RaftShift: An Experimental Study of Cost-Aware Raft Leader Transfer

Abstract

RaftShift is an experimental advisory controller for planned Raft leadership transfer. It compares forecast cumulative benefit with a component-based activation allowance, then invokes native transfer and voting. This technical report consolidates implementation, cloud, formal-model, and later measurement evidence while separating their validation scopes. Frozen three- and five-region workload experiments contain 960 evaluation runs and 960,000 operations; 22 of 24 workload cells favor RaftShift over retaining the incumbent after the declared multiplicity correction. These results do not establish the ledger's incremental advantage over simpler transfer rules. Dynamic comparisons against native transfer and a GeoLM-derived comparator remain unresolved; the cross-topology short-pulse non-inferiority gate fails. Disjoint WAN readiness validation rejects all tested activation budgets. A later corrected 64-run local measurement campaign retains all 2,496 offered requests, including 14 HTTP 503 failures charged according to the predeclared 2,000-ms loss convention. The one-time move-cost diagnostic fails in four of six horizon comparisons. These are forced-action engineering diagnostics, not a policy-superiority test or proof of RaftShift inferiority. Corrected bounded TLA+ checks are reported with their finite-state coverage and implementation limitations. Independently calibrated ledger superiority and a deployable economic cost model remain unestablished. This is an author-prepared technical report, not a peer-reviewed journal article. The accompanying ZIP contains report source and compact evidence, including per-request results; it is not a complete raw replay archive. Earlier raw cloud evidence remains in the linked artifact records. The September 12 local raw observation streams and replica disk images are not included in those earlier records. Report and original data/documentation: CC BY 4.0; software/model code: Apache-2.0, as scoped in the included license files. AI assistance and sole-author responsibility are disclosed in the report.