6. Runtime View

Prediction and Feedback

        sequenceDiagram
    participant User
    participant Streamlit UI
    participant FastAPI Backend
    participant Data Processor
    participant MLflow
    participant Prometheus
    participant Grafana
    participant MinIO

    Data Processor ->> MLflow: First training and versioning the model
    Data Processor ->> MinIO: Versioning training data and all steps
    User ->> Streamlit UI: Enter input data
    Streamlit UI ->> FastAPI Backend: Send request
    FastAPI Backend ->> Data Processor: Forward request
    Data Processor ->> MLflow: Fetch latest model
    MLflow -->> Data Processor: Return model
    Data Processor ->> Data Processor: Preprocess and predict
    Data Processor -->> FastAPI Backend: Return prediction
    FastAPI Backend -->> Streamlit UI: Send response
    Streamlit UI -->> User: Display prediction

    User ->> Streamlit UI: Provide feedback (correct/incorrect)
    Streamlit UI ->> FastAPI Backend: Send feedback data
    FastAPI Backend ->> Prometheus: Send feedback data to Prometheus
    Prometheus -->> Grafana: Track and visualize feedback metrics
    Grafana ->> FastAPI Backend: Feedback if threshold reached via webhook
    FastAPI Backend ->> Data Processor: Retrain model if threshold met
    Data Processor ->> MinIO: Versioning new training data and all steps
    Data Processor ->> MLflow: Log new model version
    MLflow -->> Data Processor: Save new model version

This scenario describes how a user interacts with the system via the Streamlit UI to request a machine learning prediction. The User give also a feedback, if the prediction was right or not.

Prediction Interaction

  1. The user inputs data via the Streamlit UI.

  2. The Streamlit UI sends the request to the FastAPI backend.

  3. The FastAPI backend forwards the request to the Data Processor.

  4. The Data Processor loads the latest model from MLflow.

  5. The Data Processor preprocesses the data and performs inference.

  6. The prediction is returned to the FastAPI backend.

  7. The FastAPI backend sends the prediction result back to the Streamlit UI.

  8. The user sees the predicted result.

Feedback Interaction

  1. After receiving the prediction, the user provides feedback (correct or incorrect) through the Streamlit UI.

  2. The feedback is sent to FastAPI via the backend, which forwards it to Prometheus.

  3. Prometheus collects feedback metrics (e.g., accuracy) and monitors them.

  4. Grafana visualizes the feedback and tracks the metrics over time, monitoring whether the threshold is met (for example, if the accuracy drops below a certain threshold). Grafana also send a webhook to the API, if the threshold met a specific value.

  5. Threshold Detection: If the feedback threshold is reached (for example, if a certain number of incorrect predictions occur), Prometheus alerts the system, triggering model retraining.

  6. Retraining the Model: The Data Processor retrains the model and logs the new model version in MLflow.