Caching Agent Skill

Caching

Put a copy of hot data closer to the reader so most requests skip the slow path.

Caching is the highest-leverage move for read-heavy systems — and the easiest to

get subtly wrong, because a cache adds a second source of truth that can serve

stale or wrong data, and can *amplify* an outage when it misbehaves.

When to reach for this

Reads dominate (a high read:write ratio from back-of-the-envelope); the same

data is read repeatedly; the datastore is the read bottleneck; or recomputation

is expensive. A cache buys read latency and offloads the origin.

When NOT to

Write-heavy or read-once data (low hit rate — pure overhead). Data that must be

exactly current with zero staleness (a cache is a stale copy by nature; when

strict freshness is required, go to the source or use consistency-coordination). Don't

add a cache before a number shows reads are the problem (YAGNI) — it's a new

failure mode and a second thing to operate.

Clarify first

• Read:write ratio and hit rate — is the working set cacheable? (→ back-of-the-envelope, 80/20.)
• Staleness tolerance — seconds? minutes? must reads see their own writes?
• Working-set size — does the hot set fit in RAM across cache nodes?
• Consistency on write — can the cache briefly disagree with the store?
• Eviction trigger — what's the access pattern (recency? frequency? time-bound?).

The options

Where to cache (often layered): client/browser → CDN edge (→ content-delivery)

→ application/in-process → distributed cache (Redis/Memcached) → database buffer

pool. This skill focuses on the application and distributed layers.

Read strategy

• Cache-aside (lazy): app checks cache, on miss reads the store and populates.

Use when reads are unpredictable; the default for most systems.

• Read-through: the cache library fetches from the store on miss. Use to keep

app code simple and caching policy centralized.

Write strategy

• Write-through: write cache and store synchronously. Use when reads right

after writes must be fresh and slower writes are acceptable.

• Write-back (write-behind): write cache now, flush to store async. Use for

write-heavy/bursty paths that tolerate a small loss window.

• Write-around: write only the store; let the cache fill on read. Use when

written data is rarely re-read soon (avoids cache churn).

Eviction policy

• LRU for recency-skewed access (most common); LFU for stable popularity;

TTL to bound staleness; FIFO rarely. Match the policy to the pattern.

Trade-offs

| Option | What it solves | What it worsens | Change it when |

|---|---|---|---|

| Cache-aside | Simple, resilient (cache down ⇒ just slower) | First read per key is a miss; risk of stale after writes | Misses are too costly → read-through + warming |

| Read-through | Centralized, clean app code | Couples app to cache lib; cold-start misses | Custom per-key load logic is needed |

| Write-through | Fresh reads after write | Slower writes; writes cached data that may never be read | Writes dominate and aren't re-read → write-around/back |

| Write-back | Fast, absorbs write bursts | Data loss window on crash; complex | Durability of recent writes is required |

| Write-around | No churn from write-only data | Recently written keys miss on first read | That data IS read right after write → write-through |

| TTL eviction | Bounds staleness automatically | Mass expiry can stampede the origin | Add jitter / soft-TTL refresh |

Behavior under stress

A cache that misbehaves doesn't just stop helping — it can take down the origin.

• Thundering herd / stampede: a hot key expires (or the cache restarts) and

thousands of concurrent misses hit the store at once. *Mitigate:* per-key locks

/ request coalescing (single-flight), early/probabilistic refresh, TTL jitter.

• Cache penetration: requests for keys that don't exist bypass the cache every

time (often malicious). *Mitigate:* cache the negative result (short TTL), or a

Bloom filter in front.

• Hot key: one key (a celebrity, a viral item) exceeds a single node's

throughput. *Mitigate:* replicate the key across nodes, add a local/L1 tier, or

shard the value.

• Eviction storm / cold cache: after a flush or deploy, hit rate craters and

the origin sees full load. *Mitigate:* warm critical keys; ramp traffic.

• Stale-after-write: the store changed but the cache didn't. *Mitigate:*

invalidate on write, or write-through, or short TTL — pick per staleness budget.

Monitor: hit rate, p99 latency, eviction rate, key distribution (hot spots),

and origin QPS during cache restarts.

How to apply

• Clarify the inputs — confirm the read:write ratio, staleness budget, and

hot-set size (see *Clarify first*). If no number yet shows reads are the

bottleneck, stop — a cache is not needed yet (→ back-of-the-envelope).

