Why We Started Putting Unpopular Assets in Memory

What this post added

Introduced a hybrid memory-SSD storage system that prioritizes placing unpopular assets in memory first before promoting them to SSDs. This strategy aims to reduce SSD wear and improve cache hit tail latency by minimizing writes to SSDs and leveraging memory for assets that are accessed infrequently or only once. The system is deployed on newer generation servers and a subset of assets, with careful consideration of memory footprint, competition with the page cache, and competition with process memory usage.