NotesSystem

VoltDB

Also: Volt Active Data · H-Store

An in-memory SQL database, from the H-Store research prototype, that executes transactions serially: one single-threaded engine per partition running deterministic stored procedures, with no locks or latches.

Also in: Serializability › In the real world

Design choices

  • Actual Serial Execution: tables are partitioned; each partition's engine runs one transaction at a time, in order.
  • Stored procedures (Java or SQL): transactions are submitted whole; they must be deterministic so replicas re-executing them stay identical.
  • Synchronous replication within a partition, with snapshots and a command log for durability.
  • Multi-partition transactions are ordered by one cluster-wide initiator and coordinate every partition they touch.
  • No lock manager, latches or buffer pool: a partition's data is only ever touched by its own thread, and all of it lives in memory. The product is now sold as Volt Active Data.

Notable trade-offs

  • Throughput depends on keeping nearly all transactions single-partition; cross-partition work doesn't scale with the cluster.
  • Data must fit in memory, and long transactions stall their partition.
  • No interactive transactions: application logic that spans several round trips must move into a stored procedure.
  • Jepsen (2016) found stale reads, dirty reads and lost updates in VoltDB 6.3 under network partitions and recovery; 6.4 fixed them and appeared to provide strict serializability.