ADR-027 Raft before 17 November, with a cut off

ADR-008 put several store nodes after the first demonstration. Two things changed. Most companies that would run QuantyKube run a highly available store today and will not consider one that is not, so the talk without it answers the first question with a plan. And the reasons for waiting are gone: HTTP/2, gRPC and Watch are built, and QuantyCore has TLS for the transport between the store nodes.

Raft is built as ADR-008 says: a deterministic simulator first, in which several members run in one process against a simulated network and disks, driven by a seed, and which must find bugs planted in it before the protocol is written against it. Then three processes on one machine with real partitions in CI, then the three rented servers, one member on each.

The store fixes from the review of 20 October come first, since Raft replicates exactly that write path, and so does the store over TCP with TLS and client certificates, which Raft, the installer and the aggregated API (ADR-028) all need.

The cut off is 7 November. If the simulator and the partition test in CI do not pass by then, the demonstration shows a single store node and the talk says why. A Raft that loses a write on stage would cost more than one that is not shown.

The cost: three weeks of work go into the part of the project that is hardest to get right, alongside everything else for the same date, and the comparison with etcd has to be made again, three members against three.