Idea: AtlasDBQueryFilter (probably in Cassandra 3)

[j-baker]

Most relevant classes: CollationController, SSTableNamesIterator.

At present, AtlasDB uses the SliceQueryFilter for its queries. This has three downsides.

1. In multicolumn range scans, we must read all historic versions.
2. Cassandra has an optimization for the names query filter - it says - sort your SSTables by max timestamp descending (provided that they do not overlap) - if you see the relevant column in the latest SSTable, you can skip looking at the others.
3. We can only select one column per RPC - if an internal use case has 5 columns, we must issue 5 RPCs to Cassandra in order to load these.

It would be relatively straightforward for us to define an AtlasDBQueryFilter which specializes Cassandra's behaviour for AtlasDB tables specifically - locating the latest write for the cell only instead of having to look at all SSTables and merge.