Muokkaa

Jaa

Monitor Azure Machine Learning

  • Artikkeli

This article describes:

  • The types of monitoring data you can collect for this service.
  • Ways to analyze that data.

Note

If you're already familiar with this service and/or Azure Monitor and just want to know how to analyze monitoring data, see the Analyze section near the end of this article.

When you have critical applications and business processes that rely on Azure resources, you need to monitor and get alerts for your system. The Azure Monitor service collects and aggregates metrics and logs from every component of your system. Azure Monitor provides you with a view of availability, performance, and resilience, and notifies you of issues. You can use the Azure portal, PowerShell, Azure CLI, REST API, or client libraries to set up and view monitoring data.

Note

The information in this document is primarily for administrators, as it describes monitoring for the Azure Machine Learning service and associated Azure services. If you are a data scientist or developer and want to monitor information specific to your model training runs, see the following documents:

If you want to monitor information generated by models deployed to online endpoints, see Monitor online endpoints.

Insights

Some services in Azure have a built-in monitoring dashboard in the Azure portal that provides a starting point for monitoring your service. These dashboards are called insights, and you can find them in the Insights Hub of Azure Monitor in the Azure portal.

Machine Learning can use Application Insights to track metrics and logs. You can send built-in metrics and logs to Application Insights, and use Application Insights features such as Live metrics, Transaction search, Failures, and Performance for further analysis. For more information, see Monitor online endpoints.

Resource types

Azure uses the concept of resource types and IDs to identify everything in a subscription. Resource types are also part of the resource IDs for every resource running in Azure. For example, one resource type for a virtual machine is Microsoft.Compute/virtualMachines. For a list of services and their associated resource types, see Resource providers.

Azure Monitor similarly organizes core monitoring data into metrics and logs based on resource types, also called namespaces. Different metrics and logs are available for different resource types. Your service might be associated with more than one resource type.

For more information about the resource types for Machine Learning, see Machine Learning monitoring data reference.

Data storage

For Azure Monitor:

  • Metrics data is stored in the Azure Monitor metrics database.
  • Log data is stored in the Azure Monitor logs store. Log Analytics is a tool in the Azure portal that can query this store.
  • The Azure activity log is a separate store with its own interface in the Azure portal.

You can optionally route metric and activity log data to the Azure Monitor logs store. You can then use Log Analytics to query the data and correlate it with other log data.

Many services can use diagnostic settings to send metric and log data to other storage locations outside Azure Monitor. Examples include Azure Storage, hosted partner systems, and non-Azure partner systems, by using Event Hubs.

For detailed information on how Azure Monitor stores data, see Azure Monitor data platform.

Azure Monitor platform metrics

Azure Monitor provides platform metrics for most services. These metrics are:

  • Individually defined for each namespace.
  • Stored in the Azure Monitor time-series metrics database.
  • Lightweight and capable of supporting near real-time alerting.
  • Used to track the performance of a resource over time.

Collection: Azure Monitor collects platform metrics automatically. No configuration is required.

Routing: You can also route some platform metrics to Azure Monitor Logs / Log Analytics so you can query them with other log data. Check the DS export setting for each metric to see if you can use a diagnostic setting to route the metric to Azure Monitor Logs / Log Analytics.

For a list of all metrics it's possible to gather for all resources in Azure Monitor, see Supported metrics in Azure Monitor.

For a list of available metrics for Machine Learning, see Machine Learning monitoring data reference.

All metrics for Azure Machine Learning are in the namespace Machine Learning Service Workspace.

Metrics Explorer with Machine Learning Service Workspace selected.

Azure Monitor resource logs

Resource logs provide insight into operations that were done by an Azure resource. Logs are generated automatically, but you must route them to Azure Monitor logs to save or query them. Logs are organized in categories. A given namespace might have multiple resource log categories.

Collection: Resource logs aren't collected and stored until you create a diagnostic setting and route the logs to one or more locations. When you create a diagnostic setting, you specify which categories of logs to collect. There are multiple ways to create and maintain diagnostic settings, including the Azure portal, programmatically, and though Azure Policy.

Routing: The suggested default is to route resource logs to Azure Monitor Logs so you can query them with other log data. Other locations such as Azure Storage, Azure Event Hubs, and certain Microsoft monitoring partners are also available. For more information, see Azure resource logs and Resource log destinations.

For detailed information about collecting, storing, and routing resource logs, see Diagnostic settings in Azure Monitor.

For a list of all available resource log categories in Azure Monitor, see Supported resource logs in Azure Monitor.

All resource logs in Azure Monitor have the same header fields, followed by service-specific fields. The common schema is outlined in Azure Monitor resource log schema.

For the available resource log categories, their associated Log Analytics tables, and the logs schemas for Machine Learning, see Machine Learning monitoring data reference.

Azure activity log

The activity log contains subscription-level events that track operations for each Azure resource as seen from outside that resource; for example, creating a new resource or starting a virtual machine.

Collection: Activity log events are automatically generated and collected in a separate store for viewing in the Azure portal.

Routing: You can send activity log data to Azure Monitor Logs so you can analyze it alongside other log data. Other locations such as Azure Storage, Azure Event Hubs, and certain Microsoft monitoring partners are also available. For more information on how to route the activity log, see Overview of the Azure activity log.

Analyze monitoring data

There are many tools for analyzing monitoring data.

