Skip to main content

Model Lifecycle

Understanding how models are created, trained, and deployed in Kanva.

Lifecycle Stages​

┌─────────┐    ┌──────────┐    ┌──────────┐    ┌──────────┐
│  Setup  │ →  │ Training │ →  │  Review  │ →  │ Deployed │
└─────────┘    └──────────┘    └──────────┘    └──────────┘

1. Setup​

In the setup stage, you:

  • Upload or connect your dataset
  • Define input features and target variable
  • Configure preprocessing options
  • Set training parameters

2. Training​

During training:

  • Kanva analyzes your data
  • Multiple model architectures are evaluated
  • Hyperparameters are optimized
  • Cross-validation ensures robust results

Training time depends on:

  • Dataset size
  • Number of features
  • Model complexity
  • Selected algorithms

3. Review​

After training completes:

  • Review model performance metrics
  • Examine feature importance
  • Test predictions on sample data
  • Compare different model versions

4. Deployed​

Once deployed:

  • Model is available via the API
  • Predictions can be made in real-time
  • Performance is monitored continuously

Model Versions​

Kanva maintains version history for your models:

VersionStatusCreatedNotes
v3Active2024-01-15Current production
v2Archived2024-01-10Previous version
v1Archived2024-01-05Initial training

Promoting Versions​

To change which version is active:

  1. Go to your project's Models tab
  2. Select the version you want to activate
  3. Click Promote to Active

The API automatically uses the active version.

Version Rollback​

If issues arise with a new model:

  1. Navigate to the previous version
  2. Click Promote to Active
  3. The previous model immediately starts serving predictions

Retraining​

Models should be retrained when:

  • New training data is available
  • Data distribution has changed (concept drift)
  • Performance metrics decline
  • Business requirements change

To retrain:

  1. Update your dataset with new data
  2. Go to Training tab
  3. Click Start New Training
  4. Review and promote the new version

API and Model Versions​

The prediction API always uses the active model version. You don't need to specify versions in API calls:

POST /api/v1/projects/{id}/predict

This design ensures:

  • Seamless updates without client changes
  • Easy rollback if issues arise
  • Consistent API interface

Monitoring​

Track your deployed model's health:

  • Prediction Volume: Requests over time
  • Latency: Response time percentiles
  • Error Rate: Failed predictions
  • Input Distribution: Feature value shifts

Set up alerts for anomalies to catch issues early.