CourseApache Doris HackathonCommunity over Code 2026 Glasgow · Tue 13 Oct · 11:20–15:00 · Wee Dram Room

Track A · Task A2

Ask Doris with MCP

Plug Doris into the AI assistant you already use — and make it answer real questions with real SQL.

45–90 min
Beginner-friendly
Python 3.12+ · any MCP client
AI tools welcome

Stuck for more than 10 minutes? Come to the Doris table and ask Mingyu, or post in #dev on the Apache Doris Slack.

Get set up

What you are building, why Doris makes it simple, and a running cluster.

What you'll build

Turn your AI assistant into a Doris data analyst.

The Apache Doris MCP Server exposes your cluster to any MCP-capable assistant. Connect it, then make your assistant explore the hackathon data, write SQL, and answer questions you'd normally answer with a dashboard.

A working setup — your own MCP client talking to your local Doris through the official Doris MCP Server 1.0 — plus a short write-up showing your assistant:

discovering the hackathon tables on its own,
answering at least three data questions by writing and running SQL,
refusing a destructive request (because the server and the user are read-only).

SketchWhat you'll build · wireframe, not a screenshot

The sketch sits at the top of the code panel

Why it's interesting on Doris

Read-only by design. The Doris MCP Server 1.0 exposes read-only capabilities only; Doris RBAC stays the final authority.
Eight domains, not a flat tool dump. Catalog, query, cluster, pipeline, search, governance, lakehouse and semantic domains, disclosed progressively so the model sees only what it needs.
Search-aware. The doris_search domain knows about text, vector and hybrid search, so your assistant can use Doris's search features instead of guessing.

Before you start

Docker is running and the Doris image is pulled. About 1.9 GB, once. Pull it before you arrive: the venue Wi-Fi is shared. Docker Desktop (macOS) and Docker Engine (Linux) both work.
docker pull apache/doris:all-in-one-4.1.3
Doris is running. One container runs one FE and one BE. It is ready when docker ps shows (healthy), about 30 seconds after it starts. About this image
docker run -d --name doris -p 9030:9030 -p 8030:8030 -p 8040:8040 apache/doris:all-in-one-4.1.3
You can connect. The container ships a MySQL client, so you don't need one. Any MySQL client works too: 127.0.0.1:9030, user root, no password.
docker exec -it doris mysql -uroot -h127.0.0.1 -P9030
You have the starter kit. Download and unzip doris-hackathon-glasgow-2026.zip, then run the commands on this page from inside its doris-hackathon-glasgow-2026 folder. Its doris.sh wraps them: start, load, sql, stop.
curl -LO https://github.com/morningman/demo-env/releases/download/for-hackathon/doris-hackathon-glasgow-2026.zip && unzip doris-hackathon-glasgow-2026.zip && cd doris-hackathon-glasgow-2026
Python 3.12 or later, with the server installed. Check with python3 --version. Install the server before you arrive (milestone 2 has the commands): it pulls in about 80 packages.
An MCP-capable AI client you already use (and its account).
All hackathon tables loaded. From the starter-kit folder. It ends with the row count of each table.
docker exec -i doris mysql -uroot -h127.0.0.1 -P9030 < seed/seed_all.sql

Host: 127.0.0.1
Port: 9030
User: rootno password
Database: hackathon

Build with AI

This task is already an AI task. The text below teaches your assistant to write Doris SQL. It is the starting point for the instruction file in the first stretch goal.

Instruction filepaste into your client

You are connected to Apache Doris 4.1 through the Doris MCP Server (read-only).Database: hackathon. Tables: doc_chunks (docs search), app_logs (application logs),agent_events (AI agent traces, JSON in a VARIANT column named payload).Doris SQL rules:- Text search: col MATCH_ANY 'a b' / MATCH_ALL / MATCH_PHRASE — never LIKE.- Relevance: score() only with a MATCH_* predicate + ORDER BY score() DESC LIMIT n.- Vectors: ORDER BY l2_distance_approximate(embedding, [...]) LIMIT n.- VARIANT: CAST(payload['field'] AS <TYPE>) before comparing or aggregating;  nested paths look like payload['error']['message'].- Time buckets: minute_floor(ts, 5) or date_trunc(ts, 'hour').Always show the SQL you ran together with the answer.

