Akka partition replay
Akka.NET, a network split, replayed
Summary
An Akka.NET cluster with its network split was run and replayed 3 times, and each replay matched its run on all 1,476 network messages, with the same 42 messages cut. Molina's replay matched the original down to the delivery time and contents of each message. The result covers one replay code, one machine and one build, and three pairs are too few to say how often a replay differs.
The full experimental record
What I ran
A cluster partition of Akka.NET nodes was run on one seed by one hypervisor build, in 3 pairs of runs, each a run and its replay. In every pair the replay matched the run: each of the 1,476 network frames (the pieces of data the servers send each other) had the same outcome (delivered or cut) and, for delivered frames, the same delivery time and the same contents (compared by a fingerprint of the data), in the same order, and both runs cut the same 42 frames, on one seed, one machine and one build.
The first moment each node gave each notice (a member up, unreachable or removed, a new leader), on that node's own clock, did not differ between a run and its replay.
Three pairs are too few to put a useful ceiling on how often a replay differs. They leave room for a replay that differs more often than not.
Each of the 3 pairs is a run and its replay: the gate compares the run with its replay within each pair, and does not compare runs across pairs.
What it shows
MIGHT. In 6 of 6 runs of the cluster partition on one seed, the run gave the same frame fates, order, cut and notice moments as the other run of its pair.
What it does not show
That the replay holds on another seed, on another machine or on another build of the hypervisor.
That anything inside a node matched beyond the frames it sent and the notices it gave.
That the worlds ran on cores reserved for them. The ledger row of this gate says isolation was not applicable, and the machine's load average when the run began, a figure the runner recorded, was 7.31, 11.52 and 14.93 over one, five and fifteen minutes.
The source of this experiment names no lease on the world farm.
Conditions
One seed and one hypervisor build across the 3 pairs (committed-tree ledger PASS).
Where it comes from
evidence/20261103T143304Z-readiness-cross-world.out
This page is the short form of the record. The finding and its limits come first, and I left out what did not fit.