Ratekeeper uses an inaccurate lower bound for the released transaction rate

Ratekeeper computes a lower bound for the number of transactions being released as the total rate of bytes being made durable across all storage and log servers divided by the max transaction size (10MB). For various reasons (e.g. replication, queue overhead, duplication of log and storage), this is not accurate and under some workloads it can lead to a significant overestimate for the number of transactions taking place. As a result, ratekeeper may allow resources to run away inadvertently.

The issue names Ratekeeper but no source file, test, or entry point. Start by locating its released-transaction-rate calculation and review how durable bytes, replication, queue overhead, and duplicated log/storage traffic affect it. Done requires a decided replacement lower bound and coverage for workloads where the current estimate overstates transactions.

Written by the indexing model from the issue text.