Week 12 · Oct 12-18, 2026 · final synthesis only
30-project synthesis, two unseen mocks, and final calibration
Rebuild the mechanism map, replay the weakest Core project, run two adaptive 45-minute unknown-system mocks, and calibrate for target-company/cohort retrieval without changing the personal route.
Loop Contract
| Goal | Finish Week 12 only as the final synthesis week matching Week 4 delivery quality. |
|---|---|
| Input scope | 12-week v3/batch specs, weekly contract, Week 4 pattern, prior weeks as references only, exact DDIA Ch13/Ch14 decision sections. |
| Execute | 30-project map, weak Core replay, two unseen mocks, Staff Q&A, crosswalk, final recall/PDF/audio, one combined self-check, production record. |
| Stop | Write reviews/week-12-production.md and stop at READY_FOR_REVIEW. No reviewer files. |
Daily Route
2026-10-12 MON
Day 1: 30-project mechanism map
Rebuild the full 30-project map from memory, then correct it against prior week artifacts as references only.
Open2026-10-13 TUE
Day 2: Weakest core replay: ChatGPT
Replay the selected weakest Core project and repair missing or weakest answer dimensions.
Open2026-10-14 WED
Day 3: Unseen mock A: incident truth platform
Run a realistic unknown-system mock focused on operational truth, derived dashboards, verification, and recovery.
Open2026-10-15 THU
Day 4: Unseen mock B: privacy-safe feedback calibration
Run a second unknown-system mock focused on feedback loops, privacy, accountability, and evaluator calibration.
Open2026-10-16 FRI
Day 5: Target company and cohort crosswalk
Calibrate the mechanism map for target-company style and external cohort topics while keeping the personal route independent.
Open2026-10-17 SAT
Day 6: Final synthesis and repair retest
Retest the weakest repaired answer and synthesize the two unseen mocks into reusable Staff trade-offs.
Open2026-10-18 SUN
Day 7: Final gate and READY_FOR_REVIEW
Complete Week 12 only, record evidence, and stop at READY_FOR_REVIEW.
OpenNavigable 30-project mechanism map
Each row is project -> core tension -> reusable mechanism -> invariant -> high-value follow-up.
BitlyDistributed Rate LimiterDistributed CacheDropboxFB News FeedFB Live CommentsInstagramWhatsAppGoogle DocsOnline ChessTicketmasterOnline AuctionRobinhoodPayment SystemAd Click AggregatorMetrics MonitoringYouTubeYouTube Top KStravaWeb CrawlerJob SchedulerPrice Tracking ServiceFB Post SearchNews AggregatorYelpUberTinderLocal Delivery ServiceLeetCodeChatGPT
| Project | Core tension | Reusable mechanism | Invariant | High-value follow-up |
|---|---|---|---|---|
| 1. Bitly CORE · W1 | 短链接写入必须唯一，读取却要极快并且可缓存。 | Slug allocation + system-of-record redirect table + cache-aside hot read path. | One slug resolves to one active destination version; cache entries are disposable projections. | How do you recover if a popular slug is cached with a stale destination after abuse takedown? |
| 2. Distributed Rate Limiter CORE · W1 | 全局限流要低延迟，但每次请求都跨区域协调会拖垮系统。 | Window/token state with atomic local admission, shard ownership, and explicit fail-open/fail-closed policy. | A request is admitted only by the owner of its limiting key and window budget. | What changes for login abuse versus optional analytics ingestion when the limiter is unavailable? |
| 3. Distributed Cache CORE · W2 | 缓存要吸收读流量，但不能被误认为业务事实来源。 | Ownership boundary between origin truth, cache projection, TTL/eviction, replication, and hot-key mitigation. | Losing the cache cannot lose business truth; rebuilding must be possible from origin state. | How do you cap hot-key load without breaking read-after-write expectations? |
| 4. Dropbox FULL · W2 | 大文件上传、metadata commit、跨设备同步必须在断线和冲突中可恢复。 | Multipart blob upload + verified metadata commit + sync cursor + conflict versioning. | A file version becomes visible once metadata points to verified blob parts exactly once. | What happens when CompleteUpload succeeds but the client never receives the response? |
