Full-text search engines: Analysis and bencmarking of distributed text-search solutions

No Thumbnail Available

URL

Journal Title

Journal ISSN

Volume Title

School of Science |

Date

2014

Major/Subject

Tietokoneverkot

Mcode

T-110

Degree programme

Language

en

Pages

64

Series

Abstract

The amount of available data has increased notably in the last few years, exposing scalability problems of storage systems. Traditional clusters built with expensive storage solutions have proven not to be a feasible solution. The amount of investment needed to build and expand such clusters is not affordable by many companies. Commodity hardware is much cheaper but fails more often. Fault tolerance has been passed to the application layer, which allows building larger clusters with less investment thus leading to more powerful systems. However the fault tolerance mechanisms have to be taken into account when designing the application. The most common mechanisms used when implementing data storage applications is replication. Creating several copies of the same data ensures that the data is still available if there is at least one replica alive. On the other hand, replication introduces new problems. Managing replicas can be complicated when modifying existing data. It is important to make sure that all the replicas store the same version of the data. Searching in huge amounts of data requires new approaches since non-distributed text search engines are not able to return relevant documents in a reduced amount of time. Scaling a text search engine requires that the storage capabilities of the cluster can be increased horizontally and that the response time does not increase drastically as the number of computers increases. The purpose of this work is to analyse two different full-text search engines, Elastic search and Cloud era's distribution of SolrCloud. Both text search engines use Lucene, a search library written in Java, under the hood to build a text search engine. However, they manage data distribution and scaling in different manners. We have prepared benchmarks to visualize how do they behave with different setups and how does the number of available nodes influence in their search and indexing performance.

Description

Supervisor

Heljanko, Keijo

Thesis advisor

Fabra Caro, Francisco Javier

Keywords

full-text search engines, distributed systems, scalability, Cloudera, Solr, SolrCloud, elastic search, Lucene, HDFS

Other note

Citation