Uber manages a massive infrastructure with thousands of microservices, catering to over 170 million users across various platforms like rides and food delivery. At its core are two proprietary databases, Docstore and Schemaless, built on MySQL, which handle an enormous volume of data—tens of petabytes and millions of requests per second. Given the scale, even minor overloads can lead to significant cascading failures, making effective overload management essential. Initially, Uber experimented with a quota-based rate-limiting system to control traffic in the query engine. This approach involved tracking usage through a central Redis cache, but it quickly revealed flaws. The overhead introduced by additional network hops and the inability to accurately measure request costs made the system inefficient. Static quotas proved ineffective in a multitenant environment, leading to frequent adjustments. The key insight from this failure was the need to manage overload closer to the storage nodes. To create a more resilient system, Uber shifted to a new load management strategy that focused on real-time signals of system load. They emphasized concurrency as a primary indicator of overload, which aligns closely with resource usage. The new design includes mechanisms like Controlled Delay (CoDel) for smarter queuing, which adapts based on request wait times rather than just queue lengths. This allows for prioritizing newer requests under stress. A Scorecard engine also helps enforce concurrency limits per tenant, ensuring no single user can monopolize resources, thus maintaining stability and fairness across the system.