Distributed Transactions - Transaction Processing - page 319

trigger-based method. Can updates be applied to a replica in a physical fashion, as
is done in the case of a backup database?
13.14 As explained in Sect. 13.8 , when the primary database server crashes, its
backup database server first completes the processing of all remaining log records
that have arrived before the crash, then performs the undo pass of ARIES , and only
then is ready to start normal transaction processing. Using techniques similar to
those used in centralized database recovery (Sect. 9.7 ), new transactions can be
accepted to the system before the undo pass has been completed and in fact already
before the processing of the remaining log records in the work queues has been
completed. Explain how these optimizations can be accomplished.
13.15 Assume that some subtransaction of a distributed transaction does no
updates. If the begin-transaction log record is only written at the time the trans-
action's first update is logged, no log records appear for a read-only subtransaction,
except for the prepare-to-commit and commit log records written in the execution
of the two-phase commit protocol. But are those log records really necessary for a
read-only subtransaction?
13.16 Sketch database schemas and transactions for key-value stores with entity
groups for the following Internet applications: (a) maintenance of incoming and
outgoing e-mail messages of different users and the user profiles; (b) a blogging
application in which different users can start a topic and to contribute to topics
startedbyothers.
Bibliographical Notes
The two-phase commit protocol is attributed to Gray [ 1978 ] and Lampson and
Sturgis [ 1976 ]. Many of the basic techniques and theoretical results pertaining to
transaction processing in distributed and replicated databases are presented in the
textbook by Bernstein et al. [ 1987 ]. The variant of the two-phase commit protocol
based on the presumed-abort assumption (Problem 13.4 ) and implemented in R
distributed database management system was defined by Mohan et al. [ 1986 ]. Gray
and Lamport [ 2006 ] present a commit protocol that uses multiple coordinators,
ensuring that progress is made if a majority of the coordinators are working.
The read-one-write-all and read-one-write-all-available replication protocols
are described by Bernstein et al. [ 1980 ], Rothnie et al. [ 1980 ], and Bernstein
et al. [ 1987 ]. The scalability problems with transactional replication are discussed
by Gray et al. [ 1996 ], among others. The notions of symmetric and asymmet-
ric replication are discussed by Jiménez-Peris et al. [ 2003 ], who also advocate
read-one-write-all-available replication against quorum consensus replication. The
discussion of remote backup databases in Sect. 13.8 is based on an article by Mohan
et al. [ 1993 ]. Lin et al. [ 2009 ] discuss snapshot isolation in the context of replicated
databases.