lokesh_karthikeyan

[lokesh_karthikeyan @_itslokesh_]

Caching Can Accidentally Destroy Your Database

Caching is supposed to reduce database load. But sometimes it does the opposite. This happens during a cache stampede.
🧵

#systemdesign #caching #software #engineering

[lokesh_karthikeyan @_itslokesh_]

The lesson
At scale, intuition isn’t enough.
Visibility into usage patterns drives the biggest savings.

Your cloud bill is an engineering problem.
Treat it like one.

[lokesh_karthikeyan @_itslokesh_]

Canva’s $3.6M Cloud Cost Lesson
Scaling to 100M+ users creates a new problem.
Cloud costs.

Canva stores more than 230 petabytes of data on Amazon S3.
Most of it is user-generated designs.

#CloudComputing #AWS #CostOptimization #canva

[lokesh_karthikeyan @_itslokesh_]

They identified:
• infrequently accessed data
• large objects
• buckets with strong savings potential
Only those datasets were migrated

The results
• 80B objects migrated in ~2 days
• 130+ PB moved to Glacier Instant Retrieval
• $300K saved per month
• $3.6M saved annuall

[lokesh_karthikeyan @_itslokesh_]

Large systems often use multiple caching layers together. Understanding where to cache is just as important as knowing how to cache.

[lokesh_karthikeyan @_itslokesh_]

When engineers hear “caching,” they usually think of Redis. But caching exists in multiple layers of a system.

In fact, many systems already use caching without developers realizing it. Here are some common caching layers.
🧵

#SystemDesign #web #software #engineering #backend

[lokesh_karthikeyan @_itslokesh_]

Each caching layer solves a different problem.
Browser cache → avoids unnecessary downloads
CDN → reduces global network latency
Redis → reduces database load
Database cache → speeds internal queries

[lokesh_karthikeyan @_itslokesh_]

In practice, engineers often mix the terminology.

What matters most is whether your data lives on one machine or many.

#databases #partitioning #sharding

[lokesh_karthikeyan @_itslokesh_]

Partitioning and Sharding Are Not the Same Thing (Even Though People Use Them That Way)
Engineers often use partitioning and sharding interchangeably.
But technically, they are different.
Understanding the difference matters when designing systems.

#systemdesign #interview

[lokesh_karthikeyan @_itslokesh_]

Example:
Shard 1 → users 1–1M
Shard 2 → users 1M–2M
Shard 3 → users 2M–3M

The key difference:
Partitioning → one database machine
Sharding → many database machines

Partitioning improves query efficiency.
Sharding improves scalability.

[lokesh_karthikeyan @_itslokesh_]

You should explain:
• why the database is a bottleneck
• what data should be cached
• how cache updates happen

Caching is powerful, but only when the problem actually requires it.

[lokesh_karthikeyan @_itslokesh_]

Most Engineers Introduce Caching Too Early in System Design Interviews
Many candidates jump straight to “let’s add Redis.”

But good system design answers don’t start with caching.
They start with identifying the actual bottleneck.
🧵

#systemdesign #interview #caching #software

[lokesh_karthikeyan @_itslokesh_]

most reads are served from memory. That allows the database to focus on writes and less frequent reads.

However, caching introduces new problems:
• stale data
• cache invalidation
• cache failures
• hot keys

That’s why in interviews you shouldn’t just say “we’ll add Redis.”