Logs record essential information about system operations and serve as a critical source for anomaly detection, which has generated growing research interest. Utilizing large language models (LLMs) within a retrieval-augmented generation (RAG) framework for log-based anomaly detection is an effective approach due to its strong generalization capabilities and efficient few-shot performance. However, the effectiveness of this method hinges on the quality of the knowledge source, which can be impacted by noise and changes within the software systems. Facing these problems, in this paper, we propose a novel log-based anomaly detection method named EagerLog, employing active learning to choose the logs for humans to label, thereby adding them to the knowledge source, thus enhancing the knowledge source and maintaining its quality. Our experiments on three open datasets (BGL, Thunderbird, Zookeeper) and one industrial dataset demonstrate that EagerLog can achieve 93.65% F1 score with approximately 10 labeled log sequences, surpassing existing methods by 15.32%.

Accurately diagnosing the fault that causes the failure is crucial for maintaining the reliability of a microservice system after a failure occurs. Mainstream fault diagnosis approaches are data-driven and mainly rely on three modalities of runtime data: traces, logs, and metrics. Diagnosing faults with multiple modalities of data in microservice systems has been a clear trend in recent years because different types of faults and corresponding failures tend to manifest in data of various modalities. Accurately diagnosing faults by fully leveraging multiple modalities of data is confronted with two challenges: 1) how to minimize information loss when extracting features for data of each modality; 2) how to correctly capture and utilize the relationships among data of different modalities. To address these challenges, we propose FAMOS, a Fault diagnosis Approach for MicrOservice Systems through effective multi-modal data fusion. On the one hand, FAMOS employs independent feature extractors to preserve the intrinsic features for each modality. On the other hand, FAMOS introduces a new Gaussian-attention mechanism to accurately correlate data of different modalities and then captures the inter-modality relationship with a crossattention mechanism. We evaluated FAMOS on two datasets constructed by injecting comprehensive and abundant faults into an open-source microservice system and a real-world industrial microservice system. Experimental results demonstrate the FAMOS's effectiveness in fault diagnosis, achieving significant improvements in F1 scores compared to state-of-the-art (SOTA) methods, with an increase of 20.33 %.

Log-based anomaly detection is essential for maintaining software availability. However, existing log-based anomaly detection approaches heavily rely on fine-grained exact labels of log entries which are very hard to obtain in real-world systems. This brings a key problem that anomaly detection models require supervision signals while labeled log entries are unavailable. Facing this problem, we propose a new labeling strategy called inexact labeling that instead of labeling an log entry, system experts can label a bag of log entries in a time span. Furthermore, we propose MIDLog, a weakly supervised log-based anomaly detection approach with inexact labels. We leverage the multiinstance learning paradigm to achieve explicit separation of anomalous log entries from the inexact labeled anomalous log set so as to deduce exact anomalous log labels from inexact labeled log sets. Extensive evaluation on three public datasets shows that our approach achieves an F1 score of over 85% with inexact labels.

Log parsing is a prerequisite for log analysis. Recently, large language models (LLMs) have demonstrated high accuracy in log parsing. However, their frequent invocations incur substantial costs. To address this issue, some methods have turned to small language models (SLMs), which offer improved efficiency but suffer from reduced accuracy due to limited model capacity. To achieve both high accuracy and efficiency, we propose CSLParser, a collaborative log parsing framework using SLMs and LLMs. CSLParser delegates most log parsing tasks to SLMs and selectively invokes LLMs to correct parsing results generated by SLMs, thereby effectively reducing the invocation cost of LLMs while maintaining high accuracy. Specifically, to enhance the accuracy of SLMs, we propose a diversified sampling strategy to select diverse samples for training, enabling SLMs to effectively handle diverse log patterns. To efficiently invoke LLMs, we design a rule-based selection strategy to identify hard cases that are challenging for SLMs to correctly parse, which are subsequently corrected by LLMs. Additionally, we propose a dynamic template updating mechanism that merges similar templates based on structural and semantic information to further enhance parsing accuracy. Extensive experiments on public large-scale log datasets show that CSLParser outperforms state-of-the-art baselines in both accuracy and efficiency.

