Caching Strategies

Caching Strategies

Purpose

A cache can reduce the arrival rate seen by an origin in L = λ × W: a hit consumes no origin connection, planner or I/O. Batching, admission control and eliminating work can also reduce origin demand, so caching is one option rather than a unique law. A stale, unbounded cache can show excellent hit rate; correctness, memory and origin protection must be measured beside it.

Workflow

Inspect the target's Maven/Gradle release/toolchain, resolved Caffeine/Spring Data/Jackson versions, runtime image and cache configuration before choosing APIs. No universal Java baseline is declared here; the configuration reference states its example baseline. Preserve project versions and do not enable preview features or upgrade dependencies to fit an example. If workload, freshness requirements or measurements are absent, identify the gap and offer a conditional decision and measurement plan rather than inventing a hit rate or safe TTL.

1. Measure source cost and capacity (latency distribution, CPU/I/O and rate) before deciding. Even a sub-millisecond lookup may matter at very high volume; latency alone is not the case.
2. Measure the access distribution and estimate h for the intended maximumSize.
3. Model saved work and latency, not hit rate alone. Estimate origin work avoided by hit distribution and compare h·T_hit + (1-h)·T_miss (including queueing/load cost) with the uncached distribution. Tail latency cannot be derived from averages.
4. Bound it—by count or a measured weight proxy. Account for keys, values, node metadata, allocator/GC headroom and concurrent load buffers; a weigher's logical bytes are not measured heap retention. Validate with heap/allocation evidence under representative occupancy.
5. Set a TTL from the business tolerance for stale data, and add jitter if entries are created in bulk. Keep the longest jittered lifetime inside that tolerance, accounting for source lag and load time. Access-based expiry does not bound the age of frequently read data.
6. Define the invalidation strategy and write an automated test for it — propagation is the part that silently stops working.
7. Instrument outcomes: request-weighted and byte-weighted hit/miss, origin rate and load latency/failures, eviction/admission, retained memory, stale-age/version and invalidation lag.

Rules

• For a simple cache-aside path, E[T] ≈ h·T_hit + (1-h)·T_miss; miss cost includes cache lookup, origin queueing/load and fill. Increasing hit rate has linear average benefit only if those distributions stay fixed; near saturation, queueing can make the system nonlinear. Hit-rate gain per byte depends on the observed popularity/size distribution, not a universal logarithm.
• A broken cache has better metrics than a correct one. No limit, no TTL and no invalidation gives the best possible hit rate. This is why hit rate never travels alone.
• Avoid putting managed/mutable JPA entities in an application cache—the cache may retain aliases, lazy proxies and persistence-context assumptions. Cache immutable projections/value snapshots with an explicit version. A provider's second-level cache is a separate coordinated mechanism, not evidence that arbitrary entity references are safe.
• In Spring's default proxy mode, @Cacheable self-invocation via this bypasses interception. AspectJ mode or direct programmatic caching differs. Test the deployed mode; extracting a collaborator is often clearer than self-injection.
• Never cache an operation with a side effect. @Cacheable on something that creates means that on a hit the thing is not created and the cache asserts that it was. For idempotency that must survive eviction, restart and retries, use a durable record with an atomic relationship to the effect; a unique key alone does not supply that relationship. Delegate the operation contract to idempotency.
• Stampede has several scopes: jitter desynchronizes bulk expiry; singleflight/LoadingCache coalesces per key only within its process/cache instance unless backed by distributed coordination; refreshAfterWrite serves an old value while a hot-key refresh runs; staged warm-up avoids a global cold cache. Probabilistic early expiration reduces the spike, it does not remove it — and the correct form is P = exp(−(expiry − now) / (β · δ)), with β in the denominator and δ representing measured recomputation duration. Validate the algorithm and clock/units rather than copying the equation without its assumptions: require positive β and δ, consistent time units, and treat already expired entries as misses rather than probabilities greater than one.
• FLUSHALL in a deploy pipeline is a stampede generator. If the service needs the cache to serve its load, the cache is an availability component, not a performance one. For a format change, version the key prefix, but stage and rate-limit warming: switching every caller to an empty namespace is also a cold-cache event. Budget old/new namespaces together.
• Spring Data Redis defaults RedisTemplate/RedisCache to JDK serialization in current documentation; override it explicitly. Prefer a typed schema/serializer. In Spring Data Redis 4, Jackson 3 uses JacksonJsonRedisSerializer<T> or GenericJacksonJsonRedisSerializer; Jackson-2-named serializers are deprecated, and the old generic serializer enabled default typing by default. Do not solve lost type information by enabling payload-selected classes (java-serialization-hardening).
• Redis pub/sub is fire-and-forget. An L1 TTL limits local retention, not end-to-end staleness: expiry may refill from an already stale L2 or origin replica. Derive the age budget across layers, propagate versions/remaining freshness, or reload from a sufficiently fresh source. Publish only after a successful commit, but recognize that an AFTER_COMMIT listener can crash before publishing. Use an outbox/CDC or version-checked reads where bounded reliable invalidation is required.
• Cache-aside has races: an old slow read can fill after a newer write invalidates, resurrecting stale data. A version field alone does not reject the old fill. Define the check against the authoritative version or invalidation watermark, including absent entries and deletes. Write-through/CDC still need ordering against concurrent fills; use a tolerated stale window only when the consistency requirement permits it.
• A key is an authorization boundary. Include tenant, locale, entitlement/principal dimensions that affect the result; canonicalize them; never let one tenant reuse another's cached response. Avoid secrets/PII in keys because keys appear in metrics, logs and admin tools.
• Negative caching protects against penetration only with a short bounded TTL and input/cardinality controls. Caching every attacker-chosen miss is itself an unbounded-memory attack.

Deliverable

For a design/review, return the cache/no-cache decision, measured inputs and assumptions, key/value contract, memory/freshness bounds, invalidation race handling and origin-outage policy. State the targeted tests and acceptance bounds. For an incident, report evidence, competing hypotheses and the next discriminating measurement; do not label an untested hypothesis a confirmed fix.

Primary sources

• Spring Data Redis object mapping and serializers
• Spring Data Redis 4 migration guide
• Caffeine refresh semantics
• Caffeine 3.2.2 API
• Redis Pub/Sub delivery
• Redis key eviction

References

• Configuring a cache — Caffeine and Spring configuration with bounds, weight, jitter and stats; the Redis settings that matter; and the near-cache (L1+L2) rules. Read when implementing or reviewing a cache.
• Cache incident triage — the symptom-to-cause table and the metric set that makes each cause visible. Read when a cache-related incident is in progress.