Last modified: April 27, 2026

This article is written in: 🇺🇸

Application-Level Caching

Application-level caching stores computed results or frequently accessed objects directly inside the running process or in a dedicated in-process store. Because data never leaves the application's memory space, reads are limited only by CPU and memory bandwidth — no network hop, no serialisation, and no socket overhead. This makes application-level caches the fastest possible cache tier, complementing distributed caches like Redis or Memcached that serve data across multiple instances.

Application-Level vs Distributed Cache

  Single Process                       Distributed Setup
  +--------------------------+         +----------------------------------+
  |  Application Instance    |         | App1   App2   App3   App4       |
  |                          |         | +---+  +---+  +---+  +---+      |
  |  +--------------------+  |         | |L1 |  |L1 |  |L1 |  |L1 |     |
  |  |  In-Process Cache  |  |         | +--++  +--++  +-+-+  +-+-+     |
  |  | (heap / off-heap)  |  |         |    |      |     |       |       |
  |  +--------+-----------+  |         |    +--+---+--+--+       |       |
  |           |              |         |       |         |       |       |
  |           v              |         |  +----+----+ +--+----+  |       |
  |  +--------------------+  |         |  | Redis / | | Redis / | |      |
  |  |  Database / Source |  |         |  |Memcached | |Memcached| |      |
  |  +--------------------+  |         |  +---------+ +---------+  |     |
  +--------------------------+         +----------------------------------+

  In-process cache = L1 (ultra-fast, node-local)
  Distributed cache = L2 (fast, cluster-wide)
  Database = source of truth

• An in-process cache is faster than any network-based store because it requires no serialisation or I/O to serve a cached entry.
• Application-level caches are bounded by the JVM heap or native memory limit of a single process, unlike distributed caches that scale across many nodes.
• When multiple application instances run behind a load balancer, each instance holds its own independent cache, meaning cache entries are not shared across nodes without a distributed layer.
• A local cache combined with a distributed cache forms a two-tier hierarchy that minimises latency for hot keys while still maintaining cluster-wide consistency.
• Cache population is event-driven in most frameworks, triggered automatically on a cache miss or explicitly pre-loaded at startup.

Common In-Process Cache Libraries

Caffeine (Java)

Caffeine is a high-performance, near-optimal Java caching library based on the W-TinyLFU eviction algorithm.

LoadingCache<String, Product> cache = Caffeine.newBuilder()
    .maximumSize(10_000)
    .expireAfterWrite(Duration.ofMinutes(5))
    .refreshAfterWrite(Duration.ofMinutes(1))
    .recordStats()
    .build(key -> productRepository.findById(key));

Product p = cache.get("product:42");  // auto-loads on miss

• The W-TinyLFU policy is optimal for workloads with skewed access patterns because it evicts items with both low frequency and low recency.
• expireAfterWrite resets the TTL after each write, ensuring stale entries are eventually removed even if they remain popular.
• refreshAfterWrite triggers a background reload after the specified interval while still serving the old value until the refresh completes.
• maximumSize bounds heap usage, and the eviction algorithm automatically selects the lowest-value entries when the limit is reached.
• Caffeine's recordStats() method exposes metrics including hit rate, load time, and eviction count via a CacheStats object.

Guava Cache (Java)

Guava Cache is the predecessor to Caffeine and is still widely used in the Java ecosystem.

LoadingCache<String, User> userCache = CacheBuilder.newBuilder()
    .maximumSize(5000)
    .expireAfterAccess(10, TimeUnit.MINUTES)
    .weakValues()
    .build(key -> userDao.load(key));

• expireAfterAccess removes entries that have not been read within the specified duration, which is useful for session-like data.
• weakValues() tells Guava to hold values via weak references, allowing the garbage collector to reclaim memory under pressure.
• Guava Cache is thread-safe by default and segments its internal structure to reduce contention under concurrent reads and writes.