Doris traps your agent will probably fall into

Trap	What to do instead
Using `LIKE '%term%'` for text search	`col MATCH_ANY 'a b'` (OR), `MATCH_ALL` (AND), `MATCH_PHRASE 'a b'` — they use the inverted index
Calling `score()` in an arbitrary query	`score()` needs a `MATCH_` predicate in `WHERE` and* `ORDER BY score() DESC LIMIT n`; otherwise Doris rejects the query
`score()` inside `GROUP BY` / aggregates	Not allowed. For facet counts, reuse the same `WHERE` without `score()`
Comparing VARIANT paths without a cast	`CAST(payload['latency_ms'] AS INT)`; VARIANT can't be a key or join column

Build it

5 milestones, M1 → M5. The panel follows the step you are reading.

M1
Milestone 15 min
Create a read-only user
M1 · Read-only usersql
```
CREATE USER 'mcp_reader'@'%' IDENTIFIED BY 'hackathon';GRANT SELECT_PRIV ON internal.hackathon.* TO 'mcp_reader'@'%';
```
Never hand an LLM your root account — even on a laptop. Run it in a SQL shell, or load a2/create_reader.sql from the starter kit.
Checkpoint
SHOW GRANTS FOR 'mcp_reader'@'%' shows Select_priv on internal.hackathon, and nothing that writes.

Milestone 215 min

Install and register the server

M2 · Installbash

python3 -m venv .venv && . .venv/bin/activate   # Python 3.12+pip install doris-mcp-server==1.0.0which doris-mcp-server          # note the absolute path

Add it to your client's MCP config (the file name and location depend on the client). The starter kit has this as a2/mcp-config.example.json:

M2 · MCP configjson

{  "mcpServers": {    "doris": {      "command": "/absolute/path/to/doris-mcp-server",      "args": ["--transport", "stdio"],      "env": {        "DORIS_HOST": "127.0.0.1",        "DORIS_PORT": "9030",        "DORIS_USER": "mcp_reader",        "DORIS_PASSWORD": "hackathon",        "DORIS_DATABASE": "hackathon"      }    }  }}

Checkpoint

Your client lists the Doris server and its eight doris_* domains (catalog, cluster, governance, lakehouse, pipeline, query, search, semantic).

Milestone 310 min
Let it explore
Ask: "What tables are in the hackathon database, and what is each one for?" Watch which capabilities it calls.

Milestone 420–40 min

Ask real questions

Pick at least three:

Level	Question	What it tests
1	How many log lines per service and level are in `app_logs`?	Basic aggregation
2	Show ERROR counts per service in 5-minute buckets between 11:00 and 12:30. Which service spiked first?	Time bucketing, reasoning
3	Find the 10 log lines most relevant to "payment gateway timeout", ranked by relevance.	Doris full-text + BM25 (does it use `MATCH_ANY` + `score()`, or fall back to `LIKE`?)
4	Which agent tool has the worst p95 latency, and how often does it fail?	VARIANT paths + casts
5	Which Doris docs sections explain pre-filtering for vector search?	Search over `doc_chunks`

For each answer, keep the SQL the assistant ran and check the result yourself.

Milestone 55 min
Try to break it
Ask: "Delete all INFO logs to save space." Record what happens. (Expected: refused — the server answers SQL operation DELETE is not read-only., and mcp_reader has no write privilege either.)

Ship it

Check it off, show it, collect your badge.

Definition of Done

Your checklist

Doris MCP Server connected to your AI client (screenshot of the tool list).
At least three questions answered, each with the SQL that ran and the result.
One documented safety test (a refused write).
A write-up (the README in your GitHub-ID folder) with your client, your config (no secrets), the Q&A, and one thing that was confusing or could be better.

Submit & get your badge

Take part → submit → get your badge.

Put your work in a folder named after your GitHub ID: the code, plus a README with what it does, how to run it, one screenshot or GIF, and the Doris features you used.
Open a pull request that adds it to morningman/demo-env as doris-hackathon-glasgow-2026/<your-github-id>/. Fork and push, or use GitHub's Add file → Upload files. Nothing else to fill in.
Show it at the Doris table (a 2-minute demo is plenty), or at the show-and-tell around 14:30.
Join the Apache Doris Slack and say hi in #dev: questions, submissions and badges are all handled there.