Azure Monitor tools

Azure Monitor supports the following basic tools:

Tools that allow more complex visualization include:

  • Dashboards that let you combine different kinds of data into a single pane in the Azure portal.
  • Workbooks, customizable reports that you can create in the Azure portal. Workbooks can include text, metrics, and log queries.
  • Grafana, an open platform tool that excels in operational dashboards. You can use Grafana to create dashboards that include data from multiple sources other than Azure Monitor.
  • Power BI, a business analytics service that provides interactive visualizations across various data sources. You can configure Power BI to automatically import log data from Azure Monitor to take advantage of these visualizations.

Azure Monitor export tools

You can get data out of Azure Monitor into other tools by using the following methods:

To get started with the REST API for Azure Monitor, see Azure monitoring REST API walkthrough.

Kusto queries

You can analyze monitoring data in the Azure Monitor Logs / Log Analytics store by using the Kusto query language (KQL).

Important

When you select Logs from the service's menu in the portal, Log Analytics opens with the query scope set to the current service. This scope means that log queries will only include data from that type of resource. If you want to run a query that includes data from other Azure services, select Logs from the Azure Monitor menu. See Log query scope and time range in Azure Monitor Log Analytics for details.

For a list of common queries for any service, see the Log Analytics queries interface.

You can use the following queries to help you monitor your Machine Learning resources:

  • Get failed jobs in the last five days:

    AmlComputeJobEvent
| where TimeGenerated > ago(5d) and EventType == "JobFailed"
| project  TimeGenerated , ClusterId , EventType , ExecutionState , ToolType

  • Get records for a specific job name:

    AmlComputeJobEvent
| where JobName == "automl_a9940991-dedb-4262-9763-2fd08b79d8fb_setup"
| project  TimeGenerated , ClusterId , EventType , ExecutionState , ToolType

  • Get cluster events in the last five days for clusters where the VM size is Standard_D1_V2:

    AmlComputeClusterEvent
| where TimeGenerated > ago(4d) and VmSize == "STANDARD_D1_V2"
| project  ClusterName , InitialNodeCount , MaximumNodeCount , QuotaAllocated , QuotaUtilized

  • Get the cluster node allocations in the last eight days::

    AmlComputeClusterEvent
| where TimeGenerated > ago(8d) and TargetNodeCount  > CurrentNodeCount
| project TimeGenerated, ClusterName, CurrentNodeCount, TargetNodeCount

When you connect multiple Machine Learning workspaces to the same Log Analytics workspace, you can query across all resources.

  • Get number of running nodes across workspaces and clusters in the last day:

    AmlComputeClusterEvent
| where TimeGenerated > ago(1d)
| summarize avgRunningNodes=avg(TargetNodeCount), maxRunningNodes=max(TargetNodeCount)
         by Workspace=tostring(split(_ResourceId, "/")[8]), ClusterName, ClusterType, VmSize, VmPriority

Alerts

Azure Monitor alerts proactively notify you when specific conditions are found in your monitoring data. Alerts allow you to identify and address issues in your system before your customers notice them. For more information, see Azure Monitor alerts.

There are many sources of common alerts for Azure resources. For examples of common alerts for Azure resources, see Sample log alert queries. The Azure Monitor Baseline Alerts (AMBA) site provides a semi-automated method of implementing important platform metric alerts, dashboards, and guidelines. The site applies to a continually expanding subset of Azure services, including all services that are part of the Azure Landing Zone (ALZ).

The common alert schema standardizes the consumption of Azure Monitor alert notifications. For more information, see Common alert schema.

Types of alerts

You can alert on any metric or log data source in the Azure Monitor data platform. There are many different types of alerts depending on the services you're monitoring and the monitoring data you're collecting. Different types of alerts have various benefits and drawbacks. For more information, see Choose the right monitoring alert type.

The following list describes the types of Azure Monitor alerts you can create:

  • Metric alerts evaluate resource metrics at regular intervals. Metrics can be platform metrics, custom metrics, logs from Azure Monitor converted to metrics, or Application Insights metrics. Metric alerts can also apply multiple conditions and dynamic thresholds.
  • Log alerts allow users to use a Log Analytics query to evaluate resource logs at a predefined frequency.
  • Activity log alerts trigger when a new activity log event occurs that matches defined conditions. Resource Health alerts and Service Health alerts are activity log alerts that report on your service and resource health.

Some Azure services also support smart detection alerts, Prometheus alerts, or recommended alert rules.

For some services, you can monitor at scale by applying the same metric alert rule to multiple resources of the same type that exist in the same Azure region. Individual notifications are sent for each monitored resource. For supported Azure services and clouds, see Monitor multiple resources with one alert rule.

Note

If you're creating or running an application that runs on your service, Azure Monitor application insights might offer more types of alerts.

Machine Learning alert rules

The following table lists common and recommended alert rules for Machine Learning.

Alert type Condition Description
Model Deploy Failed Aggregation type: Total, Operator: Greater than, Threshold value: 0 When one or more model deployments have failed
Quota Utilization Percentage Aggregation type: Average, Operator: Greater than, Threshold value: 90 When the quota utilization percentage is greater than 90%
Unusable Nodes Aggregation type: Total, Operator: Greater than, Threshold value: 0 When there are one or more unusable nodes

Advisor recommendations

For some services, if critical conditions or imminent changes occur during resource operations, an alert displays on the service Overview page in the portal. You can find more information and recommended fixes for the alert in Advisor recommendations under Monitoring in the left menu. During normal operations, no advisor recommendations display.

For more information on Azure Advisor, see Azure Advisor overview.

Related content