Skip to content

Commit e00718d

Browse files
agrawalreetikaagrawalreetika
committed
Add RFC for supporting MCP with Presto
1 parent efafb61 commit e00718d

File tree

1 file changed

+127
-0
lines changed

1 file changed

+127
-0
lines changed

‎RFC-0022-Model-Context-Protocol.md‎

Lines changed: 127 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,127 @@
1+
# **RFC-0014 for Presto**
2+
3+
## Presto Model Context Protocol Support
4+
5+
Proposers
6+
7+
* Reetika Agrawal
8+
9+
## Summary
10+
11+
This proposal introduces support for integrating Model Context Protocol (MCP) with Presto through a separate lightweight process, referred to as the MCP Server.
12+
The MCP Server acts as a protocol translation layer between AI agents using JSON-RPC and Presto’s existing HTTP-based query protocol.
13+
14+
## Background
15+
16+
Presto exposes a RESTful HTTP API for query submission and result retrieval. Clients submit a query via a POST request and receive a tokenized nextUri chain for incremental result fetching.
17+
However, AI frameworks like OpenAI MCP operate on a request–response model over JSON-RPC, where clients expect complete results in a single response, not streamed batches.
18+
Because MCP and Presto differ fundamentally in communication models, a direct integration is impractical. Additionally, direct embedding of MCP into Presto coordinators would disrupt routing and proxy layers that rely on Presto’s native HTTP semantics.
19+
20+
### Goals
21+
22+
- Enable MCP-compatible AI agents (e.g., OpenAI ChatGPT tools) to query Presto seamlessly.
23+
- Preserve all existing Presto router, proxy, and load-balancing infrastructure without modification.
24+
- Simplify authentication by reusing existing OAuth/JWT mechanisms.
25+
- Prevent unbounded queries by applying automatic limits when appropriate.
26+
27+
### Proposed Plan
28+
29+
Introduce a new lightweight service: presto-mcp-server, deployed alongside Presto coordinators and Presto Router.
30+
31+
The MCP server will:
32+
33+
- Expose a JSON-RPC 2.0 HTTP endpoint (`/mcp`, `/v1/mcp`) interface to AI clients.
34+
- Implement the core MCP primitives:
35+
- `tools/list` for tool discovery
36+
- `tools/call` for executing tools
37+
- Initially provide a single tool: `query.run`, which executes SQL queries against Presto.
38+
- Internally communicate with Presto coordinators using standard Presto HTTP APIs.
39+
- Forward OAuth/JWT Bearer tokens transparently from MCP clients to Presto, ensuring that Presto performs all authentication and authorization checks.
40+
- Translate between the two protocols, aggregating streaming results into a single response.
41+
- Remain stateless, delegating all query lifecycle management to Presto.
42+
43+
## Proposed Implementation
44+
45+
#### Core Changes
46+
47+
```json
48+
JsonRpcServlet → McpDispatcher → ToolRegistry → QueryRunTool → PrestoQueryClient → Presto Coordinator
49+
```
50+
51+
1. New Module: `presto-mcp-server`
52+
53+
- Implements JSON-RPC 2.0 protocol.
54+
- Implements core MCP primitive like `tools/list` and `tools/call`
55+
- Handles methods like query.run.
56+
57+
2. On `query.run`:
58+
59+
- Parses SQL input.
60+
- Optionally injects a LIMIT clause (if absent) to control data size.
61+
- Submits the query to Presto coordinator via /v1/statement.
62+
- Polls the returned nextUri until the query completes.
63+
- Returns final aggregated results as a single JSON-RPC response.
64+
65+
#### Example Queries
66+
67+
- `tools/list`
68+
```json
69+
{
70+
"jsonrpc": "2.0",
71+
"id": 1,
72+
"method": "tools/list",
73+
"params": {}
74+
}
75+
```
76+
77+
- `tools/call → query.run`
78+
```json
79+
{
80+
"jsonrpc": "2.0",
81+
"id": 2,
82+
"method": "tools/call",
83+
"params": {
84+
"name": "query.run",
85+
"arguments": { "sql": "SELECT 1" }
86+
}
87+
}
88+
```
89+
90+
### Rationale
91+
92+
- Introduce a new standalone service (presto-mcp-server) to avoid mixing JSON-RPC with Presto’s stateful HTTP protocol.
93+
- Translate MCP tool calls into Presto HTTP queries, using the existing StatementClient to follow nextUri pages and aggregate results into a single MCP response.
94+
- Keep the MCP server stateless, with all query lifecycle state remaining on Presto coordinators.
95+
- Forward OAuth/JWT Bearer tokens directly from MCP clients to Presto, allowing Presto to perform full authentication and authorization without changes.
96+
- Preserve all existing Presto infrastructure (Router, proxies) by keeping MCP outside the coordinator and communicating using standard Presto HTTP APIs.
97+
98+
## Backward Compatibility Considerations
99+
100+
- MCP server is a new optional component
101+
- No impact on existing Presto or Router
102+
- All existing Presto deployments remain unchanged
103+
104+
## Test Plan
105+
106+
- **Unit + Integration Tests:**
107+
Verify ToolRegistry loading, dispatcher routing, SQL execution via QueryRunTool, and JSON-RPC error handling.
108+
109+
- **End-to-End Validation:**
110+
Deploy MCP server with Router + Coordinator. Confirm Bearer token forwarding works and MCP clients (Gemini, ChatGPT) successfully execute queries.
111+
112+
## Modules involved
113+
- `presto-mcp` (new module)
114+
- `presto-client`
115+
- `airlift` modules
116+
- `presto-main`
117+
- `presto-router`
118+
119+
## Final Thoughts
120+
121+
This proposal cleanly bridges Presto with next-generation agent ecosystems (LLMs, AI workflows, model interaction tools). The MCP server architecture respects Presto’s deployment patterns, is backward-compatible, and provides a robust extension point for future interactive functionality such as:
122+
123+
- Schema browsing tools
124+
- Table metadata tools
125+
- Query explanation tools
126+
127+
## WIP - Draft PR Changes

0 commit comments

Comments
 (0)