Skip to content

Latest commit

 

History

History
110 lines (81 loc) · 3.55 KB

README_other.md

File metadata and controls

110 lines (81 loc) · 3.55 KB
Raw

mlflow-tools - Other Tools

Some other useful tools for MLflow.

MLflow Spark UDF Workaound

  • Problem
    • Currently you cannot load a SparkML model as a UDF with MLflow due to named column bug.
    • Error message: pyspark.sql.utils.IllegalArgumentException: sepal_length does not exist. Available: 0, 1, 2, 3
    • mlflow git issue 4131 - Spark UDF throws IllegalArgumentException
  • Solution

Failed Run Replayer

Failed Run Replayer - Save run details for MLflow rate limited exceptions and replay later.

Seldon MLflow MLServer

Enable Seldon MLServer to support MLflow model URIs with the models and runs scheme.

Typically production MLflow pipelines deploy models using the models URI scheme which accesses the MLflow Model Registry.

Seldon provides MLServer that supports an MLflow runtime. However MLServer only accepts file-based MLflow model URI schems such as s3, wasb, or local file path. MLServer does not support the models or runs schemes which download the requested model using the MLflowClient.

Setup

python -m venv mlserver
source mlserver/bin/activate

pip install mlserver
pip install mlserver-mlflow
pip install sklearn

git clone https://github.com/amesar/mlflow-tools
pip install -e mlflow-tools

Source Code

Create custom model-settings file

Create your settings file with an MLflow model URI using the models scheme.

For example, my-model-settings.json file.

{
    "name": "wine-classifier",
    "implementation": "mlserver_mlflow.MLflowRuntime",
    "parameters": {
        "uri": "models:/sklearn_wine/production"

    }
}

Download model and create model-settings.json

The mlflow_tools.seldon_mlflow.download_model tool does the following:

  • Downloads the model artifacts associated with the registered model to a temporary location.
  • Creates a standard model-settings.json file from your custom settings file that points to this temporary location.
python -u -m mlflow_tools.seldon_mlflow.download_model \
  --model-settings-path my-model-settings.json \
  --output-dir /tmp/my-model

Generated model-setting.json file

{
    "name": "wine-classifier",
    "implementation": "mlserver_mlflow.MLflowRuntime",
    "parameters": {
        "uri": "/tmp/my-model/sklearn-model"
    }
}

Run MLServer

mlserver start .

Score

Score with the MLflow model server JSON format (in another terminal).

curl  http://localhost:8080/invocations  \
  -H "Content-Type:application/json" \
  -d '{ "columns":   
           [ "alcohol", "chlorides", "citric acid", "density", "fixed acidity", "free sulfur dioxide", "pH", "residual sugar", "sulphates", "total sulfur dioxide", "volatile acidity" ],
        "data": [
           [ 7,   0.27, 0.36, 20.7, 0.045, 45, 170, 1.001,  3,    0.45,  8.8 ],
           [ 6.3, 0.3,  0.34,  1.6, 0.049, 14, 132, 0.994,  3.3,  0.49,  9.5 ] ] }'

[5.335031847133758, 4.5]