WEEK 09 · 2026.09.21-09.27 · WRITER GATE

第 9 周行动指南 - 索引、排序与地理搜索

FB Post Search full coverage, then transfer the same retrieval/ranking muscles to News Aggregator and Yelp.

中文深度讲义 Staff Q&A GPT Live Mock 证据验收

能力主线

本周只训练索引构建、候选召回、排序、新鲜度、地理搜索和质量信号，并且强制区分 recall latency、rank latency 和 presentation latency。

RECALL

Candidate retrieval

倒排索引、时间索引、geo cell 和粗过滤负责找到候选，不负责最终业务顺序。

RANK

Feature scoring

recency、like_count、source trust、distance、rating、review quality 和 abuse signals 进入排序。

PRESENTATION

Result assembly

snippet、highlight、media card、map pin、pagination cursor 和缓存命中属于展示延迟。

项目角色

日页面使用每个 live heading 的原始标题和 direct anchor；production record 写完整 disposition。

FULL · 41 live headings

FB Post Search

Full coverage with reading anchors and recall artifacts.

打开项目首页

TRANSFER · 20 live headings

News Aggregator / Google News

Transfer coverage: blind attempt, delta, invalid assumptions, new bottleneck and follow-up.

打开项目首页

TRANSFER · 21 live headings

Yelp

Transfer coverage: blind attempt, delta, invalid assumptions, new bottleneck and follow-up.

打开项目首页

DDIA 决策地图

只使用 Ch4 和 Ch13 的精确 decision anchors。

日	项目	改变的决定	链接
Day 1	FB Post Search	keyword search is an inverted index, not SQL LIKE	https://ddia.vonng.com/ch4/#sec_storage_full_text
Day 1	FB Post Search	recency and like sort need index/ranker fields, not presentation sorting	https://ddia.vonng.com/ch4/#sec_storage_index_multicolumn
Day 2	FB Post Search	large write volume chooses update pipeline and refresh interval consciously	https://ddia.vonng.com/ch4/#sec_storage_btree_lsm_comparison
Day 2	FB Post Search	rebuild/search index as derived state, then swap safely	https://ddia.vonng.com/ch13/#id340
Day 3	News Aggregator	feed rows and index documents are derived from crawled source articles	https://ddia.vonng.com/ch13/#id446
Day 3	News Aggregator	homepage/category feed caches are materialized views with staleness budgets	https://ddia.vonng.com/ch13/#id451
Day 4	News Aggregator	freshness SLO requires observing index lag and stale derived state	https://ddia.vonng.com/ch13/#sec_future_observing
Day 4	News Aggregator	schema/ranking evolution needs reprocessing rather than hand edits	https://ddia.vonng.com/ch13/#sec_future_reprocessing
Day 5	Yelp	geo search uses multidimensional indexes or grid cells, not naked lat/lng scans	https://ddia.vonng.com/ch4/#sec_storage_multidimensional
Day 5	Yelp	category + price + open_now + distance are composite filter/rank constraints	https://ddia.vonng.com/ch4/#sec_storage_index_multicolumn
Day 6	Yelp	text plus geo search combines full text retrieval with precise second-pass filters	https://ddia.vonng.com/ch4/#sec_storage_full_text
Day 6	Yelp	review/rating aggregate and search index are derived state that needs verification	https://ddia.vonng.com/ch13/#sec_future_verification

固定时间合同

时间	Week 9 不变量	证据
08:50-10:55	系统设计主块，先闭卷/盲做，再用 canonical anchors 修正。	每天 4 个 bounded segment。
14:30-16:15	NeetCode Graphs，连续单一 tag，每天 exactly 3 slots。	21 个 data-neetcode-slot。
20:30-21:15	recall / Q&A / transfer follow-up / mock。	day 页面和 scorecard 记录。