How it compares to VoltDB, which used similar approach and already "combines the proven power of relational processing with blazing speed, linear scalability and uncompromising fault tolerance" (according to VoldDB website) ?
There are few differences with VoltDB.
1. It's much easier to use (MySQL compatibility, ad-hoc sql no java dependency)
2. It doesn't have any issues with data skews. We are using lockfree skiplists vs b-trees partitioned by core.
3. I believe we are faster on a single box, however this claim should obviously be verified by a third party.
Congrats to the MemSQL team. It's always validating to see more entrants into the space. The comparison between the two systems is fascinating; the approach taking is really different than VoltDB, but the systems share some similar choices as well.
As for your comments:
1. Ad-Hoc SQL in VoltDB is a current area of focus in VoltDB development. We use similar compilation and expect them to be no slower than our stored procedures in the 2012 timeframe. Starting with our release next week, we'll also support multiple SQL statements in a serialized/ACID ad-hoc transaction.
2. On a single-node (or a several replicated nodes), VoltDB doesn't really have many issues with data skews. Rebalancing a distributed system is also in development, but hasn't turned out to be much of an issue with our production customers.
3. Faster at what? Single statement non-durable transactions? That's probably true for simple SQL operations for the time being. It gets murkier once you consider VoltDB features like materialized views that avoid table-scans altogether. I am also curious how durability affects performance though.
But I'm really interested to revisit this comparison once MemSQL supports real transactions and replication/sharding. For the time being, most of the VoltDB use cases we target won't run on it.
