Storage, Execution & Recovery Internals

Row, Column, and Compressed Storage

Row stores colocate attributes for one entity; column stores colocate values of one attribute across many entities. Encoding and compression exploit physical regularity, while late materialization delays row reconstruction so analytical operators carry only needed columns and compact position/selection vectors.

Row layout favors whole-record locality.
Column layout favors projection and scans.
Compression is part of execution economics.
Late materialization avoids premature width.
Updates complicate columnar locality.
Hybrid formats choose a unit of locality.

Same four orders, different locality

Physical-fit comparison
Need	Row-oriented	Column-oriented
One order, most fields	Few localized reads	Gather from several segments
Sum amount for paid orders	Reads unused columns	Reads status + amount; encoded scan
Single-row update	Localized in-place/version write	Delta/delete vector or segment rewrite
Compression	Mixed values limit runs	Homogeneous values expose dictionaries/runs/deltas