Ehcache (Java)

Ehcache is a full-featured cache framework widely used in enterprise Java and Spring applications.

<!-- ehcache.xml -->
<cache name="products"
       maxEntriesLocalHeap="10000"
       timeToLiveSeconds="300"
       timeToIdleSeconds="120"
       overflowToDisk="true">
</cache>

• Ehcache supports both heap and off-heap storage, allowing very large caches without adding GC pressure.
• Off-heap storage serialises entries to native memory that is not managed by the JVM garbage collector.
• Ehcache integrates with JCache (JSR-107), enabling portable cache code that can be swapped between providers.
• Terracotta clustering adds a distributed tier to Ehcache for sharing entries across multiple nodes without a separate Redis deployment.

Python: functools.lru_cache and cachetools

from functools import lru_cache
from cachetools import TTLCache, cached

@lru_cache(maxsize=512)
def get_config(key: str) -> str:
    return db.fetch(key)

_ttl_cache: TTLCache = TTLCache(maxsize=1000, ttl=300)

@cached(cache=_ttl_cache)
def get_user(user_id: int) -> dict:
    return user_service.fetch(user_id)

• lru_cache uses Python's built-in LRU eviction and is suitable for functions whose inputs have a small, bounded cardinality.
• cachetools.TTLCache adds time-based expiry on top of LRU eviction, removing entries that are older than the configured TTL.
• Neither cache is thread-safe by default under concurrent writes, so a lock wrapper is needed in multithreaded applications.

Node.js: node-cache and lru-cache

import LRU from 'lru-cache';

const cache = new LRU({
  max: 500,
  ttl: 1000 * 60 * 5,  // 5 minutes
});

function getProduct(id) {
  const hit = cache.get(id);
  if (hit) return hit;
  const product = db.query(id);
  cache.set(id, product);
  return product;
}

• lru-cache is a widely-used npm package providing an LRU cache with optional TTL support.
• Because Node.js runs on a single event-loop thread, in-process caches require no locking for concurrent access within the same process.
• Multiple Node.js worker processes (via cluster or PM2) each have their own isolated cache, making a distributed layer necessary for cross-process consistency.

Go: sync.Map and groupcache

var cache sync.Map

func GetUser(id string) *User {
    if val, ok := cache.Load(id); ok {
        return val.(*User)
    }
    user := db.FetchUser(id)
    cache.Store(id, user)
    return user
}

• sync.Map provides concurrent read and write access without explicit locking, optimised for the case where entries are written once and read many times.
• groupcache (from the Go team) fills the gap between a simple map and Redis by providing consistent hashing, request coalescing, and hot-key replication across a small cluster of app servers.
• Libraries like ristretto from Dgraph offer a more sophisticated eviction algorithm (TinyLFU) and metrics collection for production Go services.

Cache Patterns at the Application Layer

Cache-Aside

Application code manages loading:

  func Get(key) -> value:
    val = cache.get(key)
    if val == nil:
      val = db.fetch(key)
      cache.set(key, val, ttl)
    return val

• The calling code is responsible for populating the cache on a miss, giving it full control over what is stored and for how long.
• This pattern is lazy: data is only loaded into the cache when it is first requested, keeping memory usage proportional to actual demand.

Read-Through

• The cache itself is responsible for fetching from the backing store on a miss, using a loader function configured at setup time.
• Read-through simplifies calling code because it always reads from a single interface regardless of whether the data is cached.
• Caffeine.build(loader) and Spring's @Cacheable annotation both implement the read-through pattern transparently.

Write-Through

• Every write goes to both the cache and the backing store synchronously before the operation is acknowledged.
• Write-through keeps the cache consistent with the database at the cost of higher write latency.
• Spring's @CachePut annotation implements write-through by updating the cache entry alongside the database call.

Write-Behind (Write-Back)

• The application writes to the cache immediately and defers persistence to the backing store asynchronously.
• Write-behind improves write throughput by batching multiple updates into fewer database calls.
• The risk of data loss is higher because any crash between the cache write and the deferred flush can result in lost updates.

