​
Prerequisites

​
1. Install the UDF framework

RisingWave uses the arrow-udf as its remote UDF framework. The framework provides a Python SDK for defining and running UDFs outside of the RisingWave process. Run the following command to install arrow-udf: 
pip install arrow-udf
The minimum version of RisingWave that supports arrow-udf Python UDFs is 1.10. If you are using an older version of RisingWave, please refer to the historical version of the documentation. If you have used an older version of the RisingWave UDF SDK (risingwave 0.1), we strongly encourage you to upgrade to the latest version. You can refer to the migration guide for instructions.

​
2. Define functions in a Python file

To define UDFs in Python, you need to create a Python file and define your functions using the udf (for scalar functions) and udtf (for set-returning/table functions) decorators provided by the arrow-udf module. As an example, let’s define some simple UDFs in a Python file named udf.py:
udf.py
from arrow_udf import udf, udtf, UdfServer

# Define a scalar function that returns a single value
@udf(input_types=["INT", "INT"], result_type="INT")
def gcd(x: int, y: int) -> int:
    while y != 0:
        (x, y) = (y, x % y)
    return x

# Define a scalar function to perform some blocking operation, setting the `io_threads` parameter to run multiple function calls concurrently in a thread pool
@udf(input_types=["INT"], result_type="INT", io_threads=32)
def blocking(x):
    time.sleep(0.01)
    return x

# Define a scalar function that returns multiple values within a struct
@udf(input_types=['VARCHAR'], result_type='STRUCT<key: VARCHAR, value: VARCHAR>')
def key_value(pair: str):
    key, value = pair.split('=')
    return {'key': key, 'value': value}

# Define a table function over a Python generator function
@udtf(input_types='INT', result_types='INT')
def series(n):
    for i in range(n):
        yield i

# Define a function that accepts batch input and returns a batch output
@udf(input_types=["VARCHAR"], result_type="REAL[]", batch=True)
def text_embedding(texts: List[str]) -> List[List[float]]:
    from openai import OpenAI
    client = OpenAI("<your-api-key>")

    embeddings = [
        e.embedding
        for e in client.embeddings.create(
            model="text-embedding-ada-002",
            input=texts,
            encoding_format="float",
        ).data
    ]
    return embeddings

if __name__ == '__main__':
    # Create a UDF server and register the functions
    server = UdfServer(location="0.0.0.0:8815") # You can use any available port in your system. Here we use port 8815.
    server.add_function(gcd)
    server.add_function(blocking)
    server.add_function(key_value)
    server.add_function(series)
    server.add_function(text_embedding)
    # Start the UDF server
    server.serve()
For more examples of UDFs, such as functions handling complex data types like JSONB, see this test file in RisingWave source code.

​
3. Start the UDF server

Simply run the Python file to start the UDF server. 
python3 udf.py
The UDF server will start serving requests, allowing you to call the defined UDFs from RisingWave.

​
4. Declare external functions in RisingWave

In RisingWave, use the CREATE FUNCTION command to declare the functions you defined. Here are the SQL statements for declaring the functions defined in step 2. 
CREATE FUNCTION gcd(int, int) RETURNS int
AS gcd USING LINK 'http://localhost:8815';

CREATE FUNCTION blocking(int) RETURNS int
AS blocking USING LINK 'http://localhost:8815';

CREATE FUNCTION key_value(varchar) RETURNS struct<key varchar, value varchar> -- the field names must exactly match the ones in Python decorator
AS key_value USING LINK 'http://localhost:8815';

CREATE FUNCTION series(int) RETURNS TABLE (x int)
AS series USING LINK 'http://localhost:8815';

CREATE FUNCTION text_embedding(varchar) RETURNS real[]
AS text_embedding USING LINK 'http://localhost:8815';
The function signature in the CREATE FUNCTION statement must match the signature defined in the Python function decorator. The field names in the STRUCT type must exactly match the ones defined in the Python decorator. If you are running RisingWave using Docker, you may need to replace the host localhost with host.docker.internal in the USING LINK clause.

​
5. Use the functions in RisingWave

Once the UDFs are created in RisingWave, you can use them in SQL queries just like any built-in functions. For example: 
SELECT gcd(25, 15);
---
5

SELECT blocking(2);
---
2

SELECT key_value('a=b');
---
(a,b)

SELECT * FROM series(5);
---
0
1
2
3
4

SELECT text_embedding('Hello, RisingWave UDF!');
---
 {-0.009116887,-0.03780581,-0.014567504,0.001315606,...}

​
6. Scale the UDF Server

Due to the limitations of the Python interpreter’s Global Interpreter Lock (GIL), the UDF server can only utilize a single CPU core when processing requests. If you find that the throughput of the UDF server is insufficient, consider scaling out the UDF server.
How to determine if the UDF server needs scaling?You can use tools like top to monitor the CPU usage of the UDF server. If the CPU usage is close to 100%, it indicates that the CPU resources of the UDF server are insufficient, and scaling is necessary.
To scale the UDF server, you can launch multiple UDF servers on different ports and use a load balancer to distribute requests among these servers. The specific code is as follows:
udf.py
from multiprocessing import Pool

def start_server(port: int):
    """Start a UDF server listening on the specified port."""
    server = UdfServer(location=f"localhost:{port}")
    # add functions ...
    server.serve()

if __name__ == "__main__":
    """Start multiple servers listening on different ports."""
    n = 4
    with Pool(n) as p:
        p.map(start_server, range(8816, 8816 + n))
Then, you can start a load balancer, such as Nginx. It listens on port 8815 and forwards requests to UDF servers on ports 8816-8819.

​
Data Types

The RisingWave Python UDF SDK supports the following data types:
SQL TypePython TypeNotes
BOOLEANbool
SMALLINTint
INTint
BIGINTint
REALfloat
DOUBLE PRECISIONfloat
DECIMALdecimal.Decimal
DATEdatetime.date
TIMEdatetime.time
TIMESTAMPdatetime.datetime
INTERVALpyarrow.MonthDayNanoFields can be obtained by months(), days() and nanoseconds() from MonthDayNano
VARCHARstr
BYTEAbytes
JSONBAnyParsed / Serialized by json.loads / json.dumps
T[]List[T]
STRUCT<>Dict[str, Any]
…othersNot supported yet.

​
Migration Guide from risingwave 0.1 to arrow-udf 0.2

If you have used the Python UDF SDK in RisingWave 1.9 or earlier versions, please refer to the following steps for upgrading. Import the arrow_udf package instead of risingwave.udf. 
pip install arrow-udf

- from risingwave.udf import udf, udtf, UdfServer
+ from arrow_udf import udf, udtf, UdfServer
The type aliases FLOAT4 and FLOAT8 are removed and replaced by REAL and DOUBLE PRECISION. 
- @udf(input_types=['FLOAT4', 'FLOAT8'], result_type='INT')
+ @udf(input_types=['REAL', 'DOUBLE PRECISION'], result_type='INT')
The STRUCT type now requires field names. The field names must exactly match the ones defined in CREATE FUNCTION. The function that returns a struct type now returns a dictionary instead of a tuple. The field names of the dictionary must match the definition in the signature. 
- @udf(input_types=['VARCHAR'], result_type='STRUCT<VARCHAR, VARCHAR>')
+ @udf(input_types=['VARCHAR'], result_type='STRUCT<key: VARCHAR, value: VARCHAR>')
  def key_value(pair: str):
      key, value = pair.split('=')
-     return (key, value)
+     return {'key': key, 'value': value}