https://github.com/getindata/streaming-ml-with-flink
Demo of running SciKit model on Flink, using Mleap serialization
https://github.com/getindata/streaming-ml-with-flink
Last synced: about 1 year ago
JSON representation
Demo of running SciKit model on Flink, using Mleap serialization
- Host: GitHub
- URL: https://github.com/getindata/streaming-ml-with-flink
- Owner: getindata
- Created: 2022-03-22T15:29:18.000Z (about 4 years ago)
- Default Branch: main
- Last Pushed: 2022-03-22T15:35:01.000Z (about 4 years ago)
- Last Synced: 2025-01-24T02:31:03.203Z (over 1 year ago)
- Language: Scala
- Homepage:
- Size: 27.3 KB
- Stars: 3
- Watchers: 7
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Flink Mleap
Example how to serve ML model in Flink using Mleap
## Getting started
Before serving ML model you need to create one and export it using MLeap.
Example how to do it https://getindata.com/blog/online-ml-model-serving-using-mleap/
In `./example//src/main/resources` you can find two random forest models:
- `mleap-example-1`
- `mleap-example-2`
Those models are taking **one double as input** and **result is one double**.
## Bundle loader
We have exported model inside files, so now we need a way to load them to Flink.
To do that, we will use: `com.getindata.mleap.BundleLoader` and its `FileBundleLoader` implementation inside `lib` module.
This allows us to load ML model stored locally (it is possible to prepare loader from any cloud storage).
We provide also `GCSBundleLoader` implementation, you can load your bundle from Cloud Storage.
## Streaming API
Firstly we will use MLeap with Streaming API in package: `com.getindata.datastream`,
It is done in `FlinkDatastreamWithMleap` app in `example` module. To present it, stream of three random numbers was created, and
it is processed using `MleapProcessFunction` (from `lib` module), which simply makes prediction. Making MLeap prediction is
done in `MleapMapFunction.map` method.
## SQL
Using MLeap with SQL API is presented in package `com.getindata.sql` inside `example` module.
App `FlinkSqlWithMleap` contains e2e example how to use it. To test it `datagen` table is created with random features.
To make prediction UDF is implemented: `MLeapUDF` (in `lib` module) and MLeap is used here in `eval` method. We make `MleapUDF` very generic.
Its input arguments and output is casted to proper types based on MLeap model schema. Basically this one UDF allows using different
models inside your job. To make registrations process of UDFs easier - we implemented MLeapUDFRegistry, which can register UDFs basing
on configuration file.
For test purposes we register two UDFs:
- Predict
- Predictv2
To show it is working, simple query is executed:
`SELECT Predict(feature1) as prediction, Predictv2(feature1) as prediction2 FROM Features`