An ecommerce site hosts a TensorFlow product recommendation model on a SageMaker endpoint using three compute-optimized instances to handle expected peak traffic. At the start of each month response times spike and some users see errors. Most traffic occurs weekdays from 8 AM to 6 PM in a single time zone. Which two actions most effectively fix the problem while minimizing cost? (Choose two.)

Enabling Amazon Elastic Inference (EI) accelerators for the endpoint is effective because EI can significantly reduce inference costs and latency for deep learning models by attaching GPU-powered acceleration to CPU instances. This helps handle the increased computational demands during peak times without requiring larger or more numerous GPU instances. Configuring the endpoint to autoscale based on the InvocationsPerInstance metric is crucial because it allows the endpoint to dynamically adjust the number of instances based on actual traffic load. This prevents performance degradation and errors during peak periods (like the start of the month) while minimizing cost during off-peak times by scaling down. Creating a new endpoint configuration with two production variants doesn't directly address performance or cost issues; it's more for A/B testing or deploying different model versions. Deploying a second instance pool for blue/green deployment is a deployment strategy, not a performance or cost optimization for an existing issue. Reconfiguring the endpoint to use burstable instance types (like T-series) is generally not suitable for sustained, high-performance machine learning inference workloads due to their limited burst credits and potential for throttling.