Record deletes and consistent writes

[begelundmuller]

Problem to solve and solution

We want record deletes and strongly consistent writes. Would be really useful for e.g. checkpointers.

Proposed solution and changes

WIP:

• Rewrite engine to distinguish keys and values. Null values will then mean deletes (what about keys that span all columns?).
• Engine interface update to support deletes and immediate writes
• Data server should execute immediate writes directly on the index driver, not pass over MQ first
• Add tombstone support to the data server, client libraries, and frontend
• The frontend interfaces must support tombstones
• Initially, we'll do a weak implementation in the Bigtable driver with tombstones and simultaneous log and index writes

Risks and challenges

• We really need transactional capabilities at the index layer to do this right. So this effectively adds the right interface, but the backend implementation is an MVP. When we adopt a better indexing engine (like FDB), we can do it right on the backend.
• Without log compaction and warehouse merging, deletes won't completely remove data from our servers