Our Cache Made the App Slower. The Redis Mistake I’ll Never Repeat

Caching is supposed to be boring. You add it, things get faster, and everyone moves on. That’s the myth.

In our case, adding Redis did improve performance at first. API response times dropped sharply. The team celebrated. A few weeks later, complaints started coming in. Users reported outdated data. Some pages felt slower than before. Support escalations increased.

Nothing was “broken.” That’s what made it dangerous.

Why This Problem Didn’t Show Up Immediately

The cache worked perfectly in isolation. The issue only appeared once real usage patterns emerged.

Traffic was uneven. Some endpoints were hit constantly. Others were hit sporadically. Writes happened less often than reads, but when they did, they mattered.

The system assumed cached data would remain valid long enough to be useful. That assumption was wrong.

The Original Caching Logic

The backend cached API responses aggressively. The goal was to reduce database load and speed up list endpoints.

Copyconst cached = await redis.get(key);
if (cached) return JSON.parse(cached);

const data = await fetchFromDB();
await redis.set(key, JSON.stringify(data), "EX", 3600);
return data;

On paper, this looked fine. One hour TTL. Simple logic. No complexity.

In production, this created a slow-moving disaster.

Bug Breakdown – Stale Data Is Still Data

The real issue wasn’t correctness alone. It was trust.

Users saw values that didn’t reflect recent actions. They refreshed repeatedly. That caused cache misses and extra load. Some requests bypassed cache due to key mismatches. Others hit outdated entries.

The system oscillated between fast and slow. Predictability vanished.

Caching doesn’t just affect speed. It affects behavior.

The Root Cause We Missed

The cache key design ignored context.

The same key was used for:

• Different filter combinations

• Different user scopes

• Slightly different query shapes

This caused cache collisions and unnecessary invalidations. Worse, updates didn’t invalidate related keys at all.

The cache wasn’t wrong. It was naïve.

The Fix Was Not “Lower the TTL”

Lowering TTL was the first instinct. It helped briefly. Then load increased again.

The real fix required three changes.

First, cache keys had to reflect access patterns, not endpoints.

Second, invalidation had to be explicit, not time-based guessing.

Third, only read-heavy, low-volatility data deserved caching.

Refactoring the Cache Strategy

Instead of caching entire responses, we cached normalized slices.

Copyconst key = `company:${companyId}:orders:page:${page}`;

On writes, we invalidated only the affected namespace.

Copyawait redis.del(`company:${companyId}:orders:*`);

This reduced blast radius and restored predictability.

Another Example:

Anti-Pattern

Copyredis.set(key, data, "EX", 3600);

Production-Grade Pattern

Copyredis.set(key, data);
invalidateRelatedKeysOnWrite();

Plain explanation

Time-based expiration guesses. Explicit invalidation designs.

Why We Didn’t Cache Everything

We consciously chose not to cache:

• Rapidly changing dashboards

• User-specific transactional data

• Anything tied to real-time decisions

Caching everything increases complexity faster than it increases speed.

This decision aged well as traffic grew.

Cache Invalidation vs Freshness

Perfect cache invalidation doesn’t exist. You choose where to be wrong.

We accepted:

• Slightly higher read latency on some endpoints

• Lower cache hit rate overall

In exchange, we gained:

• Data correctness

• Stable performance

• Reduced support issues

Speed without trust is useless.

How This Changed Our Backend Discipline

After this incident:

• Every cache required an invalidation plan

• Cache keys were documented like APIs

• Performance tests included stale-data scenarios

Caching stopped being a shortcut and became part of system design.

Business Impact Nobody Expected

Once stale data disappeared:

• Refresh storms stopped

• Backend load stabilized

• Support tickets dropped noticeably

• User confidence returned

Performance issues are often behavioral, not technical.

Final Takeaway

Caching is not an optimization layer you sprinkle on top. It is a data consistency contract.

If you can’t explain how and when cached data becomes invalid, you are not optimizing. You are deferring a production incident.