Files
aevolution/docs/adr/0148-durable-cellular-migration-staging.md
dingo 0515702964
Foundation / checks (deploy) (push) Successful in 5s
Foundation / checks (ecs) (push) Successful in 2m25s
Foundation / checks (network) (push) Successful in 20m44s
Foundation / checks (client) (push) Successful in 21m57s
Foundation / checks (tsan) (push) Successful in 22m46s
Foundation / postgres-tools (push) Canceled after 0s
Foundation / server-image (push) Canceled after 0s
Foundation / image (push) Canceled after 0s
Foundation / checks (checks) (push) Canceled after 30m2s
Foundation / checks (postgres) (push) Canceled after 29m56s
M12.4: retain read-only destination migration stages
2026-10-18 09:32:50 -04:00

3.3 KiB

ADR-0148: Durable destination staging without activation

Status

Accepted and locally qualified for scoped M12.4i implementation, 2026-10-18. Executed evidence and limits are in VALIDATION. Paired result publication and activation remain subsequent work. This selects no route, automatic migration or deployment.

Decision

Extend ADR-0147 preparation with destination-authorized staging at the exact captured destination head. Lock the two regional lease rows in region order, check current destination authority and the retained source freeze, and refuse pending destination work. Retain one immutable staging record and one bounded destination pending pointer. A region cannot participate in two workflows or simultaneously be a source and destination. Both biological heads stay unchanged.

Restore both verified predecessor payloads and replay the exact retained trace. Copy the source predecessor into the destination object namespace, and upload and read back the derived destination candidate there. Record its SHA-256 and length. It is read-only evidence: it is never a committed checkpoint, owned stock, active population, census entry or permission to run biology. Recovery replays both cuts from the destination namespace and compares the complete candidate bytes. Equal aggregate material alone is insufficient.

New checkpoints, profile changes, intervals and player reservations are refused at a staged destination. Its host pauses readiness and advancement, retains exact requests after unknown replies, renews authority and recovers the pause across replacement. No lease expiry, drain or timeout releases it. Explicit source abort atomically appends the existing immutable abort and clears both pending pointers. It requires no destination authority because no stock has moved. Stage and abort serialize on the source lease row: either staging wins then abort removes both pointers, or abort wins and new staging is refused. An exact retry of a previously staged, aborted request returns that retained outcome and never reopens either pointer. Destination movement before first staging requires explicit source abort and a new preparation, without silent rebasing.

Keep all staged objects and predecessor evidence after abort. Backup already copies immutable namespace objects and shared metadata; this is not cellular clone/restore qualification. The future paired publisher must recheck both current fences, heads, staged object bindings and unsettled request, then commit both results and the common receipt atomically before activation.

Alternatives, risks and qualification

An in-memory ghost disappears on process loss. Activating the uploaded payload duplicates source stock. Advancing a destination after staging invalidates the paired candidate. Freeze the whole destination instead; the coarse pause and extra immutable storage are explicit costs, not a final movement latency policy.

Require both representations, real player-origin wildlife, full ghost replay, unchanged heads and stock, conflicting/duplicate requests, pending-work refusal, stale destination authority, replacement/restart, missing/corrupt/semantically false ghost evidence, ambiguous replies, process kills, stage/abort races and host readiness. No coupled ecology, world-cut or multigeneration claim follows.