Optimizing Queries over Partitioned Tables in MPP Systems

Optimizing Queries over Partitioned Tables
in MPP Systems
Lyublena Antova
lantova@gopivotal.com
Amr El-Helw
aelhelw@gopivotal.com
Mohamed A. Soliman
msoliman@gopivotal.com
Zhongxian Gu
zgu@gopivotal.com
Michalis Petropoulos
mpetropoulos@gopivotal.com
Florian Waas
flw@datometry.com
ABSTRACT
Partitioning of tables based on value ranges provides a pow-
erful mechanism to organize tables in database systems. In
the context of data warehousing and large-scale data analy-
sis partitioned tables are of particular interest as the nature
of queries favors scanning large swaths of data. In this sce-
nario, eliminating partitions from a query plan that contain
data not relevant to answering a given query can represent
substantial performance improvements. Dealing with parti-
tioned tables in query optimization has attracted significant
attention recently, yet, a number of challenges unique to
Massively Parallel Processing (MPP) databases and their
distributed nature remain unresolved.
In this paper, we present optimization techniques for
queries over partitioned tables as implemented in Pivotal
Greenplum Database. We present a concise and unified
representation for partitioned tables and devise optimiza-
tion techniques to generate query plans that can defer deci-
sions on accessing certain partitions to query run-time. We
demonstrate, the resulting query plans distinctly outperform
conventional query plans in a variety of scenarios.
Categories and Subject Descriptors
H.2.4 [Database Management]: Systems—Query pro-
cessing; Distributed databases
Keywords
query optimization; MPP systems; partitioning
1. INTRODUCTION
It is hardly necessary to motivate the need for systems
for processing large amounts of data. Big data and data-
driven analysis is ubiquitous in all business verticals today,
including government agencies, financial corporations, tel-
cos, insurance and retail. From running simple reports to
executing complex analytics workloads to find insights in
data, these organizations are heavily investing in big data
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 cita-
tion on the first page. Copyrights for components of this work owned by others than
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re-
publish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
SIGMOD 2014 Snowbird Utah USA
Copyright 2014 ACM 978-1-4503-2376-5/14/06 ...$15.00.
http://dx.doi.org/10.1145/2588555.2595640 .
!"#$"%&#' ()!*+,' #(,$'
-(+./01/'!"#$"%&
!"#$"%&#'()!*+,' #(,$'
2$3./014'
5'
!"#$"%&#' ()!*+,' #(,$'
Figure 1: Table orders partitioned by date
solutions, and have motivated the birth of multiple startup
companies, as well as change in direction of many traditional
database vendors to enable scalable and efficient processing
of large amounts of data.
Data partitioning is a well-known technique for achiev-
ing efficiency and scalability when processing large amounts
of data. Partitioning comes in various shapes and forms.
Many modern database systems partition the data in or-
der to store and proccess it on different nodes to achieve
parallelism [9, 11, 16]. Data can also be partitioned on a
single machine, either horizontally by rows, or vertically by
columns, to reduce the scans to acquire the data needed to
asnwer a query. Some of the well-known benefits of data
partitioning include: reduced scan time by scanning only
the relevant parts, and improved maintenance capabilities:
data can be loaded, indexed, or reformatted independently.
In this paper, we present techniques for optimizing com-
plex queries to take advantage of partitioning and select only
relevant partitions for scanning.
A common scenario that most of our customers use is to
partition the data chronologically based on a date or times-
tamp field. Consider the example in Figure 1.
It shows an example table orders containing data for the
past 2 years, which is partitioned into monthly partitions.
The idea behind this partitioning scheme is that if a query
specifies a range predicate on the partitioning key, it can
avoid scanning redundant partitions that will not satisfy the
predicate. One such query, which summarizes the orders
from the last quarter is shown in Figure 2.
To evaluate this query, we need to scan only the last three
partitions instead of all 24 that constitute the orders table.
This technique is often referred to as static partition elimi-
nation and is implemented by nearly all systems supporting
partitioning [9, 11, 16]. It consists of statically determining
which partitions need to be scanned at query optimization
time based on the predicate specified in the query.

SELECT avg(amount)
FROM orders
WHERE date BETWEEN ’10-01-2013’ AND ’12-31-
2013’
Figure 2: Example query where non-relevant parti-
tions can be eliminated statically
!"#$"%&
!"#$"%&#'()!*+,' #(,$%&#'
-'
#(,$%&#'&+''
./0/1112''
#(,$%&#'&+''
.3111103/1112''
'()$*#+,&
#(,$%&#'4$("')!+,5' #(4' #(4%!6%7$$8'
Figure 3: Fact table orders partitioned by date id,
foreign key in the dimension table date dim
Often our customers model their data using a star schema
design like the one presented in Figure 3. In this figure the
fact tables orders has a foreign key to a separate dimension
table date dim. This model is a normalized form of the one
in Figure 1 and allows for specifying additional properties of
dates like “day of week”. The fact table and dimension table
can be partitioned independently. In our case the fact table
is partitioned by the foreign key date id. Queries on this
schema often need to join the fact table with the dimension
table, and static partition elimination is not always possible.
SELECT avg(amount)
FROM orders
WHERE date id IN
(SELECT date id
FROM date dim
WHERE year = 2013 AND
month BETWEEN 10 AND 12)
Figure 4: Example query where non-relevant par-
titions can only be eliminated at query time after
evaluating part of the query
Figure 4 presents a rewriting of the original query from
Figure 2, where now the orders table has to be joined with
the dimension table to find the matching tuples. In this
query the values for the partitioning keys of orders are not
known beforehand, but are only determined at runtime af-
ter evaluating the subquery on the dimension table. This is
known as dynamic partition elimination and most database
systems either do not support it, or provide a very rudimen-
tary support that works for simple queries and schema de-
signs. See Section 5 for details. In addition to join-induced
partition elimination, there are other cases where the parti-
tioning key may not be known at optimization time, namely
in the case of prepared statements with parameters. This
case as well may require support for dynamic partition elim-
ination, as parameter values are only provided at runtime.
In this paper we propose a novel approach for optimizing
complex queries on partitioned tables. In particular, our
contributions are:
• A model for representing partitioning and queries on
partitioned tables with the following properties:
– Abstract operators following the producer-
consumer paradigm: a PartitionSelector deter-
mining which partitions need to be scanned (pro-
ducer), and a DynamicScan, which consumes the
partition ids produced by the PartitionSelector
and performs the actual scan of those partitions.
– Compactness of generated query plans: plan size
independent of the total number of partitions in
the tables, or the number of partitions that need
to be scanned.
– Support for both static and dynamic partition
elimination in a unified way.
– The approach is independent of the actual storage
format of partitioned tables and can be applied to
a variety of models for representing partitioning
in the storage layer.
• Algorithms for partition elimination, which generate
all interesting placements of PartitionSelector within
a query plan.
• Support advanced partitioning schemes such as multi-
level hierarchical partitioning.
• Performance experiments showing that the efficiency
of plans generated by our optimizer over existing ap-
proaches. We are able to identify opportunities for
partition elimination in complex queries, and achieve
multi-fold improvement of the query runtime. The re-
sults from our in-lab experiments were also confirmed
in early field trials of Pivotal’s Greenplum Database,
Pivotal’s MPP datatabase, where some customers re-
ported multi-fold improvement of query runtime due
to advanced partition elimination techniques.
The rest of the paper is organized as follows. Sec-
tion 2 presents our model for querying partitioned tables
and describes algorithms for partition elimination. Sec-
tion 3 presents the implementation of partition selection al-
gorithms in Pivotal’s MPP systems, and Section 4 demon-
strates the efficiency of the approach. Finally, we review
related work in Section 5, and conclude in Section 6.
2. OPTIMIZING QUERIES ON PARTI-
TIONED TABLES
2.1 Definitions
A table T over schema (A1, ..., An) is a set of tuples {<
a1, ..., an >}. We say that a table T is logically partitioned
on a key pk ∈ {A1, ..., An} into partitions T1, ..., Tm if there
exists a partitioning function
fT : pk → {T1, ..., Tn, ⊥}
which is used to assign tuple t to a partition Ti or the invalid
partition (denoted by ⊥) based on the value of the partition-
ing key. The latter means that the tuple cannot be mapped
to any partition. For a partitioned table T we will denote
the set of its partitions as P(T ).
Note that the partitions T1, ..., Tm need not be material-
ized on disk. For simplicity we assume that given a logical
partition object id (OID) Ti the storage layer can locate and
retrieve the tuples belonging to that partition.

!"#$%&'('$#)*+,-.,
(/01/#'/,
2$3445#(/6/'753)*+,-+,!.,
!"#$
8-,
89,
:,
8-;;,
81<6/,=5>,
?@!,=5>,
(a) Full scan
!"#$%&'('$#)*+,-.,
(/01/#'/,
2$3445#(/6/'753)*+,-+,89:;.,
(/6/'7,)89:;.,
!"#$
<-,
(b) Equality partition selection
!"#$%&'('$#)*+,-.,
(/01/#'/,
2$3445#(/6/'753)*+,-+,89:;<.,
(/6/'7,)89:;<.,
!"#$
=-,
=>,
=;,
(c) Range partition selection
!"#$%&'('$#)*+,-.,
/0&#)12345267.,
8$9::0#(;<;'=09)*+,-+,12345267.,
!"#$
5->,
5?-,
5@-,
5$A<;('$#)1.,
(d) Join partition selection
Figure 5: Implementing different scan patterns using various PartitionSelectors
In most systems the partitioning function fT implements
either a categorical, or range partitioning scheme. We also
assume the existence of a function f∗
T :
f∗
T : φ(pk) → {..., Ti1 , ...} ⊆ P(T )
which for a given predicate φ(pk) on the partitioning key,
returns a set of partition OIDs {Ti1 , .., Tim } such that if a
tuple t /∈ {Ti1 , .., Tim } then t does not satisfy φ. In other
words, function f∗
T performs partition selection for a given
predicate on the partitioning key, returning all partitions
that may satisfy the given predicate φ. Note that such func-
tion always exists, as it can trivially return P(T ). Also, for
predicates of the form pk = c, the function f∗
T is the same as
the partitioning function fT applied to the value c. Ideally,
the partition selection function should return the minimal
set of partition OIDs for which the given predicate is satis-
fied, as it will be the basis for partition pruning described
next. In Section 3 we discuss how f∗
T can be implemented
in a commercial DBMS to handle complex predicates on the
partitioning key.
2.2 Query Model for Partitioned Tables
To implement scans over partitioned tables, we extend the
set of physical query operators by two new constructs: Par-
titionSelector and DynamicScan. Those two operators come
in pairs and implement a producer-consumer model, where
the PartitionSelector computes OIDs of partitions for the
DynamicScan, which in turn is responsible for retrieving tu-
ples belonging to those partitions. The PartitionSelector can
communicate the partition OIDs to the DynamicScan via
shared memory or any other communication channel known
in the literature.
As a first example, consider a simple full table scan query
over a partitioned table. Let T be a partitioned table with
partitions T1, ..., T100, where partition Ti holds tuples for
which pk ∈ [(i − 1) ∗ 10 + 1, i ∗ 10). Figure 5(a) shows the
query plan implementing a full scan over T. Note the Parti-
tionSelector and the DynamicScan on the left and right side
of the plan, respectively. The PartitionSelector produces all
child partition OIDs T1, ..., T100, shown in the table in the
middle, and sends them for consumption to the Dynamic-
Scan, denoted by the dotted lines in the figure. At the root
of the plan tree is a Sequence operator ensuring the Parti-
tionSelector is executed before the DynamicScan.
By varying the shape of the PartitionSelectors and their
placement in the query plan, we can implement more com-
plex patterns, such as partition selection based on equality
and range predicates, as well as dynamic partition elimina-
tion. Figures 5(b) and 5(c) show the query plans for equality
and range selection over T, respectively. Here, the Partition-
Selector is annotated with the partition-selecting predicate
from the query, and the selected OIDs in the table in the
middle of these figures only include the OIDs of partitions
that may satisfy the selection predicate. Finally, Figure 5(d)
shows a query plan that accomplishes dynamic partition
elimination for a join query. Note that the PartitionSelector
is on the opposite side of the DynamicScan and its pred-
icate R.A = T.pk refers to values from table R. When the
outer (left) side of the join is executed, tuples from R will
be streamed into the PartitionSelector, where the partition
selection function will be applied to choose those partitions
determined by the value of R.A. In this case, no Sequence
operator is necessary, as the Join operator enforces the left-
to-right order of execution of children.
Below we give informal definition of the three new opera-
tors PartitionSelector, DynamicScan, and Sequence.
PartitionSelector. An operator that is annotated with a
partition table OID T, partScanId identifier, and an optional

predicate on the partitioning key of T. The PartitionSelector
is an operator with side effects: based on the OID T and
the given predicate, it computes all child partition OIDs
which satisfy the predicate using the function f∗
T defined
in Section 2.1. It then pushes them to the DynamicScan
with the same partScanId. The partScanId identifier is used
to identify the different (PartitionSelector, DynamicScan)
pairs in case the query refers to multiple partitioned tables,
or to multiple instances of the same partitioned table. If
no predicate is specified, the PartitionSelector retrieves all
child OIDs of T. A PartitionSelector may have at most one
child. If it has a child, the output of the operator is the
same as that of its child, otherwise no output is produced.
DynamicScan. An operator that is responsible for scan-
ning tuples from a partitioned table T. It is annotated with
the partitioned table OID and a partScanId. The Dynamic-
Scan operator consumes the partition OIDs provided by the
PartitionSelector with the same partScanId and retrieves the
tuples from those partitions.
Sequence. An operator that executes its children in order
and returns the result of the last child.
Our query model can be generalized by abstracting Dy-
namicScan as a table-returning function that produces a set
of rows from partitioned table based on a (run-time) pa-
rameter. Such abstraction can be used to describe several
proposals in the literature [9, 11] where partition selection
and scanning are done by different operators. For example,
an Index-Join implementation of this model performs parti-
tion selection by the outer child of the join which computes
the keys of partitions to be scanned (the parameter values),
while the inner child of the join performs partition scanning
by looking up an index defined on partition key.
A key point that needs to be addressed when implement-
ing this model in MPP databases is that partition selection
and scanning operators could be running within process-
es/threads in different machines. We describe in Section 3
our approach to address such challenge.
2.3 Placement of PartitionSelectors
In this section we describe the algorithm which computes
the placement of PartitionSelectors given a physical operator
tree with DynamicScans. Often, there are multiple ways to
place PartitionSelectors in the expression tree, but not all
placements achieve optimal partition elimination as we will
see next. We will use as a running example the query in
Figure 6 which selects all sales records placed in the last
quarter in California. The tables date dim and sales fact
are partitioned on month and date id, respectively.
SELECT *
FROM sales fact s, date dim d customer dim c
WHERE d.month BETWEEN 10 AND 12 AND
c.state=’CA’ AND d.id=s.date id AND c.id=s.cust id;
Figure 6: Query selecting all sales records from the
last quarter
The input to our algorithms is an expression tree, pro-
duced by the query optimizer, which has DynamicScan op-
erators for scanning the partitioned tables, but no Partition-
Selectors have been placed yet. A possible expression tree for
the example query of Figure 6 is given in Figure 8(a). The
goal of the algorithm is to find optimal placement of Parti-
tionSelectors, where optimality is determined with respect
to the minimum number of parts that need to be scanned.
For this example, the result of the algorithm is shown in
Figure 8(b). Note the two PartitionSelectors in the result-
ing plan. The lower one with ID 1 implements partition
elimination for the DynamicScanof date dim, while Parti-
tionSelector 2 implements dynamic partition elimination for
sales fact using values from the selection on date dim. An-
other possible PartitionSelector placement is to push Par-
titionSelector 2 on the inner side of the join. However, no
partition elimination will be done for the latter query plan.
Note also that the DynamicScan and corresponding Parti-
tionSelector do not have to be immediate children of the
same node: in Figure 8(b) they are separated by multiple
levels in the plan tree. This shows the versatility of approach
in optimizing complex queries on partitioned tables.
For simplicity, we present the recursive implementation of
the algorithm. We show in Section 3.1 how the algorithms
can be implemented to work on a compact representation of
the plan space as opposed to a complete operator tree. Note
also that the algorithms for PartitionSelector placement are
orthogonal to data distribution, where base tables are