TimeCloth: Fast Point-in-Time Database Recovery in The Cloud

In recent years, we have noted the frequent occurrence of user-triggered database recoveries in the cloud. In contrast to traditional failure-triggered recovery scenarios, they come with unique consumer-centric challenges. Unfortunately, current solutions prove inefficient, falling short either in meeting our customers' requirements or due to their close integration with the native recovery support of the underlying database engines. In this work, we present TimeCloth, a generic cloud-native recovery mechanism that achieves sublinear recovery time while meeting the specific needs of our customers. It comprises a recovery module optimized for fine-grained point-in-time recoveries and an import module enabling nearly instantaneous access to remote tables. The recovery module performs fast log filtering, parallelizes replay of non-conflicting log records and coalesce log records to reduce the volume of replay work. The import module implements a transparent FUSE-based lazy loading mechanism as well as a smart prefetcher to achieve good access performance for remote tables. Collectively, they significantly accelerate user-triggered recoveries in the cloud. TimeCloth is launched in production at Alibaba Cloud for about 15 months. We have witnesses a reduction in RTO among our customers by 44% on average and sometimes up to 92%.