Cool! Traditionally ML datasets tend to compress and dedupe very well so depending on the budget I would probably look at an appliance with a software stack compatible that performs this extremely well then offload to object store as you scale out.

What you are looking for is a scalable appliance and I would look to building out a requirements document first covering the basic questions such as, speed, capacity, data delta(growth over time), redundancy and uptime.

Once you delve deep into these questions you’ll be asking the right questions of how and what in relation to data flow. It will then build out baseline requirements for the technology stack you require.

When I’m scoping solutions, the destination hardware is always the last question answered as if you have it as the first question, the solution is doomed from the start.

There is no “cheap” way to get petabyte level of storage. What you will spend on hdds without dedupe and compression would cover the cost of an appliance for dedupe and compression. So a mixture between the two is probably the best approach if the growth rate of the data can be pre-conditioned by a dedupe appliance before offloaded to object storage.