Valkey's Madelyn Olson says async I/O threading lifted one process to about 1M requests a second

Taken at face value, Olson's figures put the gain at four to five times per process ^[18]. The interview does not describe the hardware, value sizes, client count or pipelining behind either number ^[1]. Those variables decide whether a throughput result moves from one team's test setup to another team's cluster. I'd treat the million as a hypothesis to reproduce on production-shaped traffic before it changes a capacity plan.

The control that ships with it is easier to credit. "Unlike the previous I/O-thread arrangement, thread configuration can be changed at runtime, allowing operators to add cores to help handle burst workloads," Olson said ^[9]. Under the old arrangement the setting could not be changed while the server ran ^[9]. A setting that moves live can be used during the burst itself.

I think the stronger engineering case is in the replication and cluster work. Before dual-channel replication, a primary syncing a replica sent the full copy of the dataset first and the stream of later changes after it ^[11]. The 8.x releases send the two in parallel. According to the interview, that reduces memory pressure on the primary during synchronisation ^[11]. Presumably the saving comes from no longer holding the changes made during the copy on the primary until the copy completes.

Atomic slot migration in version 9 closes a failure mode in resharding. With the old key-by-key method, a failed operation could stop a cluster with only part of its data moved ^[12]. The new method stages the data in advance and then cuts over, so a migration either completes in full or does not happen at all ^[12]. I suspect most operators assumed they already had that behaviour.

The project is more than two years past the fork from Redis ^[7]. Both developers InfoQ interviewed work at companies among the eight that supply its maintainers ^[8]. Olson works at AWS, sits on the technical steering committee and was a Redis committer before the split ^[16]. "I helped create the fork after the license change," she said ^[16]. Viktor Soderqvist works on open source at Ericsson and was also involved in creating the fork ^[17].