Release Notes



  • Many short-lived file access metrics were being created.
  • A completed backup could be improperly marked as incomplete.
  • In rare scenarios the resolvers could fail to make progress.



  • Significantly improved cluster performance in a wide variety of machine failure scenarios.


  • Clients can now load multiple versions of the client library, and will gracefully switch to the appropriate version when the server is upgraded.
  • A new operating mode for fdbbackup writes backup data files into the blob store.
  • Transactions no longer automatically reset after a successful commit.
  • Added ability to set network options with environment variables.
  • Added a new API function for determining the value to which atomic versionstamp operations in a transaction were transformed or would have been transformed.
  • Improved logic for integrating manually-assigned machine classes with other constraints on role locations.
  • Added a new machine class stateless which is the top priority location for resolvers, proxies, and masters.
  • Added a new machine class log which is the top priority location for transaction logs.
  • Trace events are now event metrics that are exposed in Scope.


  • A log could attempt to recover from a partially recovered set of logs when fast recovery was enabled.
  • A rare scenario could cause a crash when a master is recovering metadata from the previous generation of logs.
  • Streaming mode EXACT was ignoring the target_bytes parameter.