Abstract: Techniques for machine-learning of long-term seasonal patterns are disclosed. In some embodiments, a network service receives a set of time-series data that tracks metric values of at least one computing resource over time. Responsive to receiving the time-series data, the network service detects a subset of metric values that are outliers and associated with a plurality of timestamps. The network service maps the plurality of timestamps to one or more encodings of at least one encoding space that defines a plurality of encodings for different seasonal patterns. Based on the mapped encodings, the network service generates a representation of a seasonal pattern. Based on the representation of the seasonal pattern, the network service may perform one or more operations in association with the at least one computing resource.

Abstract: Techniques for artificial intelligence driven configuration management are described herein. In some embodiments, a machine-learning process determines a feature set for a plurality of deployments of a software resource. Based on varying values in the feature set, the process clusters each of the plurality of deployments into a cluster of a plurality of clusters. Each cluster of the plurality of clusters comprises one or more nodes and each node of the one or more nodes corresponds to at least a subset of values of the feature set that are detected in at least one deployment of the plurality of deployments of the software resource. The process determines a representative node for each cluster of the plurality of clusters. An operation may be performed based on the representative node for at least one cluster.

Abstract: Techniques for machine-learning of long-term seasonal patterns are disclosed. In some embodiments, a network service receives a set of time-series data that tracks metric values of at least one computing resource over time. Responsive to receiving the time-series data, the network service detects a subset of metric values that are outliers and associated with a plurality of timestamps. The network service maps the plurality of timestamps to one or more encodings of at least one encoding space that defines a plurality of encodings for different seasonal patterns. Based on the mapped encodings, the network service generates a representation of a seasonal pattern. Based on the representation of the seasonal pattern, the network service may perform one or more operations in association with the at least one computing resource.

Abstract: Techniques are disclosed for summarizing, diagnosing, and correcting the cause of anomalous behavior in computing systems. In some embodiments, a system identifies a plurality of time series that track different metrics over time for a set of one or more computing resources. The system detects a first set of anomalies in a first time series that tracks a first metric and assigns a different respective range of time to each anomaly. The system determines whether the respective range of time assigned to an anomaly overlaps with timestamps or ranges of time associated with anomalies from one or more other time series. The system generates at least one cluster that groups metrics based on how many anomalies have respective ranges of time and/or timestamps that overlap. The system may preform, based on the cluster, one or more automated actions for diagnosing or correcting a cause of anomalous behavior.

Abstract: Generating persistent multifaceted statistical distributions of event data associated with computing nodes is disclosed. From a data stream, events are identified that occur during a first time interval. Characteristics associated with the events are determined. Based on a primary characteristic, it is determined that an event corresponds to an event cluster. The event count for that cluster is incremented. It is determined that the characteristics correspond to an event descriptor of events in the cluster. Responsive to requests to view the event cluster, information about descriptors from the cluster are displayed indicating events having a particular event descriptor, or a summary of characteristics that distinguish the descriptor from other event descriptors.

Abstract: Techniques are disclosed for summarizing, diagnosing, and correcting the cause of anomalous behavior in computing systems. In some embodiments, a system identifies a plurality of time series that track different metrics over time for a set of one or more computing resources. The system detects a first set of anomalies in a first time series that tracks a first metric and assigns a different respective range of time to each anomaly. The system determines whether the respective range of time assigned to an anomaly overlaps with timestamps or ranges of time associated with anomalies from one or more other time series. The system generates at least one cluster that groups metrics based on how many anomalies have respective ranges of time and/or timestamps that overlap. The system may preform, based on the cluster, one or more automated actions for diagnosing or correcting a cause of anomalous behavior.

Abstract: Systems and methods for unsupervised training and evaluation of anomaly detection models are described. In some embodiments, an unsupervised process comprises generating an approximation of a data distribution for a training dataset including varying values for a metric of a computing resource. The process further determines, based on the size of the training dataset, a first quantile probability and a second quantile probability that represent an interval for covering a prescribed proportion of values for the metric within a prescribed confidence level. The process further trains a lower limit of the anomaly detection model using a first quantile that represents the first quantile probability in the approximation of the data distribution and an upper limit using a second quantile that represents the second quantile probability in the approximation. The trained upper and lower limits may be used to monitor input data for anomalous behavior and, if detected, trigger responsive action(s).

Abstract: Techniques are disclosed for summarizing, diagnosing, and correcting the cause of anomalous behavior in computing systems. In some embodiments, a system identifies a plurality of time series that track different metrics over time for a set of one or more computing resources. The system detects a first set of anomalies in a first time series that tracks a first metric and assigns a different respective range of time to each anomaly. The system determines whether the respective range of time assigned to an anomaly overlaps with timestamps or ranges of time associated with anomalies from one or more other time series. The system generates at least one cluster that groups metrics based on how many anomalies have respective ranges of time and/or timestamps that overlap. The system may preform, based on the cluster, one or more automated actions for diagnosing or correcting a cause of anomalous behavior.

Abstract: Generating persistent multifaceted statistical distributions of event data associated with computing nodes is disclosed. From a data stream, events are identified that occur during a first time interval. Characteristics associated with the events are determined. Based on a primary characteristic, it is determined that an event corresponds to an event cluster. The event count for that cluster is incremented. It is determined that the characteristics correspond to an event descriptor of events in the cluster. Responsive to requests to view the event cluster, information about descriptors from the cluster are displayed indicating events having a particular event descriptor, or a summary of characteristics that distinguish the descriptor from other event descriptors.

