Course contentsShow
← System Design
Lesson 521 of 1,91914. CAP TheoremFree lesson

PACELC Tradeoffs in Real Systems

Examining how popular distributed systems fall into different PACELC categories and the implications for design.

PACELC Tradeoffs in Real Systems

What you'll learn: How real-world distributed systems map to PACELC categories and why those choices shape their behavior.

Mapping Popular Systems to PACELC

Now that you understand the PACELC framework, let's see how actual systems fall into different categories and what that means for their design.

PA/EL Systems (Availability and Latency First)

Cassandra and DynamoDB are classic PA/EL systems. During partitions, they prioritize availability—reads and writes continue even if nodes can't communicate. In normal operation (Else), they optimize for low latency by returning potentially stale data with tunable consistency levels. This makes them ideal for use cases like shopping carts or social media feeds where brief inconsistency is acceptable.

MongoDB (in certain configurations) can operate as PA/EL when you use eventual consistency settings, trading strong guarantees for speed.

PC/EC Systems (Consistency Always)

HBase and BigTable are PC/EC systems. During partitions, they sacrifice availability to maintain consistency—if a region server is unreachable, writes to that region fail rather than risk conflicting updates. Even during normal operation, they prioritize consistency over latency, using synchronous replication before acknowledging writes.

Zookeeper is another PC/EC system, essential for distributed locking where correctness trumps speed.

PA/EC Systems (Mixed Strategy)

Riak can be configured as PA/EC: available during partitions but optionally consistent during normal operation by using quorum reads/writes with higher latency penalties.

Design Implications

Your PACELC choice determines your system's personality. PA/EL systems scale horizontally with great performance but require application-level conflict resolution. PC/EC systems offer simpler reasoning about data correctness but may experience downtime or higher latencies.

Key Takeaway: Real distributed systems don't just "choose two" from CAP—they make PACELC tradeoffs that define both their partition behavior AND their normal-operation characteristics, directly influencing which workloads they serve best.