Think next steps around guarantees and performance

[michal-harish]

• we already have a general at-least-once processing which a good foundation
• batching sends at the akka level should be given a though first in case it constraints this one
• exactly-once would be nice but
  1. going the route of deduplication in-memory sets is general but is a big hit on throughput and memory
  2. going the route of kafka transactions performs well but is not general, i.e:
     a) it will only work with kafka as a log storage (not a problem as long as the implementation doesn't prescribe kafka even if exactly-once is not required)
     b) kafka transactions work only on the same cluster which is acceptable
  3. is there a third way ?

In some other design work related to commits there were 2 points considered:
a) being able to do flushAll() on or produce operations
b) have a global context of all operations, i.e. their futures, which were executed