What layout models forbid
What do layout models forbid, what are the escapes, and what does each escape cost?
What the layouts forbid
- No single layout gives both O(1) access to a whole entity and an optimal scan of one attribute across entities. They need contradictory adjacency, and a byte has one address.
- No heavily encoded layout supports fine-grained in-place update, because encoding fuses values into blocks.
- No layout gives both stable physical identity and physical ordering by key.
The escapes, and what each costs
| Escape | What it buys | What it costs |
|---|---|---|
| PAX | A bounded compromise between entity and attribute access | n decode operations per point read, plus a rewrite floor of G |
| Dual storage (a row-formatted delta over a columnar base for mixed workloads, "projections", several sort orders of the same data) | Both access shapes | 2× storage, proportional write amplification and, more important, a freshness gap between the copies. Storage is cheap; the staleness is the real cost. |
| Late materialization | Buys back part of the column store's reconstruction cost | Planner complexity, and it only helps selective queries |
| Shredding | Columnar speed on documents | Validity that depends on a schema nobody declared |