He has some great points about SIGMOD/VLDB reviewing - they emphasize far too much on "algorithmic novelty", instead of impactful simple systems. However, in terms of graph processing, I too think in many instances, single machine processing is too limiting.

In general, I think people would be more convinced if there were real cases in industry where engineers switched from distributing graph algos to single-machine algos because they found performance benefits were better. In all my years, I haven't heard of a case like this (it's always been the other way around).