FAQ
Short answers about lazy-layers-cache: Redis, stampedes, invalidation across instances, Cloudflare Workers, failures, eviction, memory, speed, metrics and upgrading to 8.0.
Short answers to the questions people ask most about lazy-layers-cache 8.0.0. Every number comes from BENCHMARKS.md, which says how each was measured. For a side-by-side look at other libraries, see the comparison.
Questions
What is lazy-layers-cache?
An open-source L1/L2 cache for Node.js and Cloudflare Workers: an in-process L1 in front of a shared L2 (Redis, Cloudflare KV, Durable Objects or MongoDB), read through with getOrSet, with stampede protection inside one process and across processes, and invalidation over an event bus. It is MIT-licensed, written in TypeScript, and has one runtime dependency, msgpackr.
Do I need Redis?
No. In one process, setupCache({ namespace, redis: false }) runs an in-memory L1 only. For several instances, use Redis as L2 with Redis Pub/Sub, Cloudflare KV, or lazy-layers-cache/coordination, which keeps locks, counters, rate limits and invalidation in Cloudflare Durable Objects or MongoDB.
How does it prevent cache stampedes?
Concurrent getOrSet callers for the same missing key share one loader call in the process, and with Redis L2 a per-key lease lets one instance load while the others wait to be woken. In BENCHMARKS.md, 10,000 concurrent callers on one cold key ran the loader once.
How do other instances learn that a value changed?
Over an event bus: Redis Pub/Sub, RabbitMQ, NATS Core, NATS JetStream, Durable Object RPC or MongoDB change streams. invalidate, invalidateByPattern and invalidateTags publish a delete that every peer applies to its own L1, and a freshly loaded value can be broadcast so peers do not query the database for it. setupCache caps that broadcast at 32 KiB.
Does it work on Cloudflare Workers?
Yes. lazy-layers-cache/cloudflare is a Worker-safe entry point with a Cloudflare KV store, and lazy-layers-cache/coordination adds Durable Object locks, counters, rate limits, a shared cache tier and invalidation between Worker instances.
What happens when Redis or the event bus fails?
The cache fails open: an L2 error becomes a miss and the loader runs, circuit breakers stop repeated calls to the failing dependency, and fail-safe serves a recent value when the loader fails. With managed Redis Pub/Sub, a subscription gap makes local copies untrusted, so the cache bypasses L1 until recovery restores trust.
Which eviction policy does it use?
S3-FIFO by default, with LRU, SIEVE, FIFO and custom policies available. On the synthetic traces in BENCHMARKS.md it had the lowest mean miss ratio of the built-in policies, and on Zipf α = 1.0 traffic the in-memory store hit 78.23% with it, against 73.59% for lru-cache.
How much memory does it use?
L1 is limited by maxEntries (1,000 by default) and by maxMemory, a byte budget shared by every built-in store in the process: 20% of effective memory by default, never more than 75%. It gives memory back when the V8 heap passes 85% of its limit or RSS passes 90% of maxProcessMemory, which defaults to the container or host memory. A small encoded entry retained 76.1 bytes with LRU and 98.4 with S3-FIFO in BENCHMARKS.md.
How fast is it?
In BENCHMARKS.md a warm read-through hit costs 180 ns in reference mode and 1,610 ns in encoded mode, the default, against 6,490 ns in 0.7.0, and a miss costs 2 Redis round trips. Inserting into a full in-memory cache is slower than in lru-cache or quick-lru, because every write measures the value and charges a memory budget.
Which Node.js versions and module formats are supported?
Node.js 20 or 22+; use 22+ with NATS, because its current transitive dependency declares that minimum. The package ships ESM, CommonJS and TypeScript declarations.
Does it expose metrics?
Yes. With the observability option it serves Prometheus metrics under the lazycache prefix and a live dashboard protected by a generated token. Every cache operation is also an event you can subscribe to with cache.on().
How do I upgrade from 0.7 to 8.0?
Wire formats are unchanged, so 0.7 and 8.0 instances can share Redis and an event bus during a rolling deploy. Install the integrations you use (ioredis, amqplib, the NATS packages, lz4-napi and snappy), which are now optional peer dependencies, configure dashboard credentials if you relied on the old defaults, and read MIGRATION.md for the rest.
Compared with other caches
How lazy-layers-cache compares with bentocache, cache-manager, cacheable, cachified, async-cache-dedupe, layered-loader and lru-cache: features from each README, and measured results.
Serverful
Traditional servers, Docker, and Kubernetes where the process persists long enough for L1 to warm across requests.