<p>Accurately detecting anomalies in microservice systems has become increasingly important for rapidly growing microservice architectures. On the one hand, existing anomaly detection methods typically focus on a single type of data source (i.e., logs or traces) or overlook the correlations between different types of data, which can lead to missed anomalies and a high number of false positives. On the other hand, regarding the issue of imbalanced positive and negative samples in anomaly detection, existing research has not proposed effective solutions. To address these issues, in this paper, we propose MSNGAD, a microservice anomaly detection method based on nested graph diffusion reconstruction. MSNGAD uses nested graphs to uniformly describe the complex structure of traces and embedded log messages. It treats the log template temporal relationships as the inner graph of the nested graph, and the service invocations relationships as the outer graph. By capturing the correlations between different modalities of data, the system can detect anomalies comprehensively. Furthermore, to address the problem of imbalanced positive and negative samples in anomaly detection, MSNGAD trains a diffusion reconstruction model to augment the features of anomalous samples. This approach abandons traditional manual data augmentation strategies, effectively mitigating the unnatural disturbances to the graph structure caused by human intervention in data augmentation. Extensive evaluations were conducted on large-scale datasets, and the experimental results show that MSNGAD outperforms existing baseline models, achieving an F1-score of 0.96.</p>