across the network. Non-collocated joins might move the data sets around the network in order to prepare a consistent result set.

Apache Ignite provides machine learning training and inference functionality as well as data preprocessing and model quality estimation.[21] It natively supports classical training algorithms such as Linear Regression, Decision Trees, Random Forest, Gradient Boosting, SVM, K-Means and others. In addition to that, Apache Ignite has a deep integration with TensorFlow.[22] This integrations allows to train neural networks on a data stored in Apache Ignite in a single-node or distributed manner.

The key idea of Apache Ignite Machine Learning toolkit is an ability to perform distributed training and inference instantly without massive data transmissions. It's based on MapReduce approach, resilient to node failures and data rebalances, allows to avoid data transfers and so that speed up preprocessing and model training.[23]