Grafana Mimir, an open-source and scalable time-series database, faced challenges with uneven load distribution among its ingesters due to the consistent hashing algorithm previously used, leading to some ingesters being overwhelmed while others were underutilized. To address this, a new "spread-minimizing token generation strategy" was developed, ensuring a more balanced load across ingesters by optimizing token distribution within the hash ring. This strategy significantly improved load balancing in Grafana Cloud Metrics production clusters, reducing the load distribution spread from 15%-25% to 0.5%-1% and enabling better resource utilization and cost savings. The strategy, now default for clusters without shuffle sharding, involves assigning fresh tokens in a way that minimizes the difference in token ownership among ingesters, thereby achieving a more uniform distribution of in-memory time series. While the new strategy has not yet been applied to clusters with shuffle sharding due to additional complexities, it demonstrates potential for improving load balancing when coupled with enhancements in shard creation algorithms. The implementation is available in the dskit OSS Golang library, and the strategy is recommended for users of Grafana Mimir OSS or Grafana Enterprise Metrics in similar environments.