Database Workload Early Warning via Statistical Trend Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current DBMS monitoring systems struggle to predict subtle and complex changes in workloads that may degrade system performance, leading to potential hardware issues and performance problems, as they rely on reactive and black-and-white alerting mechanisms that fail to anticipate future issues.
Innovation Solution
A method that compares current trends of performance counter components to baseline trends over longer periods, using statistical characteristics like slope, volatility, skewness, and kurtosis, to detect deviations and provide early warnings of potential future problems, with thresholds adjusted using machine learning or statistical analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional threshold-based alerting mechanisms are used, then the system can detect obvious performance problems, but it cannot predict subtle and complex changes in workload that may degrade system performance
Solution Approach 1:
The patent transforms the monitoring approach by changing from simple threshold-based parameters to statistical parameters including mean, standard deviation, skewness, and kurtosis. These statistical parameters capture the distribution characteristics of performance counter values, enabling detection of subtle workload changes that traditional thresholds miss. The system calculates these parameters over time windows and compares them against baseline statistical profiles to predict performance degradation before it occurs.
Solution Approach 2:
The patent adds dimensional complexity by moving from single-dimensional threshold checking to multi-dimensional statistical analysis. Instead of monitoring only whether values exceed thresholds, the system analyzes the distribution shape across multiple statistical dimensions (central tendency, dispersion, skewness, kurtosis). This dimensional expansion enables detection of complex workload patterns and predicts performance issues that manifest through subtle distribution changes rather than simple threshold violations.
2Ease of operation
If reactive alerting mechanisms are used, then the system responds to obvious problems, but it does not provide early warning for future performance issues
Solution Approach 1:
The patent implements preliminary action by establishing baseline statistical profiles during normal operation and continuously comparing current statistical parameters against these baselines. The system detects deviations from expected statistical patterns before they manifest as actual performance problems. By analyzing trends in statistical parameters (mean drift, increasing variance, changing skewness), the system provides early warnings that enable proactive intervention, transforming reactive monitoring into predictive monitoring.
Solution Approach 2:
The system employs feedback mechanisms by continuously comparing current statistical parameters against historical baselines and adjusting alerts based on deviation magnitude and persistence. The feedback loop analyzes multiple time windows to distinguish temporary fluctuations from sustained workload changes, providing increasingly urgent warnings as deviations persist. This feedback-driven approach maintains operational simplicity while enabling early detection and response to emerging performance issues.
3Device complexity
If simple threshold monitoring is used, then the monitoring system remains simple to implement, but it fails to capture complex workload patterns
Solution Approach 1:
The patent applies segmentation by breaking down the complex task of workload analysis into distinct statistical components: central tendency (mean), dispersion (standard deviation), asymmetry (skewness), and peakedness (kurtosis). Each component is calculated independently and analyzed against its own baseline profile. This segmentation allows the system to implement sophisticated statistical monitoring through modular, manageable calculations that can be computed efficiently and interpreted independently, maintaining implementation simplicity while achieving high measurement precision.
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
The present invention relates to a method for determining early warning on workload in a database system. The method comprises collecting periodically data samples concerning a plurality of performance counters and determining periodically one or more data points representing a statistical characteristic of said data samples over a first time period. Based on a plurality of said data points, trend points for each of a plurality of performance counter components are determined, said trend points representing a statistical characteristic determined on basis of the respective data points. Each performance counter component's current trend, determined based on the trend points over a second time period, is compared to a determined baseline trend of the same performance counter component. Any significant deviations determined based on comparisons of the current trends and the baseline trends are classified into severity classes, and an early warning of a possible future problem in the DBMS system is given based on number of said deviations of performance counter components concerning any single performance counter being categorized into at least one of said severity classes.