caching-strategy-setup

Caching Strategy Setup

Add caching at the right layer with the parts people forget: sane keys, expiration policy, stampede protection, and an invalidation plan.

Input — point it at your project

Works on a target: a file, folder, project, or GitHub URL. Find hot reads to cache and existing cache usage: grep -rn "IMemoryCache\|IDistributedCache\|HybridCache\|GetOrCreate" --include=*.cs <target>.

Pick the layer

• IMemoryCache — single instance, fastest, not shared across nodes. Good for per-node hot data.
• IDistributedCache (Redis) — shared across instances, survives restarts, network hop. Good for multi-node.
• HybridCache (.NET 9+) — combines L1 (memory) + L2 (distributed) with built-in stampede protection; prefer it when available.
• Output caching — cache whole responses at the endpoint when appropriate.

HybridCache (recommended where available)

builder.Services.AddHybridCache();

public class ProductService(HybridCache cache, IProductRepo repo)
{
    public ValueTask<Product> GetAsync(int id, CancellationToken ct) =>
        cache.GetOrCreateAsync($"product:{id}",
            factory: async c => await repo.GetAsync(id, c),
            options: new() { Expiration = TimeSpan.FromMinutes(10), LocalCacheExpiration = TimeSpan.FromMinutes(2) },
            cancellationToken: ct);
}

The parts people forget

• Stampede / cache miss storm — without protection, many requests recompute on expiry. HybridCache handles it; with IMemoryCache add a per-key lock (SemaphoreSlim) or LazyCache.
• Keys — stable, namespaced, include all inputs that change the result (product:{id}:{culture}).
• Expiration — absolute + sliding deliberately chosen; don't cache forever.
• Invalidation — the hard part. On write, evict/refresh affected keys (cache.RemoveAsync). Tag-based eviction for groups.
• Don't cache user-specific/secret data in a shared cache without scoping the key.

Principles

• Cache to cut expensive, repeated, tolerably-stale reads — not everything.
• Every cache entry needs an expiration and an invalidation story before it ships.
• Measure hit rate; a low-hit cache is just overhead.