BYOM Handler

Http api:

POST api/handlers/byom/<engine_name>

Uploaded files

• code - python file with model code
• modules - text file with requirements

Logic:

1. Create new ml engine
2. If virtualenv is installed on computer:
   • try to create a new virtual environment using current python interpreter
   • install all requirements for model
     • if pandas is not in requirements - it is installed anyway
   • install pyarrow (for sending dataframe between different pandas versions)
3. If not virtualenv is installed:
   • try to install requirements to current environment
4. Try to import uploaded model and check predict and train methods
5. If not success:
   • remove virtual environment if it was created
   • remove ml engine

Using:

1. Add args argument to train and predict methods:

class CustomPredictor():
    def train(self, df, target_col, args=None):
       ...
       
    def predict(self, df, args=None):
       ...

2. Passing args to model training

create predictor pred 
from files (select * from byom)
predict Time
using engine='uploaded_model',
  param1=1, param2='2'

3. Passing args at predict

select * from files.byom t 
join pred3 p
using param1=1, param2='2'