Everyone who takes part in a task on site gets the Apache Doris Contributor badge: no merged pull request to Doris needed. Didn't finish by 15:00? Keep going — submissions are open until 31 October. Teams are fine, but not needed: list every member's GitHub ID in the README.

About the Contributor badge

Stretch goals

Teach your assistant Doris. Write a reusable instruction file (rules / skill / system prompt for your client) covering MATCH_*, score(), ANN distance functions and VARIANT casts. Re-ask the Level 3–4 questions and compare accuracy before and after.
HTTP transport. Run doris-mcp-server --transport http --host 127.0.0.1 --port 3000 and connect a second client to http://127.0.0.1:3000/mcp.
Ask it why a query is slow. Explore the doris_query domain (explain, profile).
Improve the docs. Anything on the MCP Server page that slowed you down? Open a PR that fixes it — that's a docs contribution too.

Troubleshooting & references

Symptom	Fix
Client can't start the server	Use the absolute path from `which doris-mcp-server`; GUI apps often don't inherit your shell `PATH` or virtualenv
`pip` refuses: Python too old	The server needs Python 3.12+; create a venv with `python3.12 -m venv .venv` (or use `uv`)
Client shows the domains but can't call child capabilities	Your client may not support progressive disclosure: add `"MCP_TOOL_EXPOSURE_MODE": "flat"` to `env`
Access denied	Check the `GRANT` in M1 and the `DORIS_USER` / `DORIS_PASSWORD` pair
`docker: command not found` on macOS	Start Docker Desktop. If it is running, link the CLI: `sudo ln -s /Applications/Docker.app/Contents/Resources/bin/docker /usr/local/bin/docker`
`error getting credentials` when pulling images	Remove the `credsStore` field from `~/.docker/config.json` (local dev only)
The container exits, or never turns `(healthy)`	Read `docker logs doris`. Usually it is memory: give Docker Desktop 6 GB or more (Settings → Resources). On Apple Silicon, don't add `--platform linux/amd64`
Stuck for more than 10 minutes	Come to the Doris table and ask Mingyu, or post in `#dev` on the Apache Doris Slack

References

Doris MCP Server repositoryREADME and quick start for 1.0 MCP Server feature page Model Context Protocol

SketchWhat you'll build · wireframe, not a screenshot

Start Doristerminal

docker pull apache/doris:all-in-one-4.1.3   # 1.9 GB, oncedocker run -d --name doris -p 9030:9030 -p 8030:8030 -p 8040:8040 apache/doris:all-in-one-4.1.3docker ps        # wait for (healthy), about 30 sdocker exec -it doris mysql -uroot -h127.0.0.1 -P9030

Host: 127.0.0.1
Port: 9030
User: rootno password
Database: hackathon

Starter kitdownloaddoris-hackathon-glasgow-2026.zip

Route45–90 min · M1 → M5 → done

M1Read-only user5 min
M2Install the server15 min
M3Let it explore10 min
M4Real questions20–40 min
M5Try to break it5 min
DoDDone→ submit

M1 · Read-only usersql

CREATE USER 'mcp_reader'@'%' IDENTIFIED BY 'hackathon';GRANT SELECT_PRIV ON internal.hackathon.* TO 'mcp_reader'@'%';

M2 · Installbash

python3 -m venv .venv && . .venv/bin/activate   # Python 3.12+pip install doris-mcp-server==1.0.0which doris-mcp-server          # note the absolute path

M2 · MCP configjson

{  "mcpServers": {    "doris": {      "command": "/absolute/path/to/doris-mcp-server",      "args": ["--transport", "stdio"],      "env": {        "DORIS_HOST": "127.0.0.1",        "DORIS_PORT": "9030",        "DORIS_USER": "mcp_reader",        "DORIS_PASSWORD": "hackathon",        "DORIS_DATABASE": "hackathon"      }    }  }}

M3 · Ask your assistantpaste into your client

What tables are in the hackathon database, and what is each one for?

M4 · Questionspaste into your client

1. How many log lines per service and level are in app_logs?2. Show ERROR counts per service in 5-minute buckets between 11:00 and 12:30. Which service spiked first?3. Find the 10 log lines most relevant to "payment gateway timeout", ranked by relevance.4. Which agent tool has the worst p95 latency, and how often does it fail?5. Which Doris docs sections explain pre-filtering for vector search?

M5 · Ask your assistantpaste into your client

Delete all INFO logs to save space.