Consolidating Concurrency Command Together With Consensus For Commits Nether Conflicts

TAPIR had investigated this query first, together with Janus does a followup. Instead of using the full general transaction model considered inward TAPIR, Janus uses i shot transactions, where each transaction contains multiple pieces together with each slice has i read/write functioning on a unmarried key. 

This reduced mightiness transaction model allows Janus to implement an abort-free transaction ordering capability. A novel insight inward Janus is that it notices it is possible to consolidate replication ordering together with transaction ordering using the same setup. To guild the transactions together with replication alike, Janus takes a dependency graph tracking approach, similar MDCC together with EPaxos. It orders replication together with transaction using dependency graph information inward the same deterministic way: commence perform a topological sort, together with the purpose transaction ids to intermission ties inward partial equivalence classes amongst cyclic dependencies. 

Since Janus tin procedure these reduced mightiness transactions inward an abort-free manner, it tin commit transactions inward 1 circular trip fourth dimension (RTT) when in that place is no conflict together with at most inward 2 RTTs otherwise. Avoiding aborts agency that Janus provides high throughput together with skillful latency benefits particularly inward broad expanse network setups. 

The Algorithm

Clients post their requests to the nearest coordinator, which becomes responsible for ordering together with coordinating the transaction. Janus has pre-accept together with commit phases inward the fast path (1 RTT commits!) together with needs bring stage inward illustration of conflicts (resulting inward 2 RTT commits). 

The SLOG work (from VLDB 2019) decentralized the centralized transaction ordering approach inward Calvin. SLOG has closed to prissy ideas most adjusting the ownership of keys together with improving the locality together with latency of commits. As for multihome transactions, SLOG relies on deterministic processing together with a locking based solution to avoid 2 stage commit. Once all parties lead maintain to the plan, processing occurs (mostly) independently on each node, amongst the organization relying on the plan's determinism inward guild to avoid replica divergence. 

There is too Ocean Vista's visibility tracking approach, which I hash out inward betoken three below.

At this point, nosotros receive got seen easily a dozen of these low-latency WAN transaction approaches. It would hold out prissy to create a graph of all these systems together with relate them to amongst honour to the approaches they share, together with how they influenced each other. 

Discussion betoken 3: Ocean Vista comparison

Ocean Vista is closed to other WAN transaction processing newspaper nosotros latterly discussed. 

Ocean Vista (OV) makes everything thence simple. First the transaction T is replicated to all the parties across the datacenters. The parties thence cheque to run across that the watermark rises higher upward T to ensure that all transactions including together with preceding T has been replicated successfully to all the parties. Then, the execution of T tin hold out done asynchronously at each party.

Since OV is simple, it tin handgrip replica together with information oculus failures easily, inward contrast to Janus. In OV, 1 RTT is plenty (modulo the additional latency for gossiping) fifty-fifty nether conflicts. And OV supports a slightly to a greater extent than full general version of transactions inward Janus, but withal to a greater extent than express than the transactions inward TAPIR.

The pick out lead maintain of inward OV is that since in that place is exclusively i visibility watermark for the system, a transaction needs to await for whatsoever before transaction fifty-fifty if in that place is no conflict. Janus ensures that you lot exclusively await for subject transactions, which provides lower latency for independent transactions, but this comes at the toll of tracking every dependency together with complicating the protocol together with recovery.

Finally, OV uses a crotch: multiversioning. Janus doesn't cry for that.

Discussion betoken 4: Atlas comparison

Another protocol nosotros latterly discussed is Atlas. Atlas provides efficient fast Paxos over WAN, without aborts, but past times reordering respecting the dependency graphs. Atlas is but consensus, together with Janus is transaction together with consensus. But of course, Janus transactions are express inward orbit to one-shot transactions, thence they may non hold out truly that far apart. 

Compared to Janus, I recollect Atlas made the fault-handling much clearer. It is withal complicated of course of pedagogy because of the dependencies involved inward fast quorums, but at to the lowest degree Atlas made the procedure explicit together with set it out clearly. 

Evaluation 

The code for Janus is opened upward source at https://github.com/NYU-NEWS/janus

The evaluation results exhibit that Janus provides depression latency together with skillful throughput fifty-fifty amongst many conflicts.

The evaluation results too exhibit that Janus has pocket-size throughput overhead when in that place are no conflicts. As nosotros mentioned earlier, the newspaper should receive got discussed how to address dissimilar mistake scenarios together with render evaluation nether straggler node together with crashed node scenarios every bit well. 

Share this post