View article

Microservice applications are increasingly embracing cloud platforms to run their services. These applications can often be impacted by anomalies. Detecting anomalies at runtime is vital to ensure that cloud-native applications meet specified requirements and ensure good Quality-of-Service. However, this is challenging to do since application owners do not always have access to the underlying cloud infrastructure and hence can not use host level metrics and hardware counters as done in the past. One potential way to address this challenge is to use a machine learning based anomaly detection approach which uses metrics that can be easily collected from applications running on public cloud platforms. In this paper, we develop a classifier using deep learning for pattern recognition of anomalies for detecting runtime software anomalies in cloud-native applications. We use textured images known as spectrograms that are obtained from time series measurement readily available from cloud-native applications. These spectrogram images are used to train a Convolutional Neural Network (CNN) classifier for anomaly detection. We evaluate our approach on two real-world datasets that capture known software anomalies in public cloud platforms. Results show that the proposed spectrogram based CNN classifier yields detection accuracy of 99% for both datasets at runtime and outperforms a competing non-image based classifier.