Docs

Observability Platform

General AI application monitoring

Data backflow

In distributed system observability scenarios, trace data generated in online environments is key to identifying system performance bottlenecks, abnormal calls, and service dependencies. As capabilities such as intelligent diagnosis, performance analysis, and trace detection rules continue to iterate, large numbers of standardized, high-quality trace samples are required for evaluation.

By exporting trace data in batches to evaluation sets, you can build a standardized and reusable trace sample library. You can then run evaluation experiments directly using trace samples from the evaluation sets. This supports scenarios such as trace diagnosis algorithm verification, detection rule optimization, anomaly detection evaluation, and fault case accumulation and reproduction. This process forms a closed loop from real online data to evaluation and verification, which improves the accuracy and reliability of diagnostic capabilities on the Observability Platform. This topic describes how to export AI trace data.

Prerequisites

AI-related applications have been integrated with Server-side Monitoring and are reporting observability data normally. For details, see AI scenario monitoring overview.

Limits

Currently, the data backflow feature is available only for AI-related applications. Both the total exported data limit and the limit per export task are 100,000 records.

You can use the data backflow feature to export only new data. Exporting historical data is temporarily not supported. To export historical data, see Offline trace export.

Procedure

In the left navigation bar, select Online Trace Export.

Click Create export task, configure the parameters based on your business requirements, and then click Confirm.

Category	Parameter	Description
Basic information	Task name	Customize the name of the export task.
Basic information		Overall data limit	Each export task supports exporting up to 100,000 records. You can adjust this value based on actual scenarios.
Data filtering	Data filtering	The range of data to be exported. Supported filter parameters include: Span type, AI span type, Service name, Model name, Interface name, Status code, Actions, and Time consumption. tip Multiple conditions can be configured for the same filter parameter. These conditions are evaluated based on OR logic. This indicates that data that meets all of the conditions is exported. If multiple filter parameters are configured simultaneously, they apply based on AND logic. This indicates that only data that meets all specified parameters simultaneously is exported.
Time range	New data feed	The data export method. Currently, only exporting new data is supported. The export rules will be applied to newly generated trace data that meets the filter conditions.
		Overall time range	Specify the total time range of the exported data based on actual scenarios.
		Repeat frequency	The repeat frequency of the export task. You can specify a time period within the total range to run the export task daily or weekly. Do not repeat: If the execution frequency is not configured, all data within the total time range will be exported. For example, if the total time range is 2026-03-03 00:00:00 to 2026-03-05 00:00:00, all data meeting the requirements within this time range are exported. Repeat by day: Specify the daily execution frequency and per-run data limit for the export task. Only data within the specified time period is exported. For example, if the total time range is 2026-03-03 00:00:00 to 2026-03-05 00:00:00 and the repeat frequency is 00:00:00 to 03:00:00 every day, only data meeting the requirements from 2026-03-03 00:00:00 to 03:00:00 and 2026-03-04 00:00:00 to 03:00:00 is exported. Repeat by week: Specify the execution frequency and per-run data limit for the export task every week. Only data within the specified time period is exported. warning Both the total data limit and the limit per export task are 100,000 records. You can customize these limits. Data exceeding the upper limit will fail to be exported.
Export settings	Export platform	Exports trace data to the agent evaluation platform.
		Data masking	On (default): Data masking is enabled. The Input and Output information in exported data is automatically masked to protect sensitive data. Off: Data masking is disabled. You can export the complete Input and Output information for debugging and troubleshooting.
		Target evaluation set	Select an existing evaluation set on the agent evaluation platform, or add a new evaluation set. When adding a new evaluation set, you need to customize the name of the target evaluation set.
		Import method	The method used to import trace data to the agent evaluation platform. Currently, only appending data is supported. This indicates that data is appended to the existing data.
		Field mapping	The field mapping relationships when data is imported to the agent evaluation platform. Sequentially select the trace span fields and the fields in the target evaluation set.

What to do next

After the export task is created, you are redirected to the task list. You can click the destination in the "Export destination" column of the target task to go to the target evaluation set and view the exported data details. You can also use the exported trace data to create an evaluation set to run evaluation experiments. For evaluation details, see What is evaluation.

The export tasks support the following basic actions.

Actions	Description
Pause/enable	Manually pause or start an export task within the specified time range.
Copy	Copy an existing export task and modify the configurations as needed to quickly create a new export task.
Edit	Edit the configurations of an existing export task. The task will be executed based on the new configurations.
Delete	Delete an existing export task. warning Deleted tasks cannot be recovered. Proceed with caution.

Last updated: 2026.08.28 17:26:59