Offline-first behavior across features

Independent feature repositories still share device storage, network, identity, battery and user expectations. Without platform policy, each feature creates incompatible “offline” semantics.

Shared policy, domain-owned conflicts

Platform provides connectivity hints, durable scheduling, account partitioning, retry/backoff, queue observability, storage quotas, transaction/idempotency primitives and common sync states. Domains own merge/conflict rules because cart quantity, chat edits and bank transfers have different semantics.

data class PendingMutation(
    val id: MutationId,
    val accountId: AccountId,
    val domain: String,
    val idempotencyKey: String,
    val createdAt: Instant,
)

Persist payload/version and state transitions too; never put secrets in generic diagnostic fields. Schedule fairly so a noisy media upload queue cannot starve a small checkout command. Apply per-domain concurrency, priority, network/battery constraints and backend rate limits.

Cross-feature invariants

• account switch isolates/cancels the correct queued data;
• logout and deletion handle pending/private data;
• storage pressure evicts rebuildable cache before irreplaceable user intent;
• sync is idempotent and process/device-restart safe;
• conflict is surfaced honestly, not overwritten silently;
• global diagnostics reveal backlog by domain without high-cardinality/PII.

Practical mistakes

• ❌ one generic last-write-wins sync engine for every domain.
• ✅ share mechanics; keep domain conflict policy explicit.
• ❌ each feature observes ConnectivityManager and retries immediately.
• ✅ centralize constrained scheduling/backoff while treating connectivity as a hint.
• ❌ clear all storage on quota pressure.
• ✅ classify durability/value and use governed eviction.

Exercise

Design fair offline scheduling for chat messages, photo uploads, cart edits and payment intent under low battery and intermittent network.