Unifying Consensus Too Atomic Commitment For Effective Cloud Information Management (Vldb19)

Zoom DistSys Reading Group.

Atomic commitment protocols (e.g., Two Phase Commit) supply ACID guarantees for transactional access to sharded data.


Consensus protocols (e.g., Paxos) replicate information across servers inward a consistent in addition to mistake tolerant manner.

Many prior industrial plant bring observed the similarities betwixt the commitment in addition to consensus problems. Both consensus in addition to atomic commit protocols aim at ensuring that ane upshot is agreed upon inward a distributed environment. However, the atmospheric condition for achieving this understanding is dissimilar inward the 2 cases. Paxos solely needs a bulk of nodes to live available for a determination to live made whereas 2PC needs votes from all the participants to determine on the lastly value.


This newspaper proposes a framework that unites in addition to explains several commitment in addition to single-leader-based consensus protocols nether ane roof, the Consensus in addition to Commitment (C&C) framework, in addition to and hence derive protocols to demonstrate the framework. Many of these derived protocols are generalizations of existing protocols, however, about of them are novel inward their ain right.

• Paxos Atomic Commitment (PAC) is a distributed atomic commitment protocol managing sharded simply non-replicated data. PAC is a variant of 3PC, in addition to integrates crash recovery in addition to normal operations seamlessly inward a elementary Paxos-like manner. Leader election stage requires a bulk quorum response, in addition to the value regain stage requires all shards to reply unless ane of the responses had Decision value truthful or its AcceptVal value set.
• Replicated-PAC (R-PAC) is for fully replicated cloud information management, which is similar to Replicated Commit. It is a variant of Gray in addition to Lamport's Paxos Commit. R-PAC is similar to PAC above, simply since all nodes keep identical data, the leader demand to solely hold off for replies from a bulk of replicas that bring the same InitVal.
• Finally G- PAC is a novel protocol for sharded replicated architectures, which is similar to other of late proposed hybrid protocols, Janus in addition to TAPIR. G-PAC integrates transaction commitment alongside the replication of information in addition to reduces transaction commit latencies past times avoiding the unnecessary layering of commitment in addition to consensus.

The newspaper compares the performance of G-PAC alongside a Spanner-like protocol, where 2PC is used at the logical information degree in addition to Paxos is used for consistent replication of logical data. The layering of Paxos in addition to 2PC increases latency, simply combining them into a apartment hierarchy inward G-PAC way that the leader needs to communicate alongside a large release of nodes, which incurs overhead. If a transaction, T, accesses n shards in addition to each shard is replicated inward r servers, at that topographic point are a amount of n ∗ r servers that are involved inward the transaction T. Method V alongside the newly defined notions of super-set in addition to super-majority, which decides the value for the transaction, is described inward Algorithm 4. Super-set: bulk of replicas (r /2 + 1) from each of the n shards. Super-majority: bulk of replicas (r /2 + 1) from bulk of shards (n/2 + 1).

The experimental results highlight the low-latency benefits of combining consensus along alongside commitment into a unmarried integrated protocol. But it looks similar the G-PAC evaluation is non done alongside fifteen nodes (i.e., three replicas inward each of the v regions), simply solely alongside v nodes (1 inward each region) where each node uses 2 nearby nodes (within these v nodes inward v dissimilar regions) to replicate.

Share this post