Nine prohibitions and their escapes

What caching forbids

Each item is something no design can do. Each comes from a derivation in an earlier article.

Freshness over a bypassable copy without paying coordination, verification, or staleness. There is no fourth currency.
Buying hit rate in a linear way. C ≈ N^h at α = 1. And at α ≤ 1, a fixed cache against a growing corpus decays, silently.
Optimal online eviction. A deterministic policy is k-competitive at best, a randomized one Θ(log k), and no policy escapes that.
Exact decisions about reusing cached query fragments. This is NP-complete at best and undecidable in general.
Exact invalidation, unless you keep the derivation graph at write granularity. Maintaining that graph costs about as much as the cache saved.
Consistency across entries without a single designated version cell.
Improving any percentile above h.
Splitting one hot key across nodes by hashing.
Caching a workload with no reuse. Nothing in the subject can replace reuse. This is the only prohibition with no escape at all.

The escapes and what each costs

Escape	Buys	Costs	Who pays
Immutability + content addressing	deletes invalidation and the write-routing dimension	prepaid write amplification + garbage collection	the writer and the janitor
Declaration (materialized views)	incrementality: the write path can be hooked	coupling; k views means k obligations on the write path	the writer
Algebra (monoids, sketches)	delta maintenance and reuse by rollup	exactness	the reader's precision
Verification (leases, ETags)	near-coherence	a round trip per read instead of a payload	the reader, cheaply
Randomization	removes coordination entirely	slightly more work, weaker worst-case guarantees	diffuse
Admission sketches	scan resistance with bounded damage	approximate frequency, one window of pollution	the one-shot traffic
Partitioning / pinning	isolation guarantees	utilization	idle capacity
Tiering	lower miss time instead of higher hit rate	latency on the miss path; a new coherence boundary	the misses
Client near-cache	kills the otherwise unfixable hotspot	a tier you cannot invalidate	staleness

(A monoid is an operation with a neutral element that can be combined in any grouping, such as sum or max. Aggregates with this property can be maintained from the change alone and rolled up. A materialized view is a stored query result that the database keeps up to date.)

Read the first row against the last. The best move in caching and the worst are the same move: put a copy somewhere the writer cannot reach. The only difference is whether you were able to make the data immutable first.

Case, resolvedConstructed case · Ifeoma at Dawnpost

The design review

At the design review Mateus asked what Dawnpost may not promise, who now pays for each trick, and whether the cache had become a dependency of the database. Where it started: What caching forbids →

Ifeoma answers Mateus with the list and the table above.

Forbidden: she can promise bounded staleness (TTL 10 minutes with jitter, plus delete on write) but not a consistent multi-article view, because Dawnpost has no designated version cell for the homepage; she can promise that the p99 stays at the database's p99 (400 ms), because a 95% hit rate improves no percentile above the 95th; and she cannot make a viral article cheap by hashing.
Escapes and who pays: archive Parquet files are immutable and content-addressed (the writer pays with compaction later); the homepage uses early probabilistic expiry (readers pay with a few seconds of staleness); the client near-cache removes the hotspot (readers pay with a 5-second lag); partitions isolate analytics (idle memory pays).
Dependency: yes, the cache is now a dependency. It is declared on the diagram with its leverage of 20, with a warm-up ramp and a load-shedding rule for when it fails.

Mateus signs off with one sentence for the runbook: "Every cache is a bet that nobody changed the data and that the database can survive without it. Write down both bets."