| 5. FB News Feed CORE · W3 | 普通用户适合写时 fanout，名人流量会把 fanout 队列打爆。 | Hybrid fanout, feed materialization, pull fallback, ranking, and celebrity skew control. | A viewer feed can be stale, but it cannot permanently omit eligible source posts after repair. | How do you prove a follower missed a post because of lag rather than ranking policy? |
| 6. FB Live Comments CORE · W3 | 直播评论要实时，但断线观众不能靠 best-effort push 保证完整性。 | Connection gateway + ordered comment log + fanout channels + reconnect gap pull. | Each comment has a stream position; clients can detect and fill gaps. | What metric tells you websocket delivery is healthy while reconnect replay is falling behind? |
| 7. Instagram TRANSFER · W3 | 图片/视频 feed 看似像 News Feed，但媒体元数据、CDN 与分页语义改变瓶颈。 | Media metadata source of truth, CDN object path, ranked feed projection, cursor pagination. | A feed item references durable media metadata; CDN freshness cannot decide post existence. | Which News Feed assumption breaks when media processing lags post creation? |
| 8. WhatsApp CORE · W4 | 用户想要即时消息，系统必须区分 durable accept、device delivery 与 soft presence。 | Conversation sequence, durable message log, per-device mailbox, connection ownership. | ACK semantics name a durable point; reconnect uses cursor to replay missing messages. | What exact state changes when Redis pub/sub drops a message? |
| 9. Google Docs CORE · W4 | 多人编辑需要保留意图，last-write-wins 会静默丢失用户操作。 | Document operation log, revision/context, OT or CRDT, room broadcast, snapshot compaction. | Every accepted operation is replayable against its context or transformed into a later valid revision. | How do you recover if snapshot compaction writes a bad document image? |
| 10. Online Chess TRANSFER · W4 | 实时棋局不能照搬聊天的 eventual ordering；move 和 clock 裁决必须权威。 | Authoritative game sequencer, server clock, matchmaking pool, derived leaderboard. | For a game and move number, exactly one legal authoritative move is accepted. | How do you fence a recovered game server from accepting stale moves? |
| 11. Ticketmaster CORE · W5 | 高并发抢票要求低延迟等待体验，但座位所有权不能 oversell。 | Seat hold state machine, conditional write/CAS, TTL expiry, checkout idempotency. | A seat is in exactly one live ownership state: available, held, sold, expired, or released. | What wins when hold expiry and payment success happen at the same boundary? |
| 12. Online Auction FULL · W5 | 竞价展示要实时，但 winner 计算需要可审计的顺序。 | Per-auction bid log, total ordering, highest-bid projection, close/winner computation. | Accepted bids are ordered and immutable; projections can lag but cannot invent winners. | How do you handle a bid accepted before close but displayed after close? |
| 13. Robinhood FULL · W5 | 用户操作、broker 状态和交易所 ACK 会跨系统错位。 | Order lifecycle, idempotency keys, outbox, exchange reconciliation, market-data stream. | Each client order intent maps to at most one active broker/exchange order identity. | What happens if exchange accepts an order while local persistence times out? |
| 14. Payment System CORE · W6 | 支付体验需要即时状态，但金钱正确性来自 ledger 和 reconciliation。 | Double-entry ledger, idempotent payment intents, outbox, settlement reconciliation. | Ledger entries balance and corrections are compensating events, never destructive rewrites. | How do you prove a six-month-old dispute after provider logs expired? |
| 15. Ad Click Aggregator CORE · W6 | 广告指标要近实时，但 event time、dedupe、late data 会破坏简单计数。 | Event log, idempotent click key, stream windows, batch true-up, OLAP serving. | A unique click contributes once to the canonical aggregate for its event-time bucket. | What is the user-visible behavior for late clicks after the dashboard already showed a total? |
| 16. Metrics Monitoring CORE · W6 | 监控系统自己也是派生数据系统，不能只监控别人。 | Time-series ingest, aggregation, alert rules, cardinality guardrails, meta-monitoring. | Alert state is trusted only with known ingest lag, rule lag, and dropped-series counters. | What signal tells you an empty dashboard is a pipeline fault rather than normal traffic? |