Log-based anomaly detection is a essential task for ensuring the reliability and performance of software systems. However, the performance of existing anomaly detection methods heavily relies on labeling, while labeling a large volume of logs is highly challenging. To address this issue, many approaches based on transfer learning and active learning have been proposed. Nevertheless, their effectiveness is hindered by issues such as the gap between source and target system data distributions and cold-start problems. In this paper, we propose LogAction, a novel log-based anomaly detection model based on active domain adaptation. LogAction integrates transfer learning and active learning techniques. On one hand, it uses labeled data from a mature system to train a base model, mitigating the cold-start issue in active learning. On the other hand, LogAction utilize free energy-based sampling and uncertainty-based sampling to select logs located at the distribution boundaries for manual labeling, thus addresses the data distribution gap in transfer learning with minimal human labeling efforts. Experimental results on six different combinations of datasets demonstrate that LogAction achieves an average 93.01% F1 score with only 2% of manual labels, outperforming some state-of-the-art methods by 26.28%. Website: https://logaction.github.io

Log-based fault diagnosis is essential for maintaining software system availability. However, existing fault diagnosis methods are built using a task-independent manner, which fails to bridge the gap between anomaly detection and root cause localization in terms of data form and diagnostic objectives, resulting in three major issues: 1) Diagnostic bias accumulates in the system; 2) System deployment relies on expensive monitoring data; 3) The collaborative relationship between diagnostic tasks is overlooked. Facing this problems, we propose a novel end-to-end log-based fault diagnosis method, Chimera, whose key idea is to achieve end-to-end fault diagnosis through bidirectional interaction and knowledge transfer between anomaly detection and root cause localization. Chimera is based on interactive multitask learning, carefully designing interaction strategies between anomaly detection and root cause localization at the data, feature, and diagnostic result levels, thereby achieving both sub-tasks interactively within a unified end-to-end framework. Evaluation on two public datasets and one industrial dataset shows that Chimera outperforms existing methods in both anomaly detection and root cause localization, achieving improvements of over 2.92%~5.00% and 19.01% ~ 37.09%, respectively. It has been successfully deployed in production, serving an industrial cloud platform.

Log-based software reliability maintenance systems are crucial for sustaining stable customer experience. However, existing deep learning-based methods represent a black box for service providers, making it impossible for providers to understand how these methods detect anomalies, thereby hindering trust and deployment in real production environments. To address this issue, this paper defines a trustworthiness metric—diagnostic faithfulness—for models to gain service providers’ trust, based on surveys of SREs at a major cloud provider. We design two evaluation tasks: attention-based root cause localization and event perturbation. Empirical studies demonstrate that existing methods perform poorly in diagnostic faithfulness. Consequently, we propose FaithLog, a faithful log-based anomaly detection system, which achieves faithfulness through a carefully designed causality-guided attention mechanism and adversarial consistency learning. Evaluation results on two public datasets and one industrial dataset demonstrate that the proposed method achieves state-of-the-art performance in diagnostic faithfulness.

Frequent software updates lead to log evolution, posing generalization challenges for current log anomaly detection. Traditional log anomaly detection research focuses on using small deep learning models (SMs), but these models inherently lack generalization due to their closed-world assumption. Large language models (LLMs) exhibit strong semantic understanding and generalization capabilities, making them promising for log anomaly detection. However, they suffer from computational inefficiencies. To balance efficiency and generalization, we propose a collaborative log anomaly detection scheme (CoorLog) that uses an adaptive coordinator to integrate SM and LLM. The coordinator determines if incoming logs have evolved. Non-evolved logs are routed to the SM, while evolved logs are directed to the LLM for detailed inference using the constructed Evol-CoT. To gradually adapt to evolution, we introduce the adaptive evolution mechanism (AEM), which updates the coordinator to redirect evolved logs identified by the LLM to the SM. Simultaneously, the SM is fine-tuned to inherit the LLM’s judgment on these logs. Extensive experiments on real-world datasets demonstrate that CoorLog achieves superior F1-scores in both intra-version and inter-version anomaly detection. Additionally, CoorLog reduces processing time by 91.63% and token consumption by 85.59% compared to using an LLM alone.

