Allegro experimented with Kafka protocol sniffing, JVM profiling, and eBPF to identify and eliminate performance bottlenecks, eventually moving all clusters to XFS filesystem for better tail latency.
The engineers at Allegro analyzed Kafka producer latency metrics, discovering unacceptable p99 latency of 1 second and p999 latency of up to 3 seconds, prompting their investigation into improving performance for low-latency messaging needs.
Collection
[
|
...
]