Development
The Sticky Note Strategy: How We Use Redis to Make Our API Blazing Fast
A practical walkthrough of the Cache-Aside pattern — what it is, how to implement it in NestJS, and the tradeoffs nobody tells you about upfront.
The Problem with Going to the Filing Cabinet Every Time
Imagine you work in a busy office. Every time someone asks you “What’s on today’s lunch menu?” you walk to the basement, unlock the filing cabinet, flip through hundreds of folders, find the answer, walk back upstairs — and give it to them.
Now imagine 500 people asking the same question, one after another.
That’s exactly what happens when every API request hits the database directly. The database is powerful, but it lives in the basement. Every trip takes time.
The real numbers
A typical PostgreSQL query with joins takes 20–200ms. A Redis read takes under 1ms. For a popular endpoint hit thousands of times per minute, that gap compounds into real infrastructure cost and user-perceived slowness.
We needed a sticky note on the desk.
Enter Redis — The Sticky Note on Your Desk
Redis is an in-memory data store. Unlike a database that reads from disk, Redis keeps everything in RAM — the fastest storage a computer has.
If the database is a filing cabinet in the basement, Redis is a sticky note pinned right on your monitor.
🗄️Filing Cabinet (Database)
Accurate, complete, durable — but slow to reach. Every query walks to the basement and back.
📝Sticky Note (Redis)
Limited space, not permanent — but instantly readable. The answer is right there on your desk.
When someone asks for data, we check the sticky note first. If it has the answer — done. If it doesn’t — we walk to the filing cabinet, get the answer, and write it on a new sticky note for next time.
This pattern is called Cache-Aside, and it is the backbone of how our system works.
Cache-Aside: The Pattern in Action
Cache-Aside (also called lazy loading) has a simple decision tree on every read request:
Here’s what this looks like in a NestJS service, using ioredis:
async getUserProfile(userId: string): Promise<UserProfile> {
const cacheKey = `user:profile:{userId}`;
// 1. check the sticky note
const cached = await this.redis.get(cacheKey);
if (cached) {
return JSON.parse(cached); // cache HIT - instant
}
// 2. walk to the filing cabinet
const user = await this.prisma.user.findUnique({
where: { id: userId },
include: { profile: true, preferences: true },
});
if (!user) throw new NotFoundException();
// 3. write a new sticky note (TTL = 5 minutes)
await this.redis.setex(cacheKey, 300, JSON.stringify(user));
return user;
}
Setting an Expiry — When the Sticky Note Goes Stale
A sticky note from last Tuesday about today’s lunch menu is worse than useless — it’s wrong. Cached data has the same problem. The fix is a TTL (Time-To-Live): Redis automatically deletes the key after a set number of seconds.
Choosing the right TTL is a judgment call, not a formula. Ask: how bad is it if this data is stale?
⚠ The TTL trap
Setting TTL too high means users see stale data long after it changed. Setting it too low means your cache hit rate collapses — you’re back to hammering the database. Start conservative (shorter), measure your cache hit ratio, and extend TTLs for data where staleness hasn’t caused problems.
What We Cache and What We Don’t
Not every piece of data belongs on the sticky note. After running this in production, here’s the rough split:
- Good cache candidates: shared data read by many users (public profiles, catalog items, config values, computed aggregates). High read frequency, low write frequency.
- Bad cache candidates: user-specific transactional data (cart contents, payment status, unread count). Stale state here causes real bugs.
- Never cache: authentication tokens (they have their own lifecycle), passwords, anything where a stale read has security implications.
🗒️The Sticky Note Rule
You’d write “Today’s lunch menu” on a sticky note. You wouldn’t write “Current bank balance” on one. Same principle.
Cache Invalidation — The Hard Part
There are only two hard things in computer science: naming things and cache invalidation. The joke is old because the problem is real.
When a user updates their profile, the cached version is now wrong. You have two options:
Option A: Delete on Write (Active Invalidation)
When the user writes new data, delete the corresponding cache key immediately. The next read will miss the cache, hit the DB, and repopulate with fresh data.
async updateUserProfile(userId: string, dto: UpdateProfileDto) {
const updated = await this.prisma.user.update({
where: { id: userId },
data: dto,
});
// tear the sticky note off the board
await this.redis.del(`user:profile:{userId}`);
return updated;
}
Option B: Let TTL Handle It (Passive Expiry)
Do nothing on write. The cache key will expire naturally after its TTL. Until then, some users see stale data. This is simpler but only works when eventual consistency is acceptable.
In practice, we use both: active deletion for data the user just changed (they expect to see their own update immediately), TTL expiry for everything else.
Structuring Redis Keys
Redis has no schema, no tables, no enforced structure. The key is literally just a string. This is powerful and dangerous — without a naming convention, a large codebase becomes impossible to reason about.
We use a colon-delimited namespace pattern:
// {resource}:{identifier}:{optional-variant}
user:profile:42 // user profile, id=42
post:detail:slug-abc // post by slug
post:list:page:1 // paginated list, page 1
leaderboard:weekly // weekly leaderboard aggregate
config:feature-flags // app-wide feature flags
Consistent naming lets you do pattern-based operations. For example, when a post is updated you can delete all cached variants: DEL post:detail:slug-abc and any list pages that might include it.
What This Looks Like in Production
After rolling out Cache-Aside on our most-hit endpoints, here’s what changed:
💡 The 84% hit rate
84% of reads never touched the database at all. That’s 84% of your queries answered in under 1ms instead of 180ms. The remaining 16% — cache misses and write paths — hit the DB as normal. The database got dramatically easier to run as a result.
What I’d Do Differently
Things I learned the hard way:
- Serialize carefully. We had a bug where
Dateobjects came back from Redis as strings. JSON doesn’t have a date type — always deserialize and validate the shape coming out of cache, not just going in. - Handle Redis being down. Redis is fast and reliable, but it can go down. Your cache layer should fail open: if Redis throws an error, log it and fall through to the database rather than crashing the whole request.
- Watch for cache stampede. If a popular key expires and 500 requests arrive simultaneously, they all miss the cache and all hit the DB at once. A mutex or probabilistic early revalidation prevents this.
- Don’t cache errors. If the database returns an error (user not found, timeout), don’t cache that result. The error might be transient — caching it means every subsequent request gets a wrong answer for the full TTL.
The main lesson: Redis doesn’t make your database faster. It makes your application ask the database less often. That distinction matters — it means Redis is a layer on top of your existing data layer, not a replacement for it. The database stays the source of truth; Redis is just the very fast first stop on the way there.
Once you internalize that, the whole design becomes obvious: keep the sticky note in sync with the filing cabinet, and your office runs like it has ten employees instead of two.
Caching is one of those things that feels like premature optimization until the day you actually need it — and then it feels like the most obvious thing in the world. The Cache-Aside pattern is a good default starting point: lazy, simple, composable, and easy to reason about even months later when you’ve forgotten why you wrote the code.
If you’re seeing slow API responses and your database query times look fine, the answer is probably a cache layer. Start small — one endpoint, one key pattern — and measure before expanding.
