There are some updates and clarifications for the approach. About Multi/Single Node Operations To improve performance for single-partition operations it's possible to separate single-partitions and multi-partitions transactions. A single-partition operation sends a put/get directly to the required node. A coordinator's logic should support only multi-node transactions. For single-partitions node also keeps an active transaction list, use updateSeq as CurrentTS to find a proper version but it's not necessary to inform a client about transactions in the RUNNING state.
About Garbage Collector "Expired Version Ejection" looks more accurate name. About Repeatable Reads isolation level To avoid caching values on client node it'd be better to cache a currentTX for keys. And the second get/select will return a value for the same currentTX. About a maxTS Need to revise the approach for reducing a space consuming as much as possible. BR, Serge On Wed, 16 Aug 2017 at 18:57, Alexey Kuznetsov <akuznet...@apache.org> wrote: > Serge, > > >#1 Could you please share any case when we need to use different modes > for different > caches? > > Atomic caches? > > > > -- > Alexey Kuznetsov >