Refresh-Ahead

• The cache pre-emptively refreshes an entry in the background before it expires, ensuring reads never block on a slow loader.
• Caffeine's refreshAfterWrite implements this pattern by triggering a background reload when a read occurs after the refresh threshold has passed.
• Refresh-ahead is most beneficial for entries with predictable, high-frequency access where expiry-induced latency spikes are unacceptable.

Invalidation Strategies

Invalidation Approaches

  Time-based (TTL)         Event-based              Version-based
  +---------------+        +---------------+         +--------------+
  | Entry expires |        | DB update     |         | Key includes |
  | after N secs  |        | triggers a    |         | version or   |
  | automatically |        | cache.delete  |         | hash suffix  |
  +---------------+        +---------------+         +--------------+
  Simple, no coupling    Precise, event-driven     No invalidation needed;
  Risk: stale data       Complexity: messaging     old keys expire naturally

• Time-to-live invalidation is simple to implement but may serve stale data for the entire TTL interval after an update.
• Event-driven invalidation reacts to domain events (e.g., "product updated") and removes or refreshes the affected key immediately.
• Version-based keys embed a hash of the data version into the cache key so a data change produces a new key and the old entry is naturally displaced.
• Spring's @CacheEvict annotation removes specific cache entries by key or clears the entire cache when a method modifies underlying data.

Two-Tier (L1 + L2) Caching

Many production architectures combine a local in-process cache (L1) with a distributed cache (L2) for maximum efficiency.

Two-Tier Cache Read Flow

  App Instance
  +--------------------------------------------+
  |  1. Check L1 (Caffeine / lru-cache)         |
  |         |                                   |
  |    HIT  +  MISS                             |
  |     |         |                             |
  |     v         v                             |
  |  Return    2. Check L2 (Redis / Memcached)  |
  |  value          |                           |
  |            HIT  +  MISS                     |
  |             |         |                     |
  |             v         v                     |
  |     Populate L1   3. Fetch from DB          |
  |     Return val        |                     |
  |                  Populate L2 + L1           |
  |                  Return value               |
  +--------------------------------------------+

• L1 serves the hottest keys with zero network latency; L2 catches misses that fall through from L1.
• L1 invalidation across nodes is challenging because a write on one instance does not automatically update the in-process caches of other instances.
• Pub/Sub invalidation broadcasts a delete message over Redis or a message queue so all instances evict their local copy when the source data changes.
• Short L1 TTLs (a few seconds) combined with longer L2 TTLs achieve a balance between staleness risk and latency reduction.
• Near-cache patterns keep a replica of a remote distributed cache entry locally, refreshed on demand, and are used by frameworks like Hazelcast and Apache Ignite.

Sizing and Capacity Planning

• The target hit rate for an in-process cache is typically 95% or higher; lower rates suggest the cache is undersized or the key space is too large.
• Access frequency histograms should guide sizing: if the top 1 % of keys account for 80 % of requests, a small cache serving those keys provides most of the benefit.
• Heap impact can be estimated by multiplying the average serialised entry size by the maximum number of entries.
• Off-heap caches in Ehcache or Chronicle Map allow large caches (tens of GB) without triggering long GC pauses.
• Monitoring eviction rate trends over time is important to detect a growing working set that is outpacing the configured cache size.

Observability

• Hit rate, miss rate, load time, and eviction count are the core metrics every application cache should expose.
• Caffeine's CacheStats object provides all four standard metrics and integrates with Micrometer for Prometheus or Datadog export.
• Tracing cache operations with OpenTelemetry spans makes it visible how much time each request spends in cache lookup versus database I/O.
• Setting up alerts on a sudden drop in hit rate is valuable for catching cache misconfigurations or deployment-induced invalidation storms.
• Logging slow cache loads (loader duration > threshold) helps identify backing queries that need their own optimisation independent of the cache.