Predictive Models for Consistency Index of a Data Object in a Replicated Distributed Database System

Consistency is a qualitative measure of database performance. Consistency Index (CI) is a quantification of consistency of a data unit in terms of percentage of correct reads to the total reads observed on the data unit in given time. The consistency guarantee of a replicated database logically depends on the number of reads, updates, number of replicas, and workload distribution over time. The objective of our work is to establish this dependency and finding their level of interactions with consistency guarantees to develop a predictor model for CI. We have implemented Transaction Processing Council-C (TPCC) online transactions benchmark on Amazon SimpleDB which is used as big-data storage. We have controlled the database design parameters and measured CI with 100 samples of workload and database design. The findings helped us to implement a prototype of CI based consistency predictor using statistical predictive techniques like a) Regression model and b) Multiple Perceptron neural network model c) Hidden Markov model. The data statistics show that the neural network based CI predictor causes less error and results in better coefficient of determination R and mean square error (MSE).The Hidden Markov model based CI predictor is capable of modelling the effect of sequence of the input workload on the probability of obtaining a desired CI. Key-Words: Consistency Index (CI), predictive models, regression, neural network, Hidden Markov model