This paper presents a formal framework for programming distributed applications capable of handling partial failures, motivated by the non-trivial interplay between failure handling and messaging in asynchronous distributed environments. Multiple failures can affect protocols at the level of individual interactions (alignment). At the same time, only participants affected by a failure or involved in its handling should be informed of it, and its handling should not be mixed with that of other failures (precision). This is particularly challenging, as through the structure of protocols, failures may be linked to others in subsequent or concomitant interactions (causality). Last but not least, no central authority should be required for handling failures (decentralisation). Our goal is to give developers a description language, called protocol types, to specify robust failure handling that accounts for alignment, precision, causality, and decentralisation. A type discipline is built to statically ensure that asynchronous failure handling among multiple endpoints is free from orphan messages, deadlocks, starvation, and interactions are never stuck.