Purpose
This article explains where the main observability data is available in the platform and which tool to use depending on the kind of investigation you need to perform.
Overview
The platform provides several observability data, each serving a different purpose. Each type answers different questions about system behavior.
The main tools are:
- VictoriaLogs for logs. Logs are timestamped records of events produced by Giotto services.
- VictoriaMetrics for metrics
- Jaeger for traces, which show show a request or operation moves through multiple services and components.
- Superset for health-check dashboards
- Ray UI for Ray-specific runtime details
- Airflow for orchestration status
A good rule of thumb is:
- use logs to understand what happened, what failed, error messages and detailed runtime behavior.
- use metrics to understand behavior over time
- use traces to understand execution paths across services
Where to see logs
Logs are available in VictoriaLogs.
To access it:
- open the platform landing UI
- click the Logs button
VictoriaLogs provides a centralized LogQL-style interface and includes logs from multiple platform components, including centralized Ray logs.
Where to see metrics
Metrics are available in VictoriaMetrics. Metrics are numerical measurements collected over time.
To access it:
- open the platform landing UI
- click the Metrics button
VictoriaMetrics exposes time-series metrics, especially from systems such as Ray and Airflow. It is useful for inspecting GPU usage, workload activity, and resource behavior over time.
Where to see traces
Traces are available in Jaeger. Traces represent the full execution path of a request, including which services were involved, how long each step took and where failures and slowdowns occurred.
Jaeger is the platform’s distributed tracing tool and is used to inspect how requests and execution paths move across multiple services. It is especially useful when behavior spans several components and you need an end-to-end runtime view.
Where to see health status
Service health is visible in Superset.
To access it:
- open the platform landing UI
- click the SuperSet button
The health dashboard shows:
- a time series of services going up and down
- a table with current status and logged errors
Email alerts are also sent when a monitored service goes down.
Where to see workflow state
Workflow state is primarily visible in Airflow.
Use Airflow when you want to understand:
- whether a flow or deployment DAG exists
- whether it was triggered
- which task failed
- where in the orchestration sequence the run stopped
Where to see Ray-specific runtime details
For Ray-based training, inference, or serving, the Ray UI is also useful.
Use Ray UI when you want to inspect:
- deployed agents in the Serve section
- execution-local logs
- compute-side runtime behavior
Which one should I use first?
A useful default is:
- start with Airflow if the question is about whether a job ran
- start with VictoriaLogs if the question is what failed
- start with VictoriaMetrics if the question is whether the system was active or using resources
- start with Jaeger if the question is how a request moved across services
- start with Superset health checks if the question is whether a service was up
Comments
0 comments
Article is closed for comments.