The OpenSearch Software Foundation, part of the Linux Foundation, has released OpenSearch 3.0, with experimental GPU-accelerated vector indexing, MCP (model context protocol) support, and modernization of its core Java code.
Vector databases are a key feature of AI workloads. An experimental feature ads GPU acceleration with a claimed speed-up in index builds of up to 9.3 times.
Native MCP support enables AI agents to communication with OpenSearch. Vendors have hastened to add MCP support to their products since without it, competitors could be at an advantage in agentic software development.
The gRPC protocol has been added to OpenSearch 3.0 as another experimental feature, providing a “new approach to data transport between clients, servers, and node-to-node,” according to today’s release from the Foundation.
Other new features include pull-based ingestion, which is an experimental approach to pulling data from streaming systems such as Apache Kafka that is said to be more efficient and controllable. Apache Lucene 10, released in October 2024, is now used for indexing and search.
There are also core upgrades to the Java code with which OpenSearch is written. The minimum Java version is now Java 21, legacy code has been removed, and use of the Java platform module system introduced, the long-term aim being to refactor the existing monolithic server module into libraries.
OpenSearch was introduced by AWS in September 2021 as a fork of ElasticSeearch and Kibana (for data analysics), following a long-running dispute between AWS and Elastic and the removal of the Elastic open source licence. In September 2024, OpenSearch joined the Linux Foundation, via a newly formed OpenSearch Software Foundation.
The battle now is for mindshare, with OpenSearch presented as a community-owned open source project, and Elastic trying to recapture community engagement with a return to open source in September last year.
A performance comparison conducted in March by security company Trail of Bits found that OpenSearch is now faster than Elasticsearch for some operations, though with exceptions and the caveat that benchmarking is a moving target because both products have frequent updates. Elastic 9.0 was released last month as a major upgrade with many new features.
OpenSearch has proved viable, though it retains an AWS flavor, and OpenSearch Foundation members include AWS but not Microsoft or Google. Users tend to prefer the permissive OpenSearch Apache 2.0 license to the AGPL copyleft license used by Elastic.