| 17. YouTube CORE · W7 | 上传成功、转码完成、CDN 可播放和搜索可见是不同发布状态。 | Blob ingest, transcode job graph, publication state machine, CDN, metadata projections. | A video is public only after required assets and metadata reach a declared publication state. | How do you safely replay a failed transcode without publishing duplicate variants? |
| 18. YouTube Top K FULL · W7 | Top K 要快又要抗刷，但它本质上是从事件派生的榜单。 | Streaming counters, approximate heavy hitters, batch correction, anti-abuse filters. | The leaderboard is a projection over canonical events and can be rebuilt with policy version. | When should you hide, mark stale, or serve an old Top K result? |
| 19. Strava TRANSFER · W7 | 运动记录在离线设备上生成，服务端同步后才成为共享事实。 | Local activity log, route point buffering, sync cursor, completed-activity projection. | An activity has one durable event history; local stats are tentative until sync reconciliation. | How does pause/resume semantics survive GPS gaps and offline uploads? |
| 20. Web Crawler CORE · W8 | 抓取要覆盖十亿页面，但不能失控、重复、违反 politeness 或丢失 frontier。 | URL frontier, host politeness queues, leases, dedupe, content fingerprints, replay. | A URL attempt is either leased, completed, retried, or dead-lettered with an auditable reason. | How do you restart after losing half the workers without recrawling a hostile domain? |
| 21. Job Scheduler CORE · W8 | 长任务需要可靠执行，但 workers 会暂停、崩溃、重复执行。 | Durable job state, leases/fencing, retry policy, idempotent task handlers, DLQ. | A worker can complete a job only with the current lease/attempt token. | How do you distinguish slow job, stuck worker, and poison input? |
| 22. Price Tracking Service TRANSFER · W8 | 价格变化检测靠周期任务，但用户信任取决于新鲜度和通知准确性。 | Scheduled crawl, change detection, price history, alert threshold, retry/backoff. | A notification references an observed price event and threshold version. | What stale-price promise can you make when retailers rate-limit crawling? |
| 23. FB Post Search FULL · W9 | 帖子创建/点赞是写流，搜索结果是倒排索引与排名派生视图。 | Inverted index, ranking features, async indexing, recency/like projections. | Search freshness has a bounded lag and can be traced to post/index offsets. | How do you answer a user who can see a post by URL but cannot search it yet? |
| 24. News Aggregator TRANSFER · W9 | 聚合新闻要低延迟、去重和新鲜度，但 publisher crawl 是外部不可靠输入。 | Publisher ingestion, canonical article clustering, feed ranking, cursor pagination. | A feed item points to a canonical article identity with crawl and ranking provenance. | How do you prevent breaking-news spikes from teaching the system to over-rank duplicate sources? |
| 25. Yelp TRANSFER · W9 | 地理搜索、评分聚合和评论唯一性混在一起时容易混淆事实和排名。 | Business geo index, review uniqueness, rating aggregate projection, query/ranking split. | One user can leave one authoritative review per business; rating is a rebuildable aggregate. | How do you repair ratings after discovering a review-ingestion bug? |
| 26. Uber CORE · W10 | GPS 是会过期的观测值，派单是必须唯一的业务状态。 | Geospatial location projection + DriverOffer CAS/TTL/fencing + durable dispatch workflow. | A driver can hold at most one live offer and a ride can have at most one accepted driver. | How do you reject a stale accept after the matcher worker paused and resumed? |
| 27. Tinder FULL · W10 | 低延迟 feed 可以预计算，但 mutual match 不能依赖两个读时机刚好相遇。 | Candidate feed projection, no-reshow filter, swipe log, unique pair match creation. | A pair has at most one Match row, regardless of concurrent right swipes. | How do you recover missed notifications without creating duplicate matches? |
| 28. Local Delivery Service FULL · W10 | 库存可用性可以缓存，实际下单必须拥有物理库存。 | Nearby DC filtering, availability cache, inventory reservation transaction, order state machine. | Checkout cannot sell more units than source-of-truth inventory reservations allow. | How do you handle cancel after picker packed the order? |
| 29. LeetCode FULL · W11 | 在线代码执行要给用户快速反馈，但隔离、配额和长任务才是安全边界。 | Submission queue, sandboxed runner, test harness, artifact capture, quota/abuse control. | Untrusted code cannot escape the sandbox or consume resources beyond its tenant quota. | What evidence proves a timeout is user code behavior rather than platform failure? |
