Understanding Negative Latency in Experiments: Is It Possible?

Zhuyi X. · 2025-08-13T18:22:13.544Z

so negative latency is possible? my guess is that it measure the time spent on calculating a metric during experiment

It seems likely that negative latency ("minus latency") is not expected or meaningful in Phoenix, as latency is calculated as the difference between end and start times, which should always be non-negative. Latency in Phoenix measures the time taken for a trace, span, or experiment run, not the time spent calculating a metric, but rather the duration of the operation itself, such as an LLM call or experiment run, using (end_time - start_time) in milliseconds according to the codebase and schema definitions (source1, source2).If you are seeing negative latency values, it could indicate a data or clock issue, as Phoenix expects start_time to be before end_time. Would you like more detail on how latency is computed or how to troubleshoot negative values?Sources: - phoenix/db/models.py - app/src/pages/experiment/ExperimentCompareMetricsPage.tsx

💡 Hint: Mention RunLLM in the thread for followups.

5 comments