Abstract

The measured memory-and-scaling case for running an entire multi-tenant cloud as one Go binary — 56 in-process subsystems over shared-nothing storage, where every tenant is its own post-quantum-encrypted SQLite file and writes shard embarrassingly with no global lock. Measured through the production open path: an actively-served tenant costs ≈0.25 MB of RAM and an idle tenant essentially zero, so memory scales with concurrency, not registrations — one billion registered tenants at 1% peak concurrency fits ≈20 commodity 128 GB nodes. Measures the shared-Postgres microservice-mesh alternative at 11–32× the memory footprint, and is explicit about writer-pinning failover, hot-tenant skew, and cross-tenant query limits.