What physical clustering forbids
What does physical clustering forbid, what are the escapes, and what does each escape cost?
What this article forbids
- No total order keeps proximity in more than one dimension. This is permanent and geometric, not an engineering limit.
- The total skip power is fixed by the number of units, so multidimensional clustering redistributes it and does not create it.
- Clustering cannot survive writes. It can only be bought again.
The escapes and their costs
| Escape | What it buys | What it costs |
|---|---|---|
| Multiple sorted copies | Several directional bets | Linear write amplification, linear storage, and divergence between the copies |
| Space-filling curves | Insurance against the total-scan case | A hard floor of U^(1−1/d) on every single-dimension query, worsening fast with d |
| Richer summaries | Predicate classes that zone maps cannot serve | Metadata that recurses into its own instance of this problem |
| Nested granularity | Both bounds on unit size | Complexity at every level |
| Correlation | Skipping that you did not pay for | It makes every estimate wrong |
The through-line of Layout models, Mutation strategies and Physical clustering. Layout bet on which columns. Mutation bet on when to reconcile. Clustering bets on which predicate. It is the only one of the three whose failure is total and not proportional. That is why it is also the only one that needs ongoing measurement and not a one-time decision: a layout choice degrades gracefully when the workload drifts, and a clustering choice falls off a cliff (the query suddenly goes from fast to scanning everything).