Scott Stults
Category: Machine-learning
2025 was a turning point for the Learning to Rank (LTR) plugin. After several years as a community-driven add-on that sat slightly outside the core OpenSearch ecosystem, LTR officially stepped into its role as a first-class OpenSearch component. Over the past year, we invested deeply in modernization, security, test coverage, release engineering, and contributor onboarding. The result is a plugin that is vastly more stable, maintainable, and aligned with how the rest of OpenSearch is built and released.
This post reflects on the work we accomplished throughout 2025 and what we are preparing for next.
Modernizing LTR for the OpenSearch Ecosystem
Our first major milestone in 2025 was aligning LTR with OpenSearch’s build, test, and release expectations. This included:
- Migrating the plugin to the OpenSearch build system and CI/CD workflows
- Modernizing Gradle configurations (first to 8.14, and later to 9.2.0)
- Improving security by phasing out the deprecated
SecurityManagerin favor of a new Java agent - Converting
.ltrstore*indices into proper system indices - Upgrading core dependencies across the board (Lucene 10 / 10.2, SLF4J,
http5client, log4j packaging fixes) - Expanding and correcting XGBoost support, including proper parsing,
rank:ndcgandrank:mapobjectives, and missing-value handling - Strengthening test coverage, flakiness detection, and hybrid float comparison handling
Over the year, we delivered several stable releases—2.19.0, 3.0 (alpha, beta, final), 3.1, 3.2, and 3.3—and are now moving toward 3.4.0-SNAPSHOT with an expanded maintainer team and new codeowners.
December Plans: Test Coverage and Future Direction
While our test coverage has improved significantly this year—reaching roughly 50%—we aren’t satisfied yet. December will focus heavily on expanding unit test coverage and strengthening confidence in new changes.
We are also beginning to explore improvements to the Feature Store APIs and the logging associated with feature execution. Several promising ideas emerged late in the year, and we expect to publish early proof-of-concepts soon.
Contributor Growth and Community Engagement
One encouraging trend in 2025 was increased involvement from the community:
- 4 contributors made 2+ commits, compared to 3 in 2024
- 5 contributors made a single commit, up from 2 in 2024
- We triaged and closed 69 GitHub issues, ending the year with 24 open items that represent ongoing work or future ideas
Despite a relatively small maintainer base, the project did not stagnate—PRs moved consistently, discussions were active, and development stayed healthy throughout the year.
On the community side:
- The
#opensearch-learn-to-rankchannel on Relevance Slack grew to 192 members, and LTR questions there remained consistent throughout the year. - We did not see LTR-related activity on the OpenSearch forums this year, and we intend to track such engagement more carefully next year.
- We aim to incorporate metrics from the Linux Foundation’s Insights tool—once we resolve why the LTR repository is not yet being included.
Looking Ahead
2025 was the year LTR finally became a first-class citizen in the OpenSearch ecosystem. We modernized almost every layer of the plugin—build system, dependencies, security model, testing infrastructure, CI, release tooling—while also expanding contributor capacity and community engagement.
In 2026, we plan to build on that foundation:
- Expanding test coverage and reliability
- Continuing feature store improvements
- Growing community adoption
- Enhancing feature logging and introspection
- Introducing new metrics and visibility into project health
LTR is now positioned for long-term sustainability and innovation within OpenSearch, and we’re excited for what comes next.