Latency Overhead Issues When Using Redis Pipelines

lettuce-pipeline-demo.zip
Here is a demonstration of using Redis pipelines, which inserts 600,000 Redis subcommands(zset) in a single pipeline.

1. With Java Agent：
Pipeline execution completed. Execution time: 11127ms
2. Without Java Agent：
Pipeline execution completed. Execution time: 5167ms

I have observed that overall latency increases by approximately 40% with the Java Agent. I infer this is due to excessive invocations of the io.opentelemetry.instrumentation.lettuce.v5_1.OpenTelemetryTracing.OpenTelemetrySpan#start() method, resulting in a noticeable rise in latency.

Subsequently, I modified the Java Agent so that when using the pipeline in spring-data-redis, then only a single internal span named “pipeline” is generated, with the remaining detailed spans being omitted. Following this change, the overall call duration returned to normal. Should we provide a switch to selectively display the detailed spans within the pipeline?

Start with the attached lettuce-pipeline-demo.zip and the reported io.opentelemetry.instrumentation.lettuce.v5_1.OpenTelemetryTracing.OpenTelemetrySpan#start() path. Compare the detailed spans produced for the 600,000-command pipeline with the single “pipeline” span described in the report. Done means a documented switch can control detailed pipeline spans while preserving normal execution latency.