Financial Services & Fintech · Greenlight's Rolling Fingerprint Refactor: Isolating an Unmonitored Six-Figure Cost
Greenlight's Rolling Fingerprint Refactor: Isolating an Unmonitored Six-Figure Cost
Reducing hidden infrastructure spend by tracing usage down to the write pattern
The Client · Greenlight
Overview
Greenlight, the category leader in family fintech — offering a debit card, investing, savings, safety, and senior-protection products to 6.5 million parents and kids with more than $2 billion in family-managed assets — came to Taller to expand engineering capacity across its mobile and backend stack. This engagement surfaced a previously unmonitored infrastructure cost and resolved it through targeted refactoring.
The Problem
A core function in the feature-management service (the system that controls which app features each customer sees and how new ones are rolled out) kept what engineers called a "rolling fingerprint" for every user-and-feature pairing. In plain terms, it stored a small, constantly updated record of the current state of each feature for each user, refreshing that record whenever something changed. Those records lived in DynamoDB, Amazon’s cloud database, which bills according to how often data is written to it. Because the system kept one record per user per feature and updated them continuously, the writes added up fast (up to 1,500 per second at peak), quietly driving costs of more than $200K per year. The expense went unflagged until it surfaced internally, and even then the team had no clear picture of what the table was costing or whether other teams actually needed data tracked at that fine a level of detail.
The Solution
The engagement began with a close investigation of the original implementation: tracing how the rolling-fingerprint logic worked, adding instrumentation to measure the actual volume of writes, and identifying where writes could be cut without affecting how the feature behaved. The team then refactored the function, rewriting its internal logic to run more efficiently while leaving its behavior unchanged, to reduce the number of database writes needed to maintain the fingerprint. The work also scoped a possible next phase: storing a single set of attributes per user rather than one per user-and-feature pairing, which would cut writes further, contingent on first confirming whether any teams actually relied on the finer, per-feature detail.
The Impact
The refactor reduced unnecessary database writes, cutting infrastructure costs by more than $200K per year on a previously unmonitored expense. New usage measurement gave the team a clear view of write patterns where none had existed before, and thorough QA confirmed the savings came with no loss of functionality.
