There are data-sets which rarely get updated so we use low cost Aurora Serverless to get occasional updates with 5 minute interval costs for inserts. Then when a user needs to use ad-hoc queries we make the data super-hot by loading those in a cluster of VRAM (HBM2 memory) for high-end GPU analytics using GPU analytics tools (such as OmniSci/MapD/Kinetica etc.).
How I built it
We parition/shard data from Aurora Storage and load each shard into a different HBM2 on a machine with https://omnisci.com installation.
See details at:
What's next for IcyHot: Serverless In-mem & GPU analytics from cold DB