Incidents in Generative AI Cloud Services: Comprehensive Study Reveals Fundamental Differences from Traditional Services
The rapid growth of generative artificial intelligence (GenAI), particularly large language models such as GPT-4, has revolutionized cloud services while also introducing new challenges in reliability and incident management. A comprehensive empirical study conducted by a team of researchers from Microsoft, the University of Illinois, and other prestigious institutions reveals fascinating insights into how incidents in GenAI cloud services differ fundamentally from those in traditional cloud services.
Dramatic Increase in Incidents with the Arrival of GPT Models
The study analyzed incidents in Microsoft's GenAI cloud services over the past four years, from June 2020 to February 2024. The results show a dramatic development directly associated with the market launch of individual GPT models. Before the release of GPT-3.5 in March 2022, GenAI incidents accounted for just 3% of all incidents across cloud services. After 2023, particularly following the launch of GPT-4 in March 2023, there was a significant increase—the number of incidents rose nearly tenfold compared with the GPT-3.5 era. This dramatic growth can be attributed to the global popularity of GPT models, which attracted millions of users worldwide. The trend appeared across all incident severity levels, with lower-impact incidents constituting the majority of cases. The growing popularity of services such as ChatGPT, which gained more than one million users within its first week of launch, contributed significantly to this trend.
Three Main Categories of GenAI Incident Symptoms
Researchers identified three fundamental categories of symptoms that characterize incidents in GenAI cloud services. Performance degradation is the most common problem, occurring in 49.8% of cases. This category primarily includes service-level degradation (27.2%), in which multiple APIs within a single GenAI service may fail simultaneously, affecting the service's overall availability and performance. The second type is API-level degradation (22.6%), in which a specific GenAI API experiences delays. Deployment failures account for 35.7% of incidents and reflect impacts on the continuity of GenAI services. This category includes model deployment failures (12.0%), such as when all user-tuned models cannot be deployed on time in a specific deployment region. Another significant component is resource deployment failures (14.4%), because GenAI cloud services depend heavily on the deployment of various types of resources, including computing, networking, and storage resources for processing, transmitting, and storing enormous volumes of data. Fine-tuning API failures (9.3%) represent the third type, occurring when problems arise while calling APIs for model fine-tuning. Invalid inference, although occurring in the smallest percentage of cases (14.5%), may represent the most fundamental challenge to the core functionality of GenAI services. Response quality degradation (10.7%) can lead to situations in which models generate low-quality content even in response to simple user prompts, or produce invalid content when the model is unable to understand the user's prompt. Content filter failures (3.8%) represent another critical issue, in which filters may contain errors that lead to the incorrect filtering of valid prompts or responses.
Infrastructure as the Main Cause of Problems
Analysis of the root causes of incidents revealed that infrastructure problems represent the most common source of issues, accounting for 27.2% of all cases. GenAI cloud services are built on complex hierarchical infrastructure comprising virtual machines, nodes, clusters, and data centers that host closely interconnected resources, including CPUs, memory, storage, and networks. Infrastructure maintenance problems (17.8%) primarily include failures of hardware components, such as worn-out GPUs, which can affect the fine-tuning and inference of GenAI services. Faulty GPUs, for example, may process requests incorrectly, leading to errors such as the generation of nonsensical outputs. Network problems (4.7%) include not only network bandwidth but also communication between VMs and nodes within clusters, including connection problems and DNS resolution failures. Storage problems (4.7%) are associated with managing enormous volumes of data, which requires robust storage solutions. Storage or I/O operation failures, such as data corruption or delays, can lead to service disruption. Configuration problems account for 24.5% of incidents and include incorrect configuration (13.1%), configuration updates (6.4%), and missing or incomplete configuration (5.0%). Code errors represent 21.5% of cases and include errors violating model data constraints (6.7%), content filter errors (2.2%), exception-handling errors (6.3%), and cross-system errors (6.3%). External usage problems (14.1%) and operational errors (12.7%) complete the range of causes.
Human Factors Dominate Incident Detection
One of the study's most surprising findings is the high percentage of incidents reported by people rather than automated systems. While people report only 13.7% of incidents in traditional cloud services, the figure is a significant 38.3% for GenAI services. This phenomenon can be explained by the fact that 45.9% of GenAI cloud services are still under development or in preview, while only 54.1% have reached general availability. Current monitoring systems for GenAI services often build on adaptations of existing frameworks designed for other types of cloud services, which may not yet fully meet the specific requirements of GenAI scenarios. For example, invalid inference incidents are often identified and reported by users, reflecting collaboration in the continued improvement of these systems. The study found approximately 25.9 unique monitors per 100 incidents reported by monitors for GenAI services, compared with 74.4% for other cloud services. The false-positive rate for incidents detected by monitors in GenAI services is significantly higher—11.0% compared with 6.6% for incidents detected by people. This higher rate of false positives is primarily due to the sensitivity of monitoring systems. These false alarms burden engineers with unnecessary investigations, slowing the resolution of genuine incidents.
Significantly Longer Incident Resolution Times
GenAI incidents generally require more time to mitigate compared with other types. Specifically, GenAI incidents take an average of 1.12 time units to resolve, compared with 0.65 time units for non-GenAI incidents. This longer resolution time applies across all mitigation categories, reflecting the complexity of resolving various GenAI incidents. The distribution of mitigation approaches differs significantly between GenAI and traditional services. While ad-hoc fixes account for 54.7% of all resolutions in other cloud services, they account for only 22.4% in GenAI services. The distribution of mitigation approaches for GenAI incidents is more balanced, requiring more diverse, sophisticated, and time-consuming methods than simple ad-hoc fixes. The study identified eight main categories of mitigation strategies. Self-healing (19.7%) refers to situations in which transient incidents are automatically mitigated because the service recovers on its own through its resilience mechanisms. Ad-hoc fixes (22.4%) include improvised, situation-specific actions applied to quickly mitigate symptoms. Rollback (15.2%) is a widely used strategy for incidents caused by changes, in which engineers revert those changes to a previous, stable version.
Unique Challenges of GenAI Cloud Services
GenAI cloud services introduce fundamentally different challenges compared with traditional cloud services. While traditional ML platforms such as AzureML, AWS SageMaker, and Google Cloud ML were primarily used for tasks such as model training, inference, and fine-tuning, GenAI services such as large language models rely on massive parameter scales, high hardware demands, and provide natural-language-driven applications such as text generation, summarization, and translation. These services also allow users to fine-tune models using user-uploaded datasets, exposing them to risks arising from changes in behavior at the model level. In addition, they provide intuitive conversational user interfaces, making them accessible to a broader audience while also adding complexity and risks in managing user interactions. These characteristics create new reliability problems related to model quality, privacy, and performance, layered on top of conventional reliability issues.
Implications for the Future
The study's findings offer practical insights for a wide range of stakeholders. For researchers, the study highlights several directions for future research, particularly in automated methods for detecting invalid inference results. Currently, invalid outputs, such as hallucinations or irrelevant responses, are difficult to detect. For model providers, in addition to the high rate of invalid inference results and the challenges of detecting hallucinations, another significant finding is that 38% of GenAI incidents are reported by people, reflecting insufficient development of monitoring tools. Providers should improve service observability for more effective problem detection and diagnosis. For service operators, the study reveals that the mitigation time for GenAI incidents is 1.83 times longer than for non-GenAI incidents, highlighting the need for automation in incident mitigation. The complexity of GenAI systems, which include extensive and interconnected layers of infrastructure, dependencies, and configurations, is a significant factor. This comprehensive study provides the first in-depth look at the characteristics of incidents in GenAI cloud services and reveals the unique challenges facing this rapidly growing field. As GenAI services become increasingly important to businesses and users worldwide, understanding and addressing these specific problems will be crucial to ensuring their reliability and stability in the future.
