Highest Write Wins flag, register, map and set

[drewhk]

Only for discussion not finished yet

[drewhk]

Bit lacking on comments now, will add more info later.

[patriknw]

@drewhk This looks great at a first overflight. Is it described in some paper, or is it invented by Varga?

It will require protobuf serialization, i.e. with ReplicatedDataSerialization.

[patriknw]

setValue is a bit strange for an immutable data type, perhaps withValue?

[drewhk]

I was thinking about this. The problem is that a name that is passive does not convey the meaning that you should not just cache True/False values but always operate on the latest value by setting it. I don't know what name would be nice here.

[drewhk]

I don't know if this is invented or reinvented by me, but I haven't read about such a structure before. I will add the serialization formats later, I just wanted a validation first so I flushed what I had here.

[drewhk]

Btw, according to this: http://www.bailis.org/blog/linearizability-versus-serializability/
both the register and flag are serializable (but not linarizable), and the set and map are serializable per entry (but not over all entries). Interestingly ORSet as far as I see is neither. It would be interesting to investigate the differences and why one of the behaviors might be preferred over the other in different use-cases.

[patriknw]

Trying to understand the end user semantics.
In the case of concurrent updates during a partition the most active replica (or side of the partition) will win. By "most active" I mean changing the flag back and forth most number of times.
Is that correct understanding?