ndif-team/ndif

Opentelemetry Tracing

Open

#137 ouverte le 29 août 2025

Voir sur GitHub
 (2 commentaires) (0 réactions) (0 assignés)Python (17 forks)github user discovery
help wanted

Métriques du dépôt

Stars
 (45 stars)
Métriques de merge PR
 (Métriques PR en attente)

Description

Description

Tracing is a form of telemetry data which tracks fine-grain information on requests as they flow through an application.

Tracing could help us in a number of ways:

  • Debugging: If someone says "my request hung," can pull a trace and immediately see: was it stuck waiting for model spin-up, or in queue, or the model crashed?
  • Development Help: Can be used to study system performance before/after a new feature was added (or help debug if something new isn't working)
  • Very fine-grained benchmarking: For example, determining if there any features of a pickled nnsight experiment which correlate directly with experiment length.

Overall, it provides a very clean way to visualize and study how user requests flow through our system. This could provide useful insights into how particular components of NDIF perform under varying conditions/contexts.

Desired outcome

Implement some level of cross-service tracing using opentelemetry, with span data collected in jaeger and visualized in grafana. Even just getting a basic prototype which works across our 3 services (api, queue, ray) would be an incredible milestone.

Guide contributeur