Architecture, data, networking, and sync interviews
Pattern comparisons
MVC/MVP/MVVM/MVI describe responsibility/state/event organization, not complete architectures. Compare state complexity, lifecycle, testability, team vocabulary and ceremony. Clean Architecture protects business policy through dependency direction; it does not require a use-case class for every repository call.
UI -> presentation/domain contracts -> repositories
implementations -> local/network
Explain source of truth, ownership and data flow before layer names.
Data questions
Repository: mediates domain data and sources; it is not merely a Retrofit forwarding wrapper. Define freshness, cache, errors, concurrency and writes.
Room transaction: protects a local atomic invariant; it does not make server/local distributed work atomic. Use idempotent commands/outbox/reconciliation.
DataStore: typed/preferences small durable state; not relational query/large datasets. SQLite/Room for structured relationships/transactions.
Paging: UI consumes PagingData; local source plus RemoteMediator can provide offline truth. Remote keys, refresh/append errors, duplicates and invalidation require explicit design.
Networking/sync questions
Define timeouts by phase/operation, cancellation, retryable errors, exponential backoff/jitter, idempotency, auth refresh serialization, pagination, error body, serialization compatibility and observability. Retrying POST is safe only with protocol/idempotency design.
Offline-first means local readable state plus durable user intent, sync state/conflict policy, freshness UX and reconciliation. Last-write-wins is only valid when domain consequences permit it.
Design drills
- Cache-first product page with stale-while-revalidate and explicit error/freshness UI.
- Offline edit from two devices: version/precondition, conflict response, merge or user choice.
- Token expiry for 100 parallel calls: one refresh, queued callers, failure/logout and no recursive interceptor loop.
- Database migration across five released schemas with large data/disk pressure and forward repair.
Red flags
- ❌ “Single source of truth means only one physical source.”
- ✅ one authoritative model per context; data can synchronize from several systems.
- ❌ “Exactly once sync.”
- ✅ at-least-once attempts plus idempotency/deduplication/reconciliation.