What is the metric being measured on the chart you shared? Can you share more details of your benchmark? In line with your findings, I would expect VARBINARY or BINARY type to be best for storing UUID data.
CockroachDB and Cloud Spanner distribute data via range sharding. This means when using an optimized COMB UUID writes/reads of recent data would only hit a couple nodes. This is why random UUID’s work better since it prevents a specific node from getting overloaded. MemSQL distributes data via hash sharding and so doesn’t have this issue. As a result, from a data distribution perspective random UUID and COMB should be relatively equivalent.
From a storage perspective, I would expect COMB to be better in MemSQL since you will likely be indexing the UUID column in row store or sorting by it or a time dimension in columnstore. Since data would be arriving in a naturally sorted order, this would make it more efficient.