• Pick the strategies from the trade-off table — choose a read strategy

(cache-aside is the default), a write strategy keyed to the staleness budget,

and an eviction policy matched to the access pattern.

• Set the key knobs — define the key naming scheme, TTL (with jitter), and the

per-key-family invalidation event. Decide negative-caching and single-flight up

front, not after the first incident.

• Stress-test the choice — walk each failure in *Behavior under stress*

(stampede, penetration, hot key, cold cache, stale-after-write) and confirm a

mitigation is in place for the ones the traffic profile can trigger.

• Size it with numbers — fit the hot set in RAM across nodes, sanity-check the

target hit rate (90%+), and confirm the node count covers peak QPS (→ *Numbers

that matter*).

• Pick a provider — default to the generic recipe; only open a provider file

if the user named a cloud (see *Choosing a provider*).

• Show the trade-off in the answer — give a `Decision | Solves | Worsens |

Change it when…` table with one row per choice you made here (e.g. the

option, its store, its failure policy); the last column is a concrete trigger

(a number or observable signal). Then name what breaks first.

Dos and don'ts

Do

• Default to cache-aside; it stays correct (just slower) when the cache is down.
• Set a TTL on every entry and add jitter so keys don't expire in lockstep.
• Add single-flight / request coalescing for hot keys *before* launch.
• Invalidate on write (or write-through, or short TTL) to a stated staleness budget.
• Size the cache to the hot set and alert on hit rate, evictions, and origin QPS.

Don't

• Don't add a cache before a number proves reads are the bottleneck (YAGNI).
• Don't cache data that must be exactly current — go to the source instead.
• Don't let a mass-expiry or cold start dump full load on the origin (warm, ramp).
• Don't ignore non-existent-key floods — cache negatives or front with a Bloom filter.
• Don't reuse a key across schema versions — bump the version (...:v2) instead.

Numbers that matter

A cache node serves ~100k–1M QPS, far above an RDBMS (~1k). Memory access is

~100 ns vs ms-scale disk — the reason caching wins. Size the cache to the hot set

(~20% of data ≈ 80% of reads). Target hit rates are usually 90%+; below that,

question whether the data is cacheable. → back-of-the-envelope.

Interface sketch

A cache entry is a contract: a key (stable, namespaced, e.g.

user:123:profile), a value (serialized; watch size), and a TTL. Decide

the invalidation event per key family (on write? on TTL? on version bump?).

Versioned keys (...:v2) make invalidation a write of a new key instead of a

delete race.

Choosing a provider

Default to the generic recipe above (Redis or Memcached, self-hosted or managed).

If the user names a cloud, read references/providers/<provider>.md for the

managed-service mapping, limits, and provider-specific trade-offs. If no file

exists for that provider, the generic recipe is the answer.

Diagram

To visualize the cache-aside read path (app → cache → miss → origin → populate)

or the stampede/fallback flow, use the in-plugin architecture-diagram skill —

cache nodes use the cache color, the origin its store color, and the miss path a

dashed arrow.

Related building blocks

• content-delivery — *pairs with* this as the edge layer above it; CDN/edge caching for static and media (that concept lives there).
• data-storage — *depends on* this as the origin a cache protects; its read replicas are an *alternative to* caching reads, and cache-node sharding mirrors its partitioning. (Consistent-hashing theory is *owned by* consistency-coordination.)
• consistency-coordination — *pairs with* this when staleness is unacceptable and read-your-writes or stronger guarantees are required.
• back-of-the-envelope — *feeds into* this skill: it supplies the read ratio and hot-set size that justify a cache.
• resilience-failure — *pairs with* this; rate limiting and circuit breakers (*owned* there) help contain the retry storms a misbehaving cache can trigger.
• system-design — *owned-concept lives in* the orchestrator: the reasoning loop, the trade-off method, and the ten failure modes.

References

• references/deep-dive.md — invalidation patterns, single-flight/coalescing, consistent hashing for cache sharding, Redis vs Memcached internals, local (L1) + distributed (L2) tiers. Read when designing the cache layer in detail.
• references/providers/{generic,aws,azure,gcp}.md — service mappings, limits, and pitfalls per environment.