Simple Log Service provides a time series forecasting feature that automatically predicts time series data. You can use the prediction results to determine future data trends and gain early insights into the status of key system or business metrics.
Starting July 15, 2025 (UTC+8), the intelligent anomaly analysis feature will no longer be available to new users. Existing users can continue to use it.
-
Scope of impact
The following core features will be unpublished: intelligent health check, text analytics, and time series forecasting.
-
Feature Migration Solutions
The machine learning syntax, scheduled query and analysis (scheduled SQL), and dashboard features of SLS can fully replace the unpublished features.
Background information
Running services generate time series data that records how metrics change over time. Monitoring and analyzing this data is widely used for system monitoring and fault localization. Time series data is typically monitored and analyzed in one of the following two ways.
-
Analyze existing time series data to detect possible anomalies by using the intelligent health check feature. For more information, see Intelligent health check.
This approach helps you promptly find and locate problems after a service becomes abnormal.
-
Predict future time series data to determine future trends by using the time series forecasting feature.
This approach provides early warnings for abnormal trends in key service metrics.
Common scenarios for time series forecasting:
-
Predict future trends of key service metrics, such as queries per second (QPS) and the number of online users. If a predicted value exceeds a threshold, an alert is triggered, allowing O&M engineers to troubleshoot potential issues before anomalies occur.
-
Predict future fluctuations in system metrics, such as CPU utilization and disk utilization. Use the predicted trends to make more reliable system adjustments. For example, scale out a system when its CPU utilization is predicted to be high.
Time series predictions are affected by factors such as bursts, instability, and internal bugs in prediction algorithms. Prediction results cannot be 100% accurate. Use them as a reference for decision-making, not as the sole basis.
Function Introduction
A time series forecasting job uses SQL to extract or aggregate time series metrics. Based on a scheduling rule, the job periodically pulls data, feeds it into a prediction model, and writes the prediction results to a destination Logstore named internal-ml-log. The results are displayed on a dashboard.
Terms
The following table describes the basic concepts of time series forecasting.
|
Term |
Description |
|
Job |
A time series forecasting task corresponds to a job, which includes data features and model parameters. |
|
Instance |
A time series forecasting job generates execution instances based on its configuration. Each instance periodically pulls data, runs an algorithm model, and distributes prediction results. For more information about how different operations affect instance scheduling and execution, see Scheduling and execution scenarios.
|
|
Instance ID |
The unique identifier of an execution instance. |
|
Creation time |
The time when an instance is created. Instances are typically generated based on the job rules that you configure. Instances are immediately generated to backfill data or catch up on delays. |
|
Running time |
The time when an instance starts to run. If a job is retried, this indicates the time when the last retry started. |
|
End time |
The time when an instance stops running. If a task is retried, this indicates the time when the last retry ended. |
|
Execution Status |
The status of the instance. Valid values:
|
|
Data feature configuration |
The data feature configuration includes the time item, entity item, and feature item. For more information, see Data feature configuration parameters. |
|
Algorithm configuration |
The algorithm configuration includes epoch, holiday, and prediction settings. The prediction configuration covers the test sequence length, confidence level, number of samples, prediction frequency, and observation duration. For more information, see Algorithm configuration parameters. |
|
Prediction result |
The prediction results are displayed on a built-in dashboard. |
Scheduling and execution scenarios
A job can generate one or more instances, but only one instance can run at a time, regardless of whether it is scheduled normally or is a retry of a failed instance. Concurrent execution of multiple instances is not supported.
The following table describes the scheduling and execution scenarios.
|
Scenario |
Description |
|
Scenario 1: Executing a time series forecasting job from the current time |
Start a job from the current time. The job reads historical data based on the configured observation duration and predicts the time series for a future period. |
|
Scenario 2: Executing a time series forecasting job from a historical point in time |
When a job is created to start from a specified historical point in time, it processes historical data based on the job rules. The algorithm model quickly consumes the historical data and gradually catches up to the current time. |
|
Scenario 3: Modifying the scheduling configuration |
After you modify the scheduling configuration, the next instance is generated based on the new configuration. The algorithm model remembers the time point of the last consumption and continues to process new data. |
|
Scenario 4: Retrying a failed instance |
If an instance fails due to insufficient permissions, a non-existent source or destination database, or an invalid configuration, the system automatically retries it. If the instance status remains STARTING, the configuration may have failed. Error logs are sent to your internal-etl-log Logstore. Check the configuration and start the instance again. After scheduling completes, the system changes the instance status to SUCCEEDED or FAILED based on the execution result. |
Usage recommendations
For efficient data conversion and prediction, specify only the required metrics to monitor. Note the following:
-
When you upload data to a Logstore, define the data format, the meaning of each field, and the observation interval to quickly configure a forecasting job.
-
Understand the changes, stability, and periodicity of the time series data for the monitored object to properly configure the algorithm parameters.
-
Align the time window of the forecasting job with an integer time, such as the nearest second, minute, or hour.