We Started Putting Unpopular Assets in Memory

Summary: They used to cache things immediately onto SSDs. This causes lots of wear because you end up writing lots of data that it turns out you didn't want to. To fix this, they first stage things in memory. If an item is popular enough and worth caching, they store it onto disk (hence reducing the number of writes on the SSD). Both with the old and new approaches, they rely on the OS page cache to keep hot items in memory.

Great article! The author (Yuchen Wu) is a master communicator.

I don't think this is an uncommon strategy. It's used in other caching scenarios where data is put into the fastest tier, then expire or age out to slower tiers as necessary.

Typical tiered memory hierarchies put more popular assets in faster tiers of memory. This post describes the opposite: putting less popular assets in faster tiers of memory to improve performance of the system.

I'm talking about the initial location (which is what this article covers).

Of course the most popular content stays in faster tiers, but the difference is placing it there first and aging it out instead of promoting it after from slower tiers.