The latter alternative has also no control over the
emerging serialisation order. It can only guarantee that the transactions of a client
are serialised correctly between all other transactions, but not precisely in which
order. However, if we do not want to sacrifice correctness??”that is the system shall
guarantee one-copy serialisability??”this is not enough.
Hence, let us in the following look at an alternative approach to cluster replication
control that combines the correctness and up-to-date guarantees of distributed
transactions with the performance of asynchronous update propagation. It is using
a primary-copy replication scheme with deferred refreshment: The coordination
middleware executes updates first on the OLTP nodes (the number of update transactions
that run in parallel on the OLTP node is not restricted). After an update
transaction finishes, as soon as a refresh is activated, the refresher propagates the
changes to the remaining replicas using decoupled refresh transactions.
In more detail, each of them refreshes one node and is activated separately. Each
node guarantees locally sterilisable executions. In addition, we have to ensure
read consistency: this means to propagate refresh transactions in a way that queryonly
transactions always see the same version during their lifetime. This has to be
handled with care because as we discussed in the previous section, the router can
send each query of a read-only transaction to a different OLAP node??”routing of
queries of the same transaction to different cluster nodes is beneficial because of
caching effects.
Pages:
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460