SQL-il tuginevate skriptimiskeelte kasutamine andmeanalüütikaks Hadoopi ökosüsteemis
Date
2016
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Selle lõputöö eesmärk on andmeanalüütika algoritmide rakendamine,\n\ret võrrelda erinevaid SQL-il põhinevaid skriptimiskeeli Hadoopi ökosüsteemis.\n\rLõputöö võrdleb erinevate raamistike efektiivsust ja algoritmide implementeerimise\n\rlihtsust kasutajal, kellel pole varasemat hajusarvutuse kogemust. Eesmärgi\n\rtäitmiseks implementeeriti kolm algoritmi: Pearsoni korrelatsioon, lihtne lineaarne\n\rregressioon ja naiivne Bayesi klassifikaator. Algoritmid implementeerti kahes\n\rSQL-il põhinevas raamistikus: Spark SQL-s ja HiveQL-s, samuti implementeeriti\n\rsamade algoritmide Spark MLlibi versioon. Algoritme testiti klastris erinevate sisendfaili\n\rsuurustega, samuti muudeti kasutatavate tuumade arvu. Selles lõputöös\n\ruuriti ka Spark SQLi ja Spark MLlibi algoritmide skaleeruvust. Algoritmide jooksutamise\n\rtulemusel selgus, et Pearsoni korrelatsioon oli HiveQL’is veidi kiirem kui\n\rteistes uuritud raamistikes. Lineaarse regressiooni tulemused näitavad, et Spark\n\rSQL ja Spark MLlib olid selle algoritmiga sama kiired, HiveQL oli umbes 30%\n\raeglasem. Kahe esimese algoritmiga skaleerusid Spark SQL ja Spark MLlibist pärit\n\ralgoritm hästi. Naiivse Bayesi klasifikaatoriga tehtud testid näitasid, et Spark\n\rSQL on selle algoritmiga kiirem kui HiveQL, hoolimata sellest, et ta ei skallerunud\n\rhästi. Spark MLlibi tulemused selle algoritmiga ei olnud piisavad järelduste\n\rtegemiseks. Korrelatsiooni ja lineaarse regressiooni implementatsioonid HiveContextis\n\rja SQLContextis andsid sama tulemuse. Selle lõputöö käigus leiti, et SQL-il\n\rpõhinevaid raamistikke on kerge kasutada: HiveQL oli kõige lihtsam samas kui\n\rSpark SQL nõudis veidi hajusarvutuse tundma õppimist. Spark MLlibi algoritmide\n\rimplementeerimine oli raskem kui oodatud, kuna nõudis algoritmi sisemise töö\n\rmõistmist, samuti osutusid vajalikuks teadmised hajusarvutusest.
The goal of this thesis is to compare different SQL-based scripting languages\n\rin Hadoop ecosystem by implementing data analytics algorithms. The thesis compared framework efficiencies and easiness of implementing algorithms with no previous\n\rexperience in distributed computing. To fulfill this goal three algorithms were\n\rimplemented: Pearson’s correlation, simple linear regression and naive Bayes classifier.\n\rThe algorithms were implemented in two SQL-based frameworks on Hadoop\n\recosystem: Spark SQL and HiveQL, algorithms were also implemented from Spark\n\rMLlib. SQLContext and HiveContext were also compared in Spark SQL. Algorithms\n\rwere tested in a cluster with different dataset sizes and different number of\n\rexecutors. Scaling of Spark SQL and Spark MLlib’s algorithm was also measured.\n\rResults obtained in this thesis show that in the implementation of Pearson’s correlation\n\rHiveQL is slightly faster than other two frameworks. Linear regression\n\rresults show that Spark SQL and Spark MLlib are with similar run times, both\n\rabout 30% faster than HiveQL. Spark SQL and Spark MLlib algorithms scaled\n\rwell with these two algorithms. In the implementation of naive Bayes classifier\n\rSpark SQL did not scale well but was still faster than HiveQL. Results for Spark\n\rMLlib in multinomial naive Bayes proved to be inconclusive. With correlation\n\rand regression no difference between SQLContext and HiveContext was found.\n\rThe thesis found SQL-based frameworks easy to use: HiveQL was the easiest\n\rwhile Spark SQL required some additional investigation into distributed computing.\n\rImplementing algorithms from Spark MLlib was more difficulty as there it\n\rwas necessary to understand the internal workings of the algorithm, knowledge of\n\rdistributed computing was also necessary.
The goal of this thesis is to compare different SQL-based scripting languages\n\rin Hadoop ecosystem by implementing data analytics algorithms. The thesis compared framework efficiencies and easiness of implementing algorithms with no previous\n\rexperience in distributed computing. To fulfill this goal three algorithms were\n\rimplemented: Pearson’s correlation, simple linear regression and naive Bayes classifier.\n\rThe algorithms were implemented in two SQL-based frameworks on Hadoop\n\recosystem: Spark SQL and HiveQL, algorithms were also implemented from Spark\n\rMLlib. SQLContext and HiveContext were also compared in Spark SQL. Algorithms\n\rwere tested in a cluster with different dataset sizes and different number of\n\rexecutors. Scaling of Spark SQL and Spark MLlib’s algorithm was also measured.\n\rResults obtained in this thesis show that in the implementation of Pearson’s correlation\n\rHiveQL is slightly faster than other two frameworks. Linear regression\n\rresults show that Spark SQL and Spark MLlib are with similar run times, both\n\rabout 30% faster than HiveQL. Spark SQL and Spark MLlib algorithms scaled\n\rwell with these two algorithms. In the implementation of naive Bayes classifier\n\rSpark SQL did not scale well but was still faster than HiveQL. Results for Spark\n\rMLlib in multinomial naive Bayes proved to be inconclusive. With correlation\n\rand regression no difference between SQLContext and HiveContext was found.\n\rThe thesis found SQL-based frameworks easy to use: HiveQL was the easiest\n\rwhile Spark SQL required some additional investigation into distributed computing.\n\rImplementing algorithms from Spark MLlib was more difficulty as there it\n\rwas necessary to understand the internal workings of the algorithm, knowledge of\n\rdistributed computing was also necessary.