> it also not feasible in practice for the same reasons ermouth outlines
If we shift the scope a bit, CRDT is ok for distributed collaborative allocation/deallocation of quantitative resources or commutative numerical entities of other kinds (ie likes or even spectral vectors). Storing *changes* of tagged numbers or vectors instead of abs values, I mean smth like +1, or [+0.2, -0.5, +0.1, ...], is more or less ok, esp if we aggregate changes received from a single source (ie user or sensor) into a single doc. The reducer for this case is obvious – built-in _sum for instance. Garbage collection is also more or less obvious – collapsing timeseries of a single doc into final result time to time, and then compacting a bucket. However, this solution is already well known and only deserves to be named CRDT for hype or marketing. ermouth
