Monitoring Limits

General Monitoring Limits

• You cannot revert to basic Droplet graphs after enabling the metrics agent. If you prefer different visualizations, export the metrics to an external tool like Grafana to create a customized, lightweight dashboard.

• You cannot export metrics data directly from the do-agent service. It collects system metrics and sends them to our backend services, where they appear in the DigitalOcean Control Panel and API.

• You cannot add or change the alert email address for individual accounts. Use a team account email or configure Slack integration to receive alerts through a shared channel.

• The metrics agent only collects system-level data and does not support application performance monitoring (APM). It does not track application errors, website traffic, or performance bottlenecks. To monitor application-level metrics, use a dedicated APM tool alongside DigitalOcean Monitoring. For example, New Relic offers comprehensive monitoring, including error tracking, transaction tracing, and performance analysis.

GPU Observability Limits

• You cannot enable or disable GPU Observability after Droplet creation. It is available only when you select Improved Metrics and Monitoring during Droplet setup with an AI/ML Ready Image.

• Alerts for GPU metrics are not yet available. You can view GPU metrics, but alerting on GPU data is currently unsupported.

• GPU temperatures maintain typical operating limits that align with vendor specifications. Sustained values near the rated maximum may indicate thermal constraints.

• GPU occupancy or memory usage with sustained values above 85–95% suggests the GPU is nearing full utilization or memory saturation.

• GPU metric limits vary by metric type. For example, °C for temperature, W for power, or % for occupancy and memory utilization.

• You cannot view power utilization as a percentage. Power usage (in watts) is available, but percentage-based power utilization metrics are not included.

• You may see differences in metric availability across GPU models. We provide an extensive set of metrics for NVIDIA A30, H100, and H200 GPUs, as well as AMD MI300 and MI325 GPUs. Profiler metrics, such as occupancy and tensor core utilization, are not available for NVIDIA L40S, RTX 6000 Ada, and RTX 4000 Ada GPUs.

• You may not see GPU Occupancy (%) or Tensor Utilization (%) on RTX-class GPUs. Some models, such as RTX 4000, 6000, and L40S, do not expose DCGM profiling metrics due to NVIDIA exporter limitations.

• You must use an AI/ML Ready Image to automatically enable GPU observability. For Basic Images, you can enable GPU metrics by manually installing the exporter, binding it to 127.0.0.1, reconfiguring do-agent to scrape it, and restarting do-agent.

• You cannot create new alert types specific to GPU metrics as GPU metrics work with the same alerting and notification system as other monitoring data.