Groww Engineering explains their load shedding implementation using Netflix's Concurrency Limits library. They discovered their configuration was too conservative, causing premature traffic rejection and slow recovery. The system monitors response time (RTT) to dynamically adjust concurrent request limits, comparing current performance against long-term baselines. They detail how the algorithm detects overload, calculates gradual adjustments using tolerance thresholds, and smooths limit changes to maintain stability while allowing faster recovery when conditions improve.
Sort: