RanKloud: a scalable ranked query processing framework on hadoop.
ABSTRACT The popularity of batch-oriented cluster architectures like Hadoop is on the rise. These batch-based systems successfully achieve high degrees of scalability by carefully allocating resources and leveraging opportunities to parallelize basic processing tasks. However, they are known to fall short in certain application domains such as large scale media analysis. In these applications, the utility of a given data element plays a vital role in a particular analysis task, and this utility most often depends on the way the data is collected or interpreted. However, existing batch data processing frameworks do not consider data utility in allocating resources, and hence fail to optimize for ranked/top-k query processing in which the user is interested in obtaining a relatively small subset of the best result instances. A naïve implementation of these operations on an existing system would need to enumerate more candidates than needed, before it can filter out the k best results. We note that such waste can be avoided by utilizing utility-aware task partitioning and resource allocation strategies that can prune unpromising objects from consideration. In this demonstration, we introduce RanKloud, an efficient and scalable utility-aware parallel processing system built for the analysis of large media datasets. RanKloud extends Hadoop's MapReduce paradigm to provide support for ranked query operations, such as k-nearest neighbor and k-closest pair search, skylines, skyline-joins, and top-k join processing.
- SourceAvailable from: Sameer Singh
Conference Paper: Parallel Large Scale Feature Selection for Logistic Regression[Show abstract] [Hide abstract]
ABSTRACT: In this paper we examine the problem of ecient feature evaluation for logistic regression on very large data sets. We present a new forward feature selection heuristic that ranks features by their estimated eect on the resulting model's performance. An approximate optimization, based on backtting, provides a fast and accurate estimate of each new feature's coecient in the logistic regression model. Further, the algorithm is highly scalable by parallelizing simultaneously over both features and records, allowing us to quickly evaluate billions of potential features even for very large data sets.Proceedings of SIAM International Conference on Data Mining (SDM'09); 01/2009
- [Show abstract] [Hide abstract]
ABSTRACT: Assume that each object in a database has m grades, or scores, one for each of m attributes. For example, an object can have a color grade, that tells how red it is, and a shape grade, that tells how round it is. For each attribute, there is a sorted list, which lists each object and its grade under that attribute, sorted by grade (highest grade first). Each object is assigned an overall grade, that is obtained by combining the attribute grades using a fixed monotone aggregation function, or combining rule, such as min or average. To determine the top k objects, that is, k objects with the highest overall grades, the naive algorithm must access every object in the database, to find its grade under each attribute. Fagin has given an algorithm (“Fagin's Algorithm”, or FA) that is much more efficient. For some monotone aggregation functions, FA is optimal with high probability in the worst case. We analyze an elegant and remarkably simple algorithm (“the threshold algorithm”, or TA) that is optimal in a much stronger sense than FA. We show that TA is essentially optimal, not just for some monotone aggregation functions, but for all of them, and not just in a high-probability worst-case sense, but over every database. Unlike FA, which requires large buffers (whose size may grow unboundedly as the database size grows), TA requires only a small, constant-size buffer. TA allows early stopping, which yields, in a precise sense, an approximate version of the top k answers. We distinguish two types of access: sorted access (where the middleware system obtains the grade of an object in some sorted list by proceeding through the list sequentially from the top), and random access (where the middleware system requests the grade of object in a list, and obtains it in one step). We consider the scenarios where random access is either impossible, or expensive relative to sorted access, and provide algorithms that are essentially optimal for these cases as well.Journal of Computer and System Sciences 01/2001; 66(4-66):614-656. · 1.09 Impact Factor
- [Show abstract] [Hide abstract]
ABSTRACT: This paper explores the problem of computing pairwise sim- ilarity on document collections, focusing on the application of \more like this" queries in the life sciences domain. Three MapReduce algorithms are introduced: one based on brute force, a second where the problem is treated as large-scale ad hoc retrieval, and a third based on the Cartesian product of postings lists. Each algorithm supports one or more approx- imations that trade eectiveness for eciency, the charac- teristics of which are studied experimentally. Results show that the brute force algorithm is the most ecient of the three when exact similarity is desired. However, the other two algorithms support approximations that yield large ef- ciency gains without signicant loss of eectiveness.Proceedings of the 32nd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2009, Boston, MA, USA, July 19-23, 2009; 01/2009
RanKloud: A Scalable Ranked Query Processing
Framework on Hadoop
K. Selçuk Candan
CIDSE, Arizona State
Tempe, AZ, 85287, USA
CIDSE, Arizona State
Tempe, AZ, 85287, USA
CIDSE, Arizona State
Tempe, AZ, 85287, USA
CIDSE, Arizona State
Tempe, AZ, 85287, USA
The popularity of batch-oriented cluster architectures like
Hadoop is on the rise. These batch-based systems success-
fully achieve high degrees of scalability by carefully allocat-
ing resources and leveraging opportunities to parallelize ba-
sic processing tasks. However, they are known to fall short in
certain application domains such as large scale media analy-
sis. In these applications, the utility of a given data element
plays a vital role in a particular analysis task, and this util-
ity most often depends on the way the data is collected or
interpreted. However, existing batch data processing frame-
works do not consider data utility in allocating resources,
and hence fail to optimize for ranked/top-k query process-
ing in which the user is interested in obtaining a relatively
small subset of the best result instances. A na¨ ıve implemen-
tation of these operations on an existing system would need
to enumerate more candidates than needed, before it can
filter out the k best results. We note that such waste can
be avoided by utilizing utility-aware task partitioning and
resource allocation strategies that can prune unpromising
objects from consideration. In this demonstration, we intro-
duce RanKloud, an efficient and scalable utility-aware par-
allel processing system built for the analysis of large media
datasets. RanKloud extends Hadoop’s MapReduce paradigm
to provide support for ranked query operations, such as k-
nearest neighbor and k-closest pair search, skylines, skyline-
joins, and top-k join processing.
∗This work is partially funded by the HP Labs Innovation
Research Program Grant “Data-Quality Aware Middleware
for Scalable Data Analysis.”
†Authors are listed in an alphabetical order.
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
EDBT 2011, March 22–24, 2011, Uppsala, Sweden.
Copyright 2011 ACM 978-1-4503-0528-0/11/0003 ...$10.00
Categories and Subject Descriptors
H.2 [Database Management]: Miscellaneous
MapReduce, Top-K, Skyline, KNN, Parallel Processing
A significant class of data processing applications that are
easy to describe, and in many cases, easy to parallelize can
be expressed in terms of a small set of primitives. This has
led to frameworks such as MapReduce [5, 1], Scope , etc.
These systems have been successfully applied in domains
such as data processing, mining, and information retrieval
[6, 8, 11]. Given an atomic task, these rely on the simple
semantic properties of the task to partition the work onto
many machines. More complex data processing tasks are
represented as workflows. Significant savings in execution
times are obtained by independently parallelizing each step
of the workflow and executing them in a batched manner.
Frameworks such as Hadoop , which is based on Google’s
MapReduce , have produced impressive results in many
data processing and analysis application domains [6, 8, 11].
On the other hand, in others domains (including those in-
volving aggregation and join tasks, which are hard to paral-
lelize) these batch processing systems lag behind traditional
DBMS solutions .
A particular shortcoming of batch-based data processing
frameworks is that they are not effective in domains where
the utility of the data elements to a particular task varies
from data instance to data instance, and users are inter-
ested not in all the data, but the ones that are best suited
for the given task. Locating such high utility data is often
known as ranked/top-k query processing. The applications
in which the data utilities vary include decision support, and
text and media analysis. In text analysis, for example, the
TF-IDF values or the frequency of the words can be consid-
ered as the utility scores of the data. While batch-oriented
Node 1Node 2Node N …….
4. Write sample1,
3. Read Merged
(a) Sample module for dataset1
Node 1Node 2Node N…….
4. Write sample2,
3. Read Merged
(b) Sample module for dataset2
3. Write Selectivity
2. Read sample1, sample2,
(c) CalcSelectivity module
3. Write Theta
Node 1Node 2Node N…….
2. Read Selectivity ,
(d) ThetaCompute module
2. Read Theta,
(e) BuildIndex module
2. Read Theta,
(f) Operator module
Figure 1: RanKloud’s execution engine: RanKloud leverages and extends the open-source Hadoop infrastructure
systems promise large scale parallelism, they are unable to
optimize for ranked query processing since they do not con-
sider variations in data utility. In order to avoid wasted
work, these data processing systems need to employ utility-
aware task partitioning and resource allocation strategies
that can prune unpromising objects from consideration.
Motivated by the above observations, this paper introduces
RanKloud, an efficient and scalable, utility-aware, parallel
processing system developed for the analysis of large scale
media datasets. RanKloud builds on Hadoop’s MapReduce
paradigm and provides: (1) adaptable, utility and rank-
based data processing (map, reduce and merge) primitives,
(2) waste and unbalance-avoidance strategies for incremen-
tal batched processing, and for utility-aware data partition-
ing and resource allocation, and (3) media processing work-
flow scheduling based on the data and utility characteristics
discovered at runtime. In particular, we demonstrate uS-
plit1, a data partitioning strategy that we developed for
processing top-k join queries in batch-oriented cluster envi-
ronments. We show how uSplit adaptively samples datasets
in order to allocate resources in a work-balanced and wasted-
work avoiding manner for top-k join processing. In addition
to uSplit, we also demonstrate efficient implementations
of the skyline and k-nearest neighbor search operators. Our
future research involves extending RanKloud to support join-
based skyline and k-closest pair queries.
The rest of the paper is structured as follows: in Section 2,
we give an overview of the existing work. Section 3 presents
the overall system architecture of RanKloud. In Section 4,
we discuss the demonstration scenarios. Lastly, we conclude
and give a sketch of future research directions in Section 5.
1The details of this work can be found in 
A top-k query can always be processed by simply enumer-
ating all results and then picking the best k. But when the
number of candidate results is large, and when the k is much
smaller than the number of candidate results, it is wasteful
to rely on processing strategies that would require the sys-
tem to enumerate all possible candidate results. Instead,
for ranked processing operations such as k-nearest neigh-
bor search, k-nearest neighbor and k-best-nearest neighbor
joins, skylines and top-k skylines, top-k joins, top-k group
and top-k co-group operations, data processing systems need
to employ data structures and algorithms that can prune un-
promising data objects from consideration, without having
to evaluate them.
Most existing algorithms that implement the above oper-
ations assume that one or both of the following data ac-
cess strategies are available: (a) sorted access – the system
is able to access the input data incrementally, in the non-
increasing order of utility scores. For example, top-k join
algorithms such as FA, TA, and NRA  assume that the in-
put data is available in utility sorted order – pipelined access
to sorted data helps identify candidate matches and prune
non-promising results faster; (b) indexed/clustered access
– for any candidate object, the system is able to quickly
identify the corresponding utility score by using an index
structure for efficient utility score look up. The index struc-
ture also helps identify similar objects quickly and prune
irrelevant objects faster. For instance, algorithms to tackle
k-nearest neighbor and k-closest pair queries assume that
the data is ordered using space filling curves  or par-
titioned using index structures (such as R-trees, KD-trees,
and Voronoi diagrams)  or hash functions . Skyline
algorithms also include sorting-based and index-based algo-
rithms . However, implementing these operators on a
bin/hadoop jar TopKoperator.jar DataAnalyzer.Sample input_params(dataset1 JoinAttr1 UtilityAttr1 dataset1_distribution SampleBudget1 numPartitions1 rangeScale1 sampleScale1 sampleType1)
output_params(segmentCount1 histogram1 sample1)
bin/hadoop jar TopKoperator.jar DataAnalyzer.Sample input_params(dataset2 JoinAttr2 UtilityAttr2 dataset2_distribution SampleBudget2 numPartitions2 rangeScale2 sampleScale2 sampleType2)
output_params(segmentCount2 histogram2 sample2)
bin/hadoop jar TopKoperator.jar DataAnalyzer.CalcSelectivity input_params(segmentCount1 segmentCount2 sample1 sample2 numPartitions1 numPartitions2)
bin/hadoop jar TopKoperator.jar DataAnalyzer.ThetaCompute input_params(Selectivity segmentCount1 segmentCount2 numPartitions1 numPartitions2 rangeScale1 rangeScale2 K mergeFunction)
bin/hadoop jar TopKoperator.jar Operator.BuildIndex input_params(dataset2 JoinAttr2 UtilityAttr2 Theta)
bin/hadoop jar TopKoperator.jar Operator.TopKjoin input_params(dataset1 JoinAttr1 UtilityAttr1 histogram1 histogram2 indexDataset2 K mergeFunction Theta)
Figure 2: An example of a user-defined query script
system such as Hadoop that supports high degrees of paral-
lelism requires further care.
Hadoop’s MapReduce framework can perform joins between
large datasets using either a map-side join or a reduce-side
join, but writing the code to do joins from scratch is fairly
involved. So rather than writing MapReduce programs, one
can opt for higher-level frameworks such as Pig , Hive,
or Cascading in which join operations are a core part of
the implementation. However, these frameworks fail to ef-
ficiently support rank-based query processing – each oper-
ator processes most of its input data, and top-k selection
involves filtering after a large number of candidates have
been produced. Time and resources are wasted in produc-
ing unnecessary, low utility results; thus, most of the work
done is wasted.Hence, via the RanKloud framework, we
aim to tackle the scalability challenges posed by large scale
media analysis applications that existing batched processing
systems fail to handle.
Figure 1 illustrates the architecture of the RanKloud frame-
work. The execution engine, which consists of the Sample,
CalcSelectivity, ThetaCompute, BuildIndex and the Op-
erator modules, is incorporated into the existing Hadoop
infrastructure. These modules drive the entire process of
executing a rank-based query submitted by the user. Along
with these modules, we integrate into Hadoop a Berke-
leyDB2component in order to provide a way to index
datasets during query processing. The query script is a
user-defined UNIX bash script that is sent to the execution
engine of RanKloud, which then invokes a MapReduce job
on the Hadoop cluster. The query result is obtained by ac-
cessing Hadoop’s Distributed File System (HDFS). Below,
we use the top-k join operator as an example to describe the
details of the architecture.
As discussed earlier, the query script is a user-defined
UNIX bash script. This gives the user the flexibility to de-
scribe a series of operations that are applied to the input
data to produce the output. The user can decide which of
RanKloud’s modules need to be called in order to execute
his/her query. Taken as a whole, the operations describe a
Berkeley DBJava Edition3.3.87–
workflow, which the RanKloud execution engine translates
into a series of MapReduce jobs.
Figure 2 shows an example query script that executes a top-
k join operation on two datasets. This script causes Ran-
Kloud’s execution engine to: (1) invoke the Sample and
the ThetaCompute modules to collect the necessary statis-
tics of the datasets, and then, (2) the TopKjoin operator is
called to obtain the k best join results.
One major difficulty we need to deal with when executing
complex media analysis workflows is that the statistics of
the intermediary, transient data within the workflow are not
available in advance. RanKloud alleviates this problem by
collecting any information needed to estimate data statis-
tics as a function of the data utility. The uSplit method
described in  is, therefore, essential in achieving waste-
avoiding ranked query processing.
Sampling and Lower Bound Computation
The uSplit sampling approach is implemented through the
following modules: (a) the Sample module – based on the
sampling budget, number of partitions of the utility space,
range scale factor, sample scale factor, input data utilities,
data distribution, and choice of either map-side or reduce-
side sampling, this module randomly picks sample tuples
from each of the datasets. It outputs the sample tuples and
histograms of the datasets; (b) the CalcSelectivity module
– this module estimates the join selectivity of the datasets
as a function of the histograms and the sample tuples picked
by the Sample module; (c) the ThetaCompute module – this
module computes the lower bound, Θk, based on the join
selectivity estimated by the CalcSelectivity module, the
merge function, and K (the number of top results). The
histograms and the Θk value generated by these modules
are sent as inputs to RanKloud’s Operator module.
The Utility-aware Partitioner explained in  is a
custom partitioner that is integrated into the existing
Hadoop infrastructure to enable top-k join queries to be
processed in parallel. This component forms a part of the
Operator module, and its main function is to estimate
the amount of work that needs to be done to find the
top-k results. The partitioner estimates the join work by
evaluating the histograms produced by the Sample module,
and based on this, it creates work-balanced partitions of the
input data that are sent to the servers for join processing.
If the histograms preexist, then the user has the option of
reusing them for future query processing.
Our implementation of the top-k join operator has the fol-
lowing modules: (a) the BuildIndex module – this module
is connected to the BerkeleyDB component to enable each
server to create the necessary index structure in order to
support efficient join processing. If an index exists, then the
user has the option of reusing it; (b) the TopKjoin module –
each server calls this module in order to execute its portion
of the join task based on the the merge function, K, and the
lower bound, Θk. The results from the individual servers
are then combined to select the top-k results.
Top-K Join Operator
This section describes the demo set up and the scenarios. We
use real (IMDB3and Flickr images4) and synthetic datasets
of varying sizes and data distributions. RanKloud is inte-
grated into Hadoop running on the Ubuntu operating sys-
tem. This demonstration presents three possible scenarios.
This scenario demonstrates how RanKloud’s uSplit ap-
proach is designed to handle joins between two datasets.
A complete user-defined query, similar to the one shown
in Figure 2, is executed to show how this workflow is exe-
cuted on our framework. We will compare our technique to
Hadoop and Pig-Latin join algorithms to demonstrate the
efficiency of our method. Lastly, we will show how uSplit’s
data statistics and indexes can be reused in RanKloud for
top-k join queries that are executed repeatedly.
Two-way Top-k Join Processing
Through this scenario, we demonstrate how the uSplit par-
titioning strategy is extended to handle joins between mul-
tiple datasets. The sampling and lower bound computation
modules have the ability to collect the necessary data statis-
tics and estimate the lower bound, Θk, even for high dimen-
sional join processing. The top-k join operator is extended
to efficiently process joins between multiple datasets based
on the statistics it receives.
Multi-way Top-k Join Processing
In this scenario, we show RanKloud’s ability to handle skyline
and k-nearest neighbor search operations on large datasets.
Parallel versions of these operators are implemented to build
a complete ranked query processing framework. Our future
research goals involve extending RanKloud to support join-
based skyline and k-closest pair queries. While the overall
adaptive approach is similar to the top-k join operator, the
specifics of work partitioning and pruning strategy will differ
from one operator to another.
Other Ranked Processing Operators
This demonstration introduces a new data processing frame-
work on Hadoop called RanKloud. RanKloud supports ranked
query operations on large media datasets in a waste-avoiding
manner by treating the utilities of the data to be an integral
part of the analysis process. We demonstrate the effective-
ness of uSplit in handling top-k join operations. uSplit
considers the ranked semantics of the analysis operations,
as well as the data and utility characteristics discovered at
runtime, in deciding on a data partitioning and resource al-
location strategy. We also demonstrate the effectiveness of
RanKloud in supporting other ranked operators such as the
skyline and k-nearest neighbor search operators. Our future
research will involve extending RanKloud to support join-
based skyline and k-nearest neighbor queries, in which the
data points on which the skyline/k-nearest neighbor query
is executed are not available, but an explicit join operation
needs to be carried out in order to discover these points.
 Amazon. Elastic MapReduce.
 A. Andoni and P. Indyk. Near-optimal hashing
algorithms for approximate nearest neighbor in high
dimensions. Commun. ACM, 51(1):117–122, 2008.
 R. Chaiken, B. Jenkins, P. A. Larson, B. Ramsey,
D. Shakib, S. Weaver, and J. Zhou. SCOPE: Easy and
efficient parallel processing of massive data sets. Proc.
VLDB Endow., 1:1265–1276, 2008.
 A. Corral, Y. Manolopoulos, Y. Theodoridis, and
M. Vassilakopoulos. Closest pair queries in spatial
databases. In SIGMOD, pages 189–200, 2000.
 J. Dean and S. Ghemawat. MapReduce: Simplified
data processing on large clusters. In OSDI, 2004.
 T. Elsayed, J. Lin, and D. W. Oard. Pairwise
document similarity in large collections with
MapReduce. In HLT, pages 265–268, 2008.
 R. Fagin, A. Lotem, and M. Naor. Optimal
aggregation algorithms for middleware. J. Comput.
Syst. Sci., 66(4):614–656, 2003.
 J. J. Lin. Brute force and indexed approaches to
pairwise document similarity comparisons with
MapReduce. In SIGIR, pages 155–162, 2009.
 C. Olston, B. Reed, U. Srivastava, R. Kumar, and
A. Tomkins. Pig latin: a not-so-foreign language for
data processing. In SIGMOD, pages 1099–1110, 2008.
 A. Pavlo, E. Paulson, A. Rasin, D. J. Abadi, D. J.
DeWitt, S. Madden, and M. Stonebraker. A
comparison of approaches to large-scale data analysis.
In SIGMOD, pages 165–178, 2009.
 S. Singh, J. Kubica, S. Larsen, and D. Sorokina.
Parallel large scale feature selection for logistic
regression. In SDM, pages 1171–1182, 2009.
 A. Vlachou, C. Doulkeridis, and Y. Kotidis.
Angle-based space partitioning for efficient parallel
Skyline computation. In SIGMOD Conference, pages
 Yahoo! Hadoop. http://hadoop.apache.org, 2008.
 B. Yao, F. Li, and P. Kumar. K nearest neighbor
queries and kNN-joins in large relational databases
(almost) for free. In ICDE, pages 4–15, 2010.
 R. Yu, M. Nagendra, P. Nagarkar, K. S. Candan, and
J. W. Kim. Workload-balanced processing of top-K
join queries on cluster architectures. Technical Report: