Ritvik Pandya presents short-lived execution graphs built from distributed tracing in payment APIs. Normal paths, ordering and retries provide a baseline that can make deviations easier to localize than a single aggregate latency or error-rate measure.
The system uses an inexpensive first stage before deeper statistical and structural checks. Pandya distinguishes unusual behavior from legitimate drift, such as a changing customer or traffic mix, and explains why client-aware baselines and deployment context matter.
Operational considerations include asynchronous processing, missing or delayed traces, new endpoints and staged rollout. The talk concerns learned and statistical reliability tooling rather than an LLM agent, and its benefits are presented as the speaker's engineering account.
Watch on YouTube
