Caching trades storage for transit, and reuse rate decides the direction

What it is

Caching is the most common barter of sustainable design: store content closer (on device, at the edge) in exchange for not having to transit it later. Whether the trade pays depends on the reuse rate — the probability the same data gets used again. At high hit rates, one act of storage amortizes across many saved transits for a net gain; at low hit rates, the write, the retention, and the eventual eviction all burn energy while few saved transits ever happen — the cache becomes net expenditure. Caching is not a default environmental win but a transaction to be accounted against access patterns; treating it as habitual optimization ("a cache can't hurt") does precisely the reverse for long-tail content.

Why it happens

The two sides have asymmetric cost structures. Transit cost accrues on every miss, scaled by content size and link type (per-GB energy on mobile networks far exceeds fiber); storage cost has three phases — write-once (transit plus medium write), retention (the standing power of static storage, local and cloud alike), and invalidation (rewrites and cleanup at TTL expiry or capacity eviction). Net gain = expected hits × per-transit cost − (write + retention + management cost), with reuse rate as the only free variable: popular content (high reuse) almost always pays; long-tail content (low reuse) almost always loses — and eviction machinery makes its storage cost recur: store, expire, store again. Prefetching pushes the trade to its extreme: it is a bet placed before any reuse has happened, wagering on "will be used"; for unpredictable browsing, the miss rate climbs with prefetch scope, and an unwatched prefetched video is pure loss.

Where it stops holding

Two bounds apply to energy-frame caching conclusions. First, user value comes first: offline availability and instant launch are real experiential gains, and some low-reuse caches are worth keeping for experience — the honest account then reads "paying energy for experience," not counting it as environmental savings. Second, the relative costs of storage and transit shift with technology: storage density and efficiency improve faster than some links, so historical conclusions expire — the accounting must be redone on current hardware parameters, not inherited rules of thumb.

Applying it

• Differentiate cache policy by access distribution: head content (durable high reuse) gets long TTL and generous prefetch; long-tail gets short TTL or no caching, driven by measured hit distributions rather than uniform defaults.
• Cap prefetch scope with behavioral data: prefetch only content whose next-play probability clears a threshold, and track unwatched-prefetch energy loss as a monitored metric, shrinking prefetch when it crosses.
• Impose eviction discipline on cache capacity: use re-store rate (the share of content repeatedly evicted and re-written) as the health metric — a high re-store rate signals mis-sized capacity or TTL design, pure waste.
• Verify: quarterly accounting of net cache benefit (transit saved minus the three-phase storage cost) for head and long-tail classes; a persistently negative long-tail shrinks its caching scope, and the conclusion feeds the next capacity plan.

Related

• Same group: P4.15.2 Video bitrate and resolution are the largest single variable · P4.08.2 Autoplay and prefetch amplifying traffic
• Adjacent: I2 Trade-offs in state synchronization and refresh · P4.15.3 Linear accumulation of inference cost
• Search terms: caching trade-off · hit rate · prefetch energy cost