import logging
import os

from flask import jsonify

from . import app, request
from .tasks.train_model import train_models

logging.getLogger().setLevel(logging.INFO)

DATABASE_NAME = os.getenv('DATABASE_NAME')


@app.route('/train_models', methods=['POST'])
def invoke_train_models():
    """
    Expects some feature engineering, but can do without too

    Trains multiple ML models and stores results in mlflow s3.
    Accessible through UI mlflow.fansifter.cloud.

    Returns modelling_runs dictionary with mlflow run_ids to search data from mlflow tracking server later.
    """
    data = request.json if request.json else request.data
    logging.info(f"input data: {repr(data)}")
    schema = data['schema']
    collection_ids = data['collection_ids']
    bucket = data['bucket']
    file_key = data['file_key']

    """ We used this in legacy to differentiate between "RFM modelling" and "other". When feeding input fields
    from whatever is calling model training, we no longer have a need for this.
    """
    modelling_config = {
        'plot_2d': False,
        'plot_3d': False,
        'plot_silhouette': False,
        'importance': True,
        'boxplot': False,
        'save_model': False,
        'modelling_type': 'clustering',
        # -1 means no pca dimensionality reduction is done. Multiple values mean for each of those values a model will be trained
        'pca_components': [4],  # TODO! All experiments with -1 (no pca) were inferior to PCA, but there is more room to experiment
        'drop_corr_threshold': 0.99  # Drop features that are highly correlated. That's why - https://www.trchome.com/docs/5-cluster-analysis-gets-complicated/file
    }
    modelling_runs = train_models(schema, collection_ids, bucket, file_key, modelling_config)

    return jsonify(modelling_runs), 200