Defines two lambdas, get-bandit-tests and query-lambda, that form a state machine (using Step Functions) to record view and acquisition data for multi-armed bandit experiments. RRCP channel tests (Epics + Banners) that are marked as bandit experiments will be picked up and their performance data written to a DynamoDB table. The state machine is scheduled to run once an hour.
Diagram showing how this state machine (shown here as "bandit sampling") fits into the existing architecture.
The multi-armed bandit decision-making happens in SDC. (Original here)
The first lambda in the state machine is get-bandit-tests. It queries the DynamoDB table support-admin-console-channel-tests- for Epic and Banner tests that have a multi-armed bandit methodology in their methodologies field. It returns a list of these tests as its output.
The next lambda, query-lambda, takes the tests from get-bandit-tests as its input. It queries the BigQuery tables fact_page_view_anonymised and fact_acquisition_event to get the views and acquisitions in GBP per variant per test. It uses the hour prior to the lambda execution as its date range.
It writes this data to a DynamoDB table support-bandit-. The data recorded consists of a test name (prefixed with the channel), timestamp and an array of performance data per variant.
Diagram showing the architecture of this state machine.