[RFC] Migrating existing OpenSearch plugins from the main OpenSearch repository into their own dedicated repositories

[pgtgrly]

Please describe the end goal of this project

Glossary

• OpenSearch server Version: The release version of OpenSearch server without any plugins.
• Plugin Version: The release version of plugins, this is independent of OpenSearch server version.
• Managed Plugins: Plugins developed and released by OpenSearch Project
• Core Repository Plugins : Managed Plugins in the OpenSearch core repository (eg analysis-nori)
• Default Plugins : Managed Plugins that are bundled with standard OpenSearch distribution (eg security plugin).

Motivation

This document proposes, migrating existing plugins in core OpenSearch repository into their own dedicated repository

Background

In addition to the background in https://github.com/opensearch-project/OpenSearch/issues/17127, There are a few more things we need to dive into before we discuss the problem statement.

Default and Core repository plugins

Currently the core OpenSearch repository contains 28 plugins:

1. analysis-icu
   

2. analysis-kuromoji
   

3. analysis-nori
   

4. analysis-phonenumber
   

5. analysis-phonetic
   

6. analysis-smartcn
   

7. analysis-stempel
   

8. analysis-ukrainian
   

9. cache-ehcache
   

10. crypto-kms
    

11. discovery-azure-classic
    

12. discovery-ec2
    

13. discovery-gce
    

14. identity-shiro
    

15. ingest-attachment
    

16. mapper-annotated-text
    

17. mapper-murmur3
    

18. mapper-size
    

19. repository-azure
    

20. repository-gcs
    

21. repository-hdfs
    

22. repository-s3
    

23. store-smb
    

24. telemetry-otel
    

25. transport-grpc
    

26. transport-nio
    

27. transport-reactor-netty4
    

28. workload-management
    

Whereas the standard opensearch distribution contains the following 23 plugins:

1. opensearch-alerting
   

2. opensearch-anomaly-detection
   

3. opensearch-asynchronous-search
   

4. opensearch-cross-cluster-replication
   

5. opensearch-custom-codecs
   

6. opensearch-flow-framework
   

7. opensearch-geospatial
   

8. opensearch-index-management
   

9. opensearch-job-scheduler
   

10. opensearch-knn
    

11. opensearch-ml
    

12. opensearch-neural-search
    

13. opensearch-notifications
    

14. opensearch-notifications-core
    

15. opensearch-observability
    

16. opensearch-performance-analyzer
    

17. opensearch-reports-scheduler
    

18. opensearch-security
    

19. opensearch-security-analytics
    

20. opensearch-skills
    

21. opensearch-sql
    

22. opensearch-system-templates
    

23. query-insights
    

There is no overlap between the plugins that are part of the standard distribution plugins and the plugins which are present in the core repository.

Most of the plugins present in core repository, are not subject to active changes after initial development and go through dependency version bumps that need to be executed by the core maintainers.
These core plugins are not separately released in maven and require the consumers to build it by themselves to install in their OpenSearch cluster. The core plugins lack documentation as well.

Plugin artifact publication

Managed plugins are published in multiple repositories:

• OpenSearch Managed repositories: Repositories like artifacts.opensearch.com and ci.opensearch.com. However it is not possible to browse these repositories to navigate to a plugin zip. One has to know the exact URL to be able to download the plugin.
• Maven: Only the default managed plugin zip [example] is published in maven. Core repository zips are not hosted there [example]

Previous discussion

reta published a RFC in 2021 to migrate the plugins out of the core repository #1754. He did a PoC as well where he moved the plugins out of the core repository into a single repository as well. In Issue 7962 It was called out that moving the plugins out of core repository would removed multiple 3P dependencies from the core repo.

Problem Statement

The current inclusion of 28 plugins within the OpenSearch core repository creates several operational challenges. These plugins lack active maintenance, proper documentation, and Maven publication, requiring users to build them manually. Updates to these plugins, primarily dependency version bumps, unnecessarily burden core maintainers due to their tight coupling with the core repository, which also prevents independent versioning and release management. Additionally, the presence of plugin-specific third-party dependencies bloats the core repository. This RFC proposes migrating these plugins to dedicated repositories to improve maintainability, documentation, and release processes while reducing complexity in the core repository.

Proposed Approach

We propose migrating the all the plugins from the core repository to their own repository. The migration process will begin by identifying owners for each plugin who will be responsible for their respective repositories. New dedicated repositories will be created following the naming convention 'opensearch-project/opensearch-' and will include standard components such as Apache 2.0 LICENSE, Code of Conduct, CONTRIBUTING.md, GitHub Actions CI for tests and scanning, and comprehensive README documentation. The code migration will preserve Git history using git subtree split, replace existing plugin folders in the core repository with redirecting READMEs, and update official documentation accordingly. Each repository will have its own CI/CD infrastructure for running integration tests, enabling independent releases that align with core major versions, and maintaining plugin-specific documentation. This migration can proceed independently of the OpenSearch 3.0 release timeline, starting with a single plugin to establish a standard operating procedure for subsequent migrations.

Pros compared to the status quo:

• Decouples plugin code from core repository, removing bloatware and maintenance overhead
• Creates a standard release process and release notes for each plugin
• Creates documentation for each plugin
• This can be coupled with RFC [RFC] Decoupling OpenSearch Distribution, Server and Plugin versioning #17127 to maintain further independent releases

Cons:

• Will require additional head count to ensure maintenance and release of each plugin repository

Alternatively, we can just migrate the plugins to a single separate repository as reta proposed. This will remove the bloatware from the core repository

FAQ

Q. What will the plugins folder in core repository be used for once these plugins are migrated out?
A. It can be used as a step in migrating features from plugins to core opensearch (like security), it can also contain example and temples for custom plugins.

Q. How will this affect the release process for the core OpenSearch distribution?
A. This will not impact the standard distribution as none of the core repo plugins are shipped with standard distribution

[andrross]

Just one note...the existing issue #1754 was focused only on the repository-xxx plugins, of which there are 4 in this repo. This is the first issue I've seen suggesting that we move all plugins out of this repository.

Now don't get me wrong, I'm a huge fan of doing anything to reduce the amount of code being maintained here (see #16887), but there are a lot of plugins here that are essentially a single class. Examples include the analysis-xxx and mapper-xxx plugins. I suspect the small plugins with zero/minimal additional dependencies are likely not worth the overhead of a dedicated repository. I wonder if there is a middle ground where the heavyweight plugins could be moved out, thereby providing most of the benefits of reducing bloat and maintenance burdens in this repository, while also reducing the additional overhead. It might also be worthwhile to consider grouping plugins in separate repositories, such as putting repository-s3, discovery-ec2, and crypto-kms plugins together in a repo. All depend on the AWS SDK and maintainers are very likely to be the same individuals.

Q. How will this affect the release process for the core OpenSearch distribution?
A. This will not impact the standard distribution as none of the core repo plugins are shipped with standard distribution

I don't think this is accurate. The bundled plugins today are special in that they can be installed by name. If I extract a distribution tarball, I can then run ./bin/opensearch-plugin install analysis-icu to install the analysis-icu plugin. I think this will not work for a non-bundled plugin.

[peterzhuamazon]

Infra side this is doable but we do need to understand that:

1. All new plugins also need to have separate integTest onboard to Jenkins, which is already causing a lot of resources to run now on the existing 23 plugins.
2. We will not be able to install plugins with opensearch-plugin install repository-s3 directly if they are not core-plugins anymore. As non-core-plugins as of now does not have the mechanic to do so. User would need to manually download the zip and install, or we need to bundle them like the other 23 plugins.
3. Adding above 2 would require changing the default install url logic in core + changing folder structure of the existing bundled plugins, which is a pretty big breaking change I would say.
4. Owners of these new plugins would need to participate in the release process from now on.
5. There are plans to validate core plugin installations at one point from infra side: Validate native plugin installation opensearch-build#2859
6. I agree with @andrross that we might be able to combine similar functioned plugins into a single repo or plugin scope to reduce overhead.

cc @getsaurabh02

Thanks.

[lukas-vlcek]

@peterzhuamazon Just a quick comment on point 2)
OpenSearch plugins can be published into maven https://opensearch.org/blog/opensearch-plugin-zips-now-in-maven-repo/ and since this update the groupId can be fully customized which means that opensearch-plugin utility could be extended to search over (pre-)configured MVN repositories as well (which might not be seen as a breaking change IMO).

[peterzhuamazon]

@peterzhuamazon Just a quick comment on point 2) OpenSearch plugins can be published into maven https://opensearch.org/blog/opensearch-plugin-zips-now-in-maven-repo/ and since this update the groupId can be fully customized which means that opensearch-plugin utility could be extended to search over (pre-)configured MVN repositories as well (which might not be seen as a breaking change IMO).

Right, tho we are currently still pulling from S3 instead of maven for released core plugin installation.
We can definitely try to use maven but that requires each plugin repo to maintain a corresponding build task carefully.
Also maven is not as easy to manage as S3 so I do have a bit concern on that front. cc @prudhvigodithi .

[krisfreedain]

Catch All Triage - 1 2