Lakehouse: Manus? MCP? Let’s Talk About Lakehouse and AI

env DORIS_HOST=<doris-host> DORIS_PORT=<query-port> DORIS_USER=<doris-user> DORIS_PASSWORD=<doris-pwd> uv run --with mcp-doris --python 3.13 mcp-doris

Then, click “Add” to add an Apache Doris MCP Server. A small green dot indicates that the MCP server is running normally.

2. Interacting With Doris MCP Server

The currently implemented Doris MCP Server is relatively simple, only implementing the ability to list databases, tables, and execute query SQL. Now we can open Cursor’s “Chat” page and switch to “Agent” mode. 

Note that if you’re using an older version of Cursor, you’ll need to open the “Composer” page to interact with the MCP server.

Next, we can interact with the Doris database using natural language.

In the video demonstration, I asked Cursor to list the databases in Doris and analyze the contents of the TPCH database, and finally, I used natural language to ask a question about the TPCH dataset:

In TPCH database, find which supplier should be selected to place an order for a given part in a given region.

Cursor, through multiple interactions with the MCP server, ultimately provided and executed an SQL example:

SELECT 
    s.s_suppkey, 
    s.s_name,
    s.s_address,
    s.s_phone,
    p.p_partkey,
    p.p_name,
    ps.ps_supplycost,
    ps.ps_availqty,
    n.n_name AS nation,
    r.r_name AS region
FROM 
    tpch.supplier s
    JOIN tpch.nation n ON s.s_nationkey = n.n_nationkey
    JOIN tpch.region r ON n.n_regionkey = r.r_regionkey
    JOIN tpch.partsupp ps ON s.s_suppkey = ps.ps_suppkey
    JOIN tpch.part p ON ps.ps_partkey = p.p_partkey
WHERE 
    r.r_name = 'REGION_NAME' AND  -- Replace with desired region
    p.p_partkey = PART_KEY        -- Replace with desired part
ORDER BY 
    ps.ps_supplycost ASC,         -- Primary: lowest cost
    ps.ps_availqty DESC           -- Secondary: highest availability
LIMIT 1;

As we can see, through the MCP server, we quickly implemented capabilities similar to ChatBI. Additionally, we can add more MCP servers with different capabilities, and Cursor can simultaneously use these servers’ capabilities to complete complex combined tasks.

We can further enrich the capabilities of the Doris MCP Server by providing interfaces for querying system information to help us diagnose the operation issues of the Doris cluster. We can also allow the MCP server to read database, table, and field comments to better understand the meaning of data in the database and provide more accurate answers.

Lakehouse: What Is the Big Deal?

Returning to our main topic, let’s discuss the relationship between Lakehouse and AI. And why Lakehouse can become the data foundation in the AI era.

Actually, the MCP server example above doesn’t showcase any particular advantages of Lakehouse or analytical databases in this scenario. It seems we could achieve similar capabilities by replacing Doris with MySQL or other systems with data access capabilities. So what exactly are the advantages of Data Lake and data analytics engines like Doris in the AI domain?

Lakehouse in the AI Era

First is the Data Lake. In my previous article, I mentioned two core points: data sharing and diverse workloads and collaboration. It allows different compute engines to work collaboratively on shared data and integrates seamlessly with various tools and platforms. This means that during AI development, data scientists and engineers can conveniently use familiar tools to process data without complex data migration and format conversion, ensuring data consistency and real-time access, driving efficient AI workflow operation.

On the other hand, support for diverse data format storage is also an important factor that enables Data Lake to serve AI scenarios. Beyond tabular data, storage of unstructured data and vector data allows us to confidently meet the data requirements of complex scenarios like Manus, which requires multiple capabilities and data collaboration. 

Of course, the widely applied open lake formats such as Iceberg, Paimon, etc., essentially still belong to the structured and semi-structured tabular format category, which is just a part of the entire data lake storage capability. I will discuss how a unified catalog service manages all data on the lake in future series articles.

Next is the data analytics engine. For analytics engines like Apache Doris, on the surface, it seems that from the data era to the AI era, the main functionality they provide hasn’t changed significantly: data analysis through SQL language. However, in the AI era, application scenarios place higher demands on the capabilities of analytics engines:

Higher Performance

From the MCP server demonstration earlier, we can see that when we pose questions to LLM using natural language, the model breaks down the problem into multiple sub-problems to process. With DeepSeek’s popularization of deep thinking capabilities, a complex problem implies a longer chain of thought and more sub-problems. 

If readers carefully observe the model’s processing of the problem in the example, they can see that it repeatedly interacts with the MCP Server (or Doris) to eventually arrive at the solution through exploration. Behind a model solving a problem are numerous data interaction actions, even explosive request generation. 

Therefore, the response capability of the backend data source directly affects the user experience. Analytics engines like Doris, with their ultra-high performance and concurrent processing capabilities, make these interactive AI scenarios possible with an acceptable user experience.

Richer SQL Expression Capabilities

SQL language provides reliable problem expression capabilities based on relational algebra. From the development trajectory of SQL to NoSQL and back to SQL in the big data field, we can see that SQL remains one of the most important capabilities in the data analytics domain. 

On the basis of two-dimensional relational data, through UDFs, we can extend more capabilities, such as the ability to analyze unstructured data or the ability to call LLM to analyze data. The remote UDF provided by Apache Doris further decouples “server-side computing capabilities” and “client-side computing capabilities,” enabling more flexible data analysis functions.

Open Data API or Open Data Format?

In mainstream open data lakehouse architectures, data is typically presented in its openness in the form of Open Data Format. Representative projects include Iceberg, Hudi, Paimon, Delta Lake, etc. These projects usually start from a storage perspective, designing a unified format standard for upstream systems to interface with. 

On the other hand, compute engines or data warehouse systems represented by BigQuery and Redshift expose their stored data externally through Open Data API.

The following table lists the differences between these two “Open” forms: