Updated
Updated · Google Research · Jun 25
Google Unveils Elastic Caching, Cutting Spanner I/O Costs to 0.5% as Memory Runs Up to $3 per GiB
Updated
Updated · Google Research · Jun 25

Google Unveils Elastic Caching, Cutting Spanner I/O Costs to 0.5% as Memory Runs Up to $3 per GiB

2 articles · Updated · Google Research · Jun 25

Summary

  • Google Cloud and Google Research said their new linear elastic caching system dynamically changes cache size and reduced total cache ownership costs in Spanner tests while keeping added I/O costs to just 0.5%.
  • The approach treats page eviction as a ski-rental problem, assigning each page a time-to-live with a lightweight decision-tree model and falling back to LRU only when the cache physically fills.
  • Google said the method targets a cloud economics problem in which fixed-size caches are either too small for performance or overprovisioned for peak demand, with some serverless providers charging up to $3 a day for 1 GiB of memory.
  • Public-trace evaluations also showed elastic caching consistently beat fixed-size baselines, with savings growing as memory costs rise relative to cache-miss costs.
  • Presented at CIDR 2025 and now described as part of Spanner production deployments, the work points to more granular, pay-as-you-go infrastructure using cost-aware caching instead of static peak provisioning.

Insights

As AI models demand more memory, can Google's new caching logic prevent soaring cloud costs?
Google slashed its cache costs by 40%. When will other cloud providers offer similar savings?
Google's new cache trades hits for dollars. What hidden performance risks does this strategy introduce?