Abstract: Techniques for training and evaluating seasonal forecasting models are disclosed. In some embodiments, a network service generates, in memory, a set of data structures that separate sample values by season type and season space. The set of data structures may include a first set of clusters corresponding to different season types in the first season space and a second set of clusters corresponding to different season types in the second season space. The network service merges two or more clusters the first set and/or second set of clusters. Clusters from the first set are not merged with clusters from the second set. After merging the clusters, the network service determines a trend pattern for each of the remaining clusters in the first and second set of clusters. The network service then generates a forecast for a metric of a computing resource based on the trend patterns for each remaining cluster.

Abstract: Generating persistent multifaceted statistical distributions of event data associated with computing nodes is disclosed. From a data stream, events are identified that occur during a first time interval. Characteristics associated with the events are determined. Based on a primary characteristic, it is determined that an event corresponds to an event cluster. The event count for that cluster is incremented. It is determined that the characteristics correspond to an event descriptor of events in the cluster. Responsive to requests to view the event cluster, information about descriptors from the cluster are displayed indicating events having a particular event descriptor, or a summary of characteristics that distinguish the descriptor from other event descriptors.

Abstract: Techniques for artificial intelligence driven configuration management are described herein. In some embodiments, a machine-learning process determines a feature set for a plurality of deployments of a software resource. Based on varying values in the feature set, the process clusters each of the plurality of deployments into a cluster of a plurality of clusters. Each cluster of the plurality of clusters comprises one or more nodes and each node of the one or more nodes corresponds to at least a subset of values of the feature set that are detected in at least one deployment of the plurality of deployments of the software resource. The process determines a representative node for each cluster of the plurality of clusters. An operation may be performed based on the representative node for at least one cluster.

Abstract: Systems and methods are described for efficiently detecting an optimal number of behaviors to model software system performance data and the aspects of the software systems that best separate the behaviors. The behaviors may be ranked according to how well fitting functions partition the performance data.

Abstract: Techniques for artificial intelligence driven configuration management are described herein. In some embodiments, a machine-learning process determines a feature set for a plurality of deployments of a software resource. Based on varying values in the feature set, the process clusters each of the plurality of deployments into a cluster of a plurality of clusters. Each cluster of the plurality of clusters comprises one or more nodes and each node of the one or more nodes corresponds to at least a subset of values of the feature set that are detected in at least one deployment of the plurality of deployments of the software resource. The process determines a representative node for each cluster of the plurality of clusters. An operation may be performed based on the representative node for at least one cluster.

Abstract: The present disclosure describes a flexible technique to learn patterns in time series data that recur over time. The patterns may be used for simulation, predicting future behavior, or detecting anomalies in a system in which the data is collected. The technique incrementally detects daily, weekly, monthly, and yearly patterns. Each pattern is built over time instead of requiring all the data to be available at the beginning of the analysis. Instead of modeling each pattern explicitly, each pattern is described in the context of a day and formed based on time series data collected over an entire day. An example use of the technique is detecting load patterns in a computer system. A metric of system load such as CPU utilization may be collected periodically over a day. The techniques presented herein capture multiple daily models, each representing a different load pattern.

Abstract: Techniques for efficiently generating aggregate distribution approximations are disclosed. In some embodiments, a system receives a plurality of piecewise approximations that represent different distributions of a set of values on at least one computing resource. Based on the plurality of piecewise approximations, a set of clusters are generated, within volatile or non-volatile memory, that approximate an aggregate distribution of the set of metric values on the at least one computing resource. The set of clusters is transformed, within volatile or non-volatile memory, to an aggregate piecewise approximation of a function for the set of metric values on the at least one computing resource.

Abstract: Techniques for generating distribution approximations with low memory footprints are disclosed. In some embodiments, a system receives a first set of values that measure one or more metrics of at least one computing resource. A set of clusters are generated, within volatile or non-volatile memory, that approximate a distribution of the first set of values measuring the one or more metrics of the at least one computing resource. The set of clusters is transformed, within volatile or non-volatile memory, to a piecewise approximation of a function for the first set of values.

Abstract: Techniques are described for applying what-f analytics to simulate performance of computing resources in cloud and other computing environments. In one or more embodiments, a plurality of time-series datasets are received including time-series datasets representing a plurality of demands on a resource and datasets representing performance metrics for a resource. Based on the datasets at least one demand propagation model and at least one resource prediction model are trained. Responsive to receiving an adjustment to a first set of one or more values associated with a first demand: (a) a second adjustment is generated for a second set of one or more values associated with a second demand; and (b) a third adjustment is generated for a third set of one or more values that is associated with the resource performance metric.

Abstract: Techniques are described for generating predictive alerts. In one or more embodiments, a seasonal model is generated, the seasonal model representing one or more seasonal patterns within a first set of time-series data, the first set of time-series data comprising data points from a first range of time. A trend-based model is also generated to represent trending patterns within a second set of time-series data comprising data points from a second range of time that is different than the first range of time. A set of forecasted values is generated based on the seasonal model and the trend-based model. Responsive to determining that a set of alerting thresholds has been satisfied based on the set of forecasted values, an alert is generated.

Abstract: Systems and methods are described for efficiently detecting an optimal number of behaviors to model software system performance data and the aspects of the software systems that best separate the behaviors. The behaviors may be ranked according to how well fitting functions partition the performance data.