Log-based anomaly detection is an essential task in maintaining software reliability. Existing log-based anomaly detection approaches often consist of three key phases: log parsing, event embedding, and model construction. Event embedding efficiently extracts semantic information from log events and produces vector representations of log events. However, existing event embedding methods suffer from two key problems. First, semantic noises are buried in log events leading to inevitable gaps between the obtained semantics from log events and their essential meanings. Second, there exists a gap between general semantic embedding and the specific embedding requirement of anomaly detection tasks. To mitigate these problems and improve the quality of representations of log events, we propose a novel anomaly detection approach named LLMeLog. It leverages the capabilities of large language models (LLMs) to enrich the contents of log events with in-context learning techniques. Then it utilizes the enriched log events to fine-tune a pre-trained BERT model. At last, it trains a transformer-based anomaly detection model with the event representations produced by the pre-trained BERT model. Evaluation results on three public log datasets show that LLMeLog achieves the best performance across all datasets, boasting F1-scores exceeding 99%. Besides, when using only 10% of labeled data as training data, our approach can still achieve over 90% F1-scores.

Log-based anomaly detection plays a crucial role in maintaining the reliability of software systems. Unsupervised models are more suitable for real-world usage because they do not rely on huge data labeling efforts. However, their effectiveness is limited because of the lack of supervision of data labels. To balance model effectiveness and labeling efforts, existing approaches enhance model capabilities by incorporating relatively few but key human labels as a golden signal, thereby improving the model ability with acceptable labeling efforts. However, these methods still face limitations of complex human labels and insufficient utilization of human knowledge. In this paper, we introduce LogCAE, a two-stage log anomaly detection approach based on active learning and contrastive learning. It utilizes an unsupervised model to learn from unlabeled log data without human labels and incorporates human knowledge through active learning during online optimization. We employ contrastive learning to optimize the representation of log samples in feature space for more efficient usage of human labels. We conducted experiments on three distinct public log datasets (Thunderbird, BGL, and Zookeeper). The results show that our method improves 12.93% F1-score on average with 6.06% labeled data samples. Besides, our approach is more effective in utilizing human labels than state-of-the-art approaches.

Log-based anomaly detection is an essential aspect of maintaining software reliability, particularly in the context of microservice systems. However, existing log-based anomaly detection approaches rely on historical anomalous labeled data or require huge labeling efforts. This makes existing log-based anomaly detection approaches inefficient. In this paper, we propose AcLog, a novel anomaly detection approach that incorporates human knowledge to enhance model ability based on the framework of deep active learning. It incorporates an unsupervised model to learn from normal log data rather than historical anomalous labeled data and leverages active learning to incorporate human knowledge as a golden signal to augment the quality of training log data. Experiment results on three open log datasets and one log dataset collected from a real-world microservice system show that our approach improves over 7% F1-score with 5% labeled training data on average.

Log-based anomaly detection is becoming more and more important for maintaining the availability of modern microservice systems. Existing supervised/semi-supervised log anomaly detection models require a large amount of human-labeled logs for training which are hard to collect in real-world systems. Unsupervised models often perform poorly without explicit anomaly labels. To improve the performance of unsupervised models, in this paper, we first make an empirical study of existing unsupervised models to tackle the reason why they often produce unsatisfied results. We find that anomaly detection results produced by existing unsupervised models are significantly affected by two key problems including Not-Cover (NC) problem and Suspicious-Noise (SN) problem. To solve these problems, we propose a novel augmentation framework called AFALog. AFALog leverages the idea of active learning to incorporate human knowledge so as to augment data quality. It can support almost all existing unsupervised models and improve their performance. Our experiments on two open datasets and one dataset collected from a real-world microservice system demonstrate that DALog improves the F1-score by an average of 6.61%, with only 5.9% labeled training data.