Consistency Management in Cloud Storage Systems - Large Scale and Big Data: Processing and Management - page 345

10.6 HARMONY: AUTOMATED SELF-ADAPTIVE CONSISTENCY
Most of the existing adaptive consistency approaches require a global knowledge
of the application access pattern (e.g., consistency rationing requires the data to be
categorized in advance). However, with the tremendous increase in data size along
with the significant variation in the service's load, this task is hard to accomplish
and will add an extra overhead to the system. Moreover, these approaches cover a
small set of applications where operation orderings are strictly required. In this con-
text, we introduce our approach Harmony [11], an automated self-adaptive approach
that considers applications tolerance rate for stale reads (i.e., Harmony complements
other adaptive approaches as it targets applications with stale reads consideration
rather than operation orderings, see Table 10.3).
Rather than relying on a standard model based only on the access pattern to define
the consistency requirement of an application—which is the case for most existing
work—Harmony, instead, uses the stale read rate of the application to precisely
define such requirement: by doing so, Harmony retains the same rate of stale reads
regardless the variation in the access loads and patterns of an application or the
changes on the network latency of the system. Consequently, Harmony embraces an
intelligent estimation model to automatically identify the key parameters affecting
the stale reads such as the system states (network latency) and application's require-
ments (current access pattern). Harmony, therefore, elastically scales up/down the
number of replicas involved in read operations to maintain a low (possibly zero) tol-
erable fraction of stale reads, hence, improving the performance of the applications
while meeting the desired consistency level.
In the rest of this section, we will first present design trend in Harmony of using
stale reads rate to define the consistency requirement of an application. Then, we
discuss how to estimate the amount of stale reads in the system. We then describe
the Harmony implementation, and integration into Cassandra cloud storage system
[34] and present detailed results of experimental evaluations on Amazon EC2 [3].

TABLE 10.3
Adaptive Consistency Approaches

Cloud Storage
System:
Implemented
Within

Consistency
Specification
Level

Testbed for
Evaluating the
Solution

Automated
Consistency

Tradeoff

RedBlue
Consistency
[8]

Operations Gemini Amazon EC2 in

different
availability
zones

No Consistency-

performance

Consistency
Rationing
[47]

Data Amazon S3 Amazon S3 Ye s Consistency-

cost

Harmony
[11]

Operations Apache

Cassandra

Grid'5000 and
Amazon EC2

Ye s Consistency-

performance