Faster recovery of distributed stream processing services through coordinated data placement of stream state checkpoints

Fault-tolerant distributed stream-processing systems (DSPS) typically requirethe use of a distributed data store (DDS) to maintain remote state snapshotsthat can be used to survive failures of nodes that are hosting DSPS tasks. SuchDDSs are independent systems that use their own data placement and replicationpolicies. As a result, recovering DSPS state for a task generally requirestransferring data over the network to the recovery node where that task will beplaced, as well as copying data between les within that node. In this paper,we propose coordination between DSPS and DDS by which DDS placement andreplication policies are aware of and take into account DSPS data semanticsand alignment of the location of remote state with nodes that will host recoverytasks, avoiding network transfers during recovery. We additionally proposeoptimizations to reduce data movement costs within the recovery node. We implementour proposed improvements in the Flink DSPS and Hadoop File SystemDDS and demonstrate experimentally that they jointly result to up to ~4.3xfaster recovery time compared to the baseline (unaligned) implementations.