Purpose
This article explains how to access the platform metrics interface and use it to inspect infrastructure and runtime metrics such as GPU activity.
Overview
The platform uses VictoriaMetrics as its centralized metrics system. It exposes a query interface compatible with the same query style commonly used with Prometheus, allowing users to inspect time-series metrics across the platform.
VictoriaMetrics collects and exposes many metrics from key systems, especially Ray and Airflow, which makes it useful for monitoring execution health, system performance, and infrastructure usage such as GPU activity.
When to use VictoriaMetrics
Use VictoriaMetrics when you want to:
- inspect infrastructure usage over time
- monitor GPU activity
- review metrics coming from Ray
- review metrics coming from Airflow
- understand runtime behavior through time-series data
- validate whether services or workloads were active during a time window
VictoriaMetrics is especially useful when the question is about performance, load, or resource usage rather than logs or functional behavior.
How to access the metrics UI
To open the metrics interface:
- open the platform landing UI
- click the Metrics button
This opens the VictoriaMetrics UI.
How the interface works
VictoriaMetrics uses a query language compatible with the Prometheus-style metric querying model. In practice, this means users can inspect metrics using the same query approach commonly used for Prometheus and PromQL-style time-series exploration.
This makes the interface familiar to users who already work with infrastructure monitoring tools in the Prometheus ecosystem.
What metrics are available
VictoriaMetrics collects and exposes many metrics across the platform. In particular, it includes metrics from:
- Ray
- Airflow
These metrics make it possible to inspect:
- workload activity
- resource utilization
- scheduler or execution patterns
- GPU usage and similar infrastructure signals
For example, users can inspect GPU activity to confirm whether a compute-heavy job actually used the expected hardware during execution.
Typical workflow for metric inspection
A common way to use VictoriaMetrics is:
- identify the workflow, service, or time window of interest
- open the Metrics interface
- query the relevant metric
- inspect the time-series behavior
- correlate the results with logs, Airflow runs, or Ray execution state
This makes VictoriaMetrics especially valuable when you want to understand not just whether something happened, but how the system behaved over time.
Common use cases
Checking GPU activity
One common use case is verifying whether a Ray training or inference workload actually used GPU resources during execution.
Monitoring Airflow-related metrics
Airflow-exposed metrics can help users understand whether workflow activity was present during a given time window.
Monitoring Ray-related metrics
Ray metrics can help explain runtime behavior in distributed compute or serving scenarios.
VictoriaMetrics versus logs
Logs and metrics serve different purposes.
A useful mental model is:
- logs explain what happened in detail
- metrics show patterns, load, and system behavior over time
If you need to understand whether GPUs were active, whether a service was under load, or whether a task caused resource pressure, VictoriaMetrics is usually the right first stop.
Best practices
Use metrics together with logs
Metrics help identify the time window and system behavior, while logs help explain the exact events behind it.
Focus on the relevant subsystem
If you are debugging a workflow execution problem, look first at Airflow and Ray metrics. If you are debugging serving behavior, focus more on runtime and infrastructure activity.
Use metrics to confirm assumptions
If a job was expected to use GPUs, VictoriaMetrics is one of the best places to verify that assumption.
Compare time-series behavior with workflow lifecycle
Metrics are much easier to interpret when compared with known launch, execution, and failure times.
Comments
0 comments
Article is closed for comments.