Skip to main content

Case study — large news feed with Paging

Requirements​

Personalized/topical feed, fast cached open, pagination/refresh, offline reading, media, bookmarks, ranking freshness, insertion/removal, ads/content policy and accessible adaptive UI.

Compose Lazy list <- PagingSource(Room items/feed entries)
                            ^
                            |
                    RemoteMediator <-> Feed API/cursor
Article details/bookmarks -> repositories/local source of truth

Separate article entity from feed-entry rank/reason/cursor metadata. One article may occur in multiple feeds. RemoteMediator transactionally stores page and remote keys; refresh policy considers age, user action and invalidation. Opaque server cursors avoid client offset errors when ranking changes.

Stable item IDs prevent wrong UI state. Do not use list position as identity. Deduplicate by content/entry contract, preserve scroll anchor across refresh, model refresh versus append errors separately and cache media with bounded quota.

Offline read uses Room truth and displays freshness. Bookmark/action commands are idempotent and sync later. Ranking/exposure analytics fire when meaningfully visible, deduplicated per contract—not on every recomposition/bind. Server validates ad/content eligibility.

Failure/testing/operations​

Test empty/duplicate/overlapping/reordered pages, cursor expiry, deleted article, refresh during append, process death, huge DB, low storage, offline bookmark, account switch, font scale and fast scroll/jank. Observe time-to-content, cache hit/age, page latency/error, duplicate/gap rate, DB/size, image failures, slow frames and exposure data quality.

Scale/trade-offs​

Precomputed feeds/cursors improve read latency at storage/fan-out cost; fan-out-on-read improves flexibility at latency/compute cost. Hybrid by user activity. Aggressive prefetch improves smoothness but costs data/battery/backend; tune from evidence.