| 30. ChatGPT CORE · W11 | 模型回答、检索上下文、流式体验和 tenant 边界会同时影响信任。 | Request router, model serving, retrieval/context budget, streaming, evals, fallbacks, tenant isolation. | Every answer is produced under explicit context, policy, model, tool, and tenant boundaries. | How do you evaluate and roll back a retrieval change that improves average score but leaks private context in edge cases? |
DDIA Ch13/Ch14 decisions
| Decision question | Exact DDIA section | Design consequence |
|---|---|---|
| How do 30 prior projects become a reusable mechanism map instead of a name list? | Ch13 - 观察派生数据状态 | Every cached feed, dashboard, search result, ranking, or model answer is treated as derived state with observable lag, completeness, and rebuild path. |
| Which project outputs can be cached/materialized and which remain operational truth? | Ch13 - 物化视图和缓存 | The map separates source-of-truth state machines from materialized views so interview answers stop over-trusting caches. |
| How should the weakest core replay prove correctness rather than confidence? | Ch13 - 及时性与完整性 | The repair must state what can be stale, what must be complete, and which invariant metric proves the boundary. |
| How do we avoid believing a repaired answer just because it sounds polished? | Ch13 - 信任但验证 | The repair protocol requires audit evidence, counterexample injection, and a replacement score rule. |
| Mock A: where must correctness live in an incident-truth system? | Ch13 - 数据库的端到端原则 | Incident correctness is checked at workflow and reconciliation boundaries, not delegated to a single queue, dashboard, or database. |
| Mock A: how does the system remain reviewable after a wrong derived incident state? | Ch13 - 为可审计性而设计 | All derived incident decisions carry provenance, actor, model/rule version if any, and correction history. |
| Mock B: who is accountable for automated feedback and candidate scoring? | Ch14 - 责任与问责 | The design must assign accountable owners, audit logs, appeal/review workflow, and evaluator calibration. |
| Mock B: how can feedback loops distort future evaluations? | Ch14 - 反馈回路 | Cohort scores and training recommendations must be measured for self-reinforcing bias and rubric drift. |
| What privacy boundary applies to target-company and cohort retrieval? | Ch14 - 隐私与追踪 | The crosswalk retrieves only topic and rubric signals; it does not import private cohort transcripts into the personal route. |
| When is collection still unsafe even if it is useful for calibration? | Ch14 - 监视 | Do not collect hidden behavioral traces merely to improve scoring; require visible purpose and minimization. |
| How does the cohort plan stay opt-in and non-blocking? | Ch14 - 同意与选择自由 | External cohort participation creates retrieval prompts only; it cannot reorder the personal Week 12 sequence. |
| How is interview data reuse bounded? | Ch14 - 隐私与数据使用 | Mock recordings, transcripts, and score notes have explicit retention, purpose, and deletion rules. |
| How do we repair a bad index, ranking, dashboard, or eval after code/schema evolution? | Ch13 - 应用演化后重新处理数据 | Keep raw events and provenance so derived state can be reprocessed after rubric, schema, or bug fixes. |
| When should reads become explicit feedback events? | Ch13 - 读也是事件 | Reads can improve ranking/evaluation only when consent, purpose, and anti-feedback-loop controls are present. |
| Which Week 12 decisions require policy beyond engineering taste? | Ch14 - 立法与自律 | The final answer names policy, accountability, privacy, and audit choices when systems affect people. |
Applied AI source gate
No decorative AI source filling is scheduled. Applied-AI material is used only through the Week 11 ChatGPT replay when it changes eval, containment, tenant-boundary, or operational-decision scoring. Week 12's linked decision sources remain DDIA Ch13/Ch14.
Detailed lecture notes, audio, recall scripts, PDFs, Staff Q&A, and mock packs are archived locally and are intentionally not published on this site.