We're seeing read latency spikes (very high read latency p95, up to the ~300 ms range) in our Scylla backend.

These seem to be correlated to blob read operations like contains_blob.

We suspect that what's happening is that:

• blob queries are relatively slow, because blobs are big
• all blobs are assigned the same root_key, which we use as a Scylla partition key
• being on the same partition means that all blob queries are handled on one CPU, where they block all other queries to that CPU
• this results in the queries at the tail of the query queue being serviced very late

As a solution, we suggest setting the partition key of blobs to be just the blob ID.

This is a different behaviour from that expected by linera-views (which correlates things under the same root_key, e.g. chain ID, for fast multi-key queries on related objects), so we'd like to pull out the ‘key/value store abstraction’ layer (linera_views::backend) from the ‘persistent views’ layer of linera-views and use that directly for blobs.