Hi Tanishq,

  Thanks for the proposal! This is a solid direction for Hive to become a 
modern, managed platform. 
   I'm +1 for a single MCP inside Hive if we are just talking about HS2 and 
Metastore. As of now we might just need one MCP server product, modular 
internally, and keeping the MCP inside the Hive helps align with thrift 
evolution, ASF governance, and release packaging. 

  About the HMS metadata discovery, we can collapse the tools to search_tables 
+ get_partitions_by_filter with the HIVE-29681. Instead of keyword match and 
repeated thrift calls, it provides a hybrid, contextual search table experience 
with full, near up to date table details.

Best,
Zhihua

On 2026/08/24 11:15:06 Tanishq Chugh wrote:
> Hi Everyone!
> 
> I have recently been exploring around the development of a Model Context
> Protocol (MCP) Server for Hive. There is a ticket logged for Metadata
> Catalog MCP as well: HIVE-29370 [1]. While working on the same ticket, I
> realized that enabling the MCP to execute Read queries would be highly
> beneficial, as it would bring on the natural language querying capabilities
> through LLMs.
> 
> However, there are two critical questions / decisions to make:
> 
> 1. Should we have separate MCP servers for Metadata (HMS) & Query Execution
> (HS2), or one MCP server that can serve both?
> 2. Where do we plan to store the MCP code? Should we store it in our
> Apache/Hive repository itself or in a separate repository?
> 
> I have the opinion of going ahead with a single MCP, because a
> standalone HS2 MCP serving query execution capabilities would still require
> metadata for query formation by LLMs. Although metadata can be fetched
> using HiveQL queries in this case, running a query for every metadata call
> seems costly.
> I have developed a High Level Design document (attached to this mail) for
> an MCP server that serves both Metadata and Query execution capabilities,
> with the option to run individually as an HMS / HS2 MCP.
> 
> I would like to ask you all to review the design, and would appreciate your
> opinions and suggestions regarding this.
> 
> Thanks!
> 
> Best,
> Tanishq Chugh
> 
> [1] https://issues.apache.org/jira/browse/HIVE-29370
> 

Reply via email to