OmniTable
paper Your tags
Your notes
Ant Group's blueprint for the data layer under its Ling and Ring models, where petabyte-scale corpora sit in hundreds of physical tables, feature engineering is hand-orchestrated per table, and lineage is largely absent. OmniTable is a unified wide-table layer built on logical unification with physical separation: one logical schema over multi-source data and thousands of derived features via logical-physical mapping, declarative feature lifecycle management, and lineage throughout. The 19-author paper reports it serving about 35 PB and cutting a curation cycle from 14 days to 2.5.