Group communication is a middleware technology that may simplify development of highly available applications through replication. In this paper we propose a replicated service implementation based on partitionable group communication. Unlike its primary-partition counterpart, a partitionable group communication system never blocks an application because network failures may have caused it to partition. Modern Internet-based services with workloads that are typically biased in favor of read-only requests can exploit this property and remain available in multiple partitions despite the presence of replicated state. The proposed service offers several interesting features including sequential consistency, load balancing, and "atmost -once" semantics for updates despite unreliable communication between clients and servers. In our implementation, clients may simply resubmit updates to any available server if they suspect that the original request may have been lost due to a broken TCP connection. The possibility to recover from communication failures by having clients reissue requests to servers without incurring the risk of multiple executions is an increasingly important requirement for many Internet-based services. The service makes use of novel abstractions for partitionable group communication that may be interesting in their own right.