AvroCoder vs SerializableCoder Benchmark Results

Overview

We created a benchmark to compare SerializableCoder (Java Serialization) and AvroCoder (Apache Avro) for a typical POJO in an Apache Beam pipeline.

Benchmark Setup

• Object Type: TestObject (POJO with String, int, List, Map, boolean)
• Count: 1,000,000 objects
• Iterations: 10 (after warmup)
• Environment: Local execution

Results (1,000,000 Objects)

Metric	SerializableCoder	AvroCoder	Difference
Avg Serialize Time	~2360 ms	~384 ms	~6.1x Faster
Avg Deserialize Time	~9567 ms	~754 ms	~12.7x Faster
Avg Size per Object	~492 bytes	~59 bytes	~8.3x Smaller
Total Size	468.85 MB	56.04 MB	~8.3x Smaller

Conclusion

AvroCoder is vastly superior to SerializableCoder for both performance and storage efficiency.

• CPU: Serialization and deserialization are much faster, which reduces CPU usage and processing time in Dataflow.
• Network/Storage: The serialized size is nearly an order of magnitude smaller, which significantly reduces network I/O (shuffle) and storage costs.

Recommendation: Always prefer AvroCoder (or Schema-based coders) over SerializableCoder for custom types in production pipelines.

How to Run

cd beam-coder-benchmark
mvn clean package -DskipTests
mvn exec:java -Dexec.mainClass="com.example.Benchmark"