bsafronov Migrate System
Reach the requested transition stage while preserving the guarantees needed by remaining consumers and recoverable data.
- Identify which readers and writers can exist at each relevant stage, which representation is authoritative, and what a rollback would actually restore. Include writes made after migration begins: reverting code or restoring an old backup alone may lose those writes or leave them unreadable.
- Choose the smallest viable transition from the actual rollout constraints. An additive form, backfill, consumer switch, and removal are useful only where coexistence requires them. A coordinated offline conversion can instead use a verified maintenance window. Keep transitional code temporary and give its removal a concrete compatibility condition.
- Make partial progress safe to resume. Define checkpoints, retry identity, and handling of concurrent writes or failed batches from the storage guarantees; do not infer atomicity from a successful happy path. Preserve source data until recovery no longer depends on it. Make incomplete, rejected, and conflicting records observable instead of silently skipping them as complete.
- Prove old and new paths at the stages the task reaches, using representative data and failure points. Exercise interruption and retry, overlapping writers where allowed, and rollback after new writes. Check the resulting data and consumer behavior, not just that the migration command exited successfully. Distinguish local simulation from storage, deployment, or production proof.
- Perform only the requested stage. Preparation and reversible implementation use the user's existing authorization. Execute destructive contraction only when it is in that authorization and its compatibility and recovery conditions hold; a plan may describe it without executing it. Carry forward an explicit authorized destructive phase rather than asking for the same approval again.
Stop when that stage's compatibility, data preservation, and recovery claims have supporting evidence, or name the specific claim still unproved. Interface choice, work decomposition, specialized library conversion, and requested test-first sequencing can contribute independent outcomes; none is a prerequisite. Keep ordinary structure-preserving edits outside this transition lifecycle.