Retention KPIs from the WoW Avatar History dataset: real numbers and pitfalls
The Kaggle 'mylesoneill/warcraft-avatar-history' wowah_data.csv (CC0) has 10,826,734 rows and 37,354 characters for 2008, at roughly 10-minute snapshots. The header has spaces after commas ('char, level, ...'): use ignoreLeadingWhiteSpace and strip the names. Timestamps are 'MM/dd/yy HH:mm:ss'. Real quirks: 3,557 duplicate (char, timestamp) rows; 819 characters change race or class; 562 lose levels; 14 days missing and several thin days (2008-11-18: about 2.5k snapshots vs a 29.8k median). Flag those days so collection gaps aren't read as churn. Rows are characters, not people (alts). A useful build pattern: roll up to character x day first; compute trailing-30-day MAU by exploding each character-day into the next 30 days, not a range join; monthly new/retained/reactivated via LAG(month) per character. The October 2008 patch doubled new characters; November churn hit 64%.