KeyDB v6.3.0 Release

KeyDB version 6.3.0 is the first open source release to include previously KeyDB Enterprise Features! This release is the culmination of many years of work to move away from KeyDB’s global lock and enable better scaling across cores.

Active Replication PSYNC

Active Replication has been a feature of many KeyDB releases, but always required a full sync while connecting to a new peer. In this release KeyDB now supports partial synchronization to enable fast cluster healing. In addition KeyDB is able to update its replication offset based upon the knowledge of peer nodes, meaning a full sync from one node is enough to permit partial syncs to other nodes within a mesh.

This change greatly reduces the time to add a new node to an active replication cluster. In our testing we’ve seen the time reduced from well over 10 minutes to a minute and a half for a 12 node mesh cluster.

Async Commands

Async commands are commands which can execute without the global lock. This feature must be enabled first in your configuration file by setting “enable-async-commands yes”. When enabling async commands the consistency is modified slightly, in particular writes from other clients may take a configured amount of time to become visible to other clients. This time is configured with the max-snapshot-slip configuration parameter.

The ordering rules with async commands are as follows: Writes are always totally ordered among clients Writes are always immediately visible by the client that performed the write

For most applications the slight modification to the consistency model of Active Replication will not be visible. If your application relies on ordering between different clients this feature is disabled by default. In addition a new command called “lfence” is provided for clients that do rely on consistency across clients but are willing to insert fences to assert total ordering at the correct times.

The following commands support async operation: GET MGET

Additional read-only commands will get added over time. If a command is of specific interest to you please post an issue to help us prioritize your use case.

Async Rehash

KeyDB relies upon a spinlock to synchronize threads. In prior versions of KeyDB the CPU time consumed while waiting for the lock to be acquired was wasted. KeyDB now has the ability to perform rehash during this time using otherwise wasted CPU time for a more useful purpose. In many cases this can almost completely hide the overhead of rehashing. No configuration is required to enable this functionality, it operates by default.

In Process Background Saving

Historically Redis has used the “fork” system call to create a new process which performed the background save. This made the code easier to follow as the kernel handled the hard task of copy-on-write of memory while background saving was in progress. However the downside to this approach is that it is not possible to accurately estimate the total amount of memory required including background saving.

KeyDB now uses a new “snapshot” system to create snapshots of the database at a specific period and copy-on-write new changes within the process. As a result the maxmemory setting is now a global setting and will include memory used for background save.

For backwards compatibility the semantics of maxmemory are modified slightly during a background save. While a background save is in progress KeyDB will permit memory consumption to exceed maxmemory by up to 20%. This is to simulate the old behavior of background save memory consumption not being counted towards the maxmemory setting while permitting easy calculation of a total upper bound for memory usage.

IStorage Interface

This release does not yet include KeyDB’s FLASH feature based upon RocksDB. However the release does include KeyDB’s IStorage interface which is the backbone of our persistent storage feature. By implementing this interface for your custom storage solution you will enable all KeyDB features.

Thanks To Our Contributors:

Thank you to everyone who contributed to KeyDB with bug reports, design and testing:

Ben Schermel Firaenix: Bug Report #393 - Crash in sorted sets with long names Kerog: #401 Error after include *.confg Gvsafronov: Fedora 35 Compile Failure Talkabout: Always providing helpful and detailed bug reports, not limited to #378 Inakisoriamrf: #383 crash Alebcay: #384 macOS build break Server2245: #379 Install issue on RHEL7 Antiarchitect: #380, and many others Tchernomax: #352 missing support for systemd Heng Kuang: Detailed bug reports and design suggestions Kajaruban Surendran: Detailed performance analysis of async rehash and other features Paul Chen: Your enthusiasm and willingness to help debug issues has been deeply appreciated.

Special thanks to Huawei Canada Research Team for your detailed and in-depth bug reports and feature requests and design participation.

Thanks to the EQ Alpha Team that made this release possible: Malavan Sotheeswaran, Vivek Saini, Christian Legge, Peter Liang, Ben Schermel, John Sully