Questions about hot storage

Hi everyone,

Recently I have read " How Rockset Separates Compute and Storage Using RocksDB"(https://rockset.com/blog/separate-compute-storage-rocksdb/) and have a question about hot stoarge.
In this article, compute node retrives block from hot storage node. Does it act like a secondarycache in rocksdb? Compute node already has a PSC which caches data in local SSD, did these secondarycaches cascade?

Yes, from a data movement perspective hot storage acts like a read-through persistent cache. It is shared across all rocksdb instances (primary vs. secondary instances used to increase availability, as well as leader vs. follower VIs).

Hot storage blurs the lines between cache and storage because we use application level knowledge and cache migration strategies to get the cache miss rate to an extremely low level.

When compute nodes have a secondary cache then hot storage is an additional chained caching layer, yes. Note that the I/O bandwidth that is available to the disk on a compute node makes it difficult to use the PSC effectively – we can neither read it nor write it as fast as we would like. The long term trajectory of our architecture is toward a design with no PSC.

:ok_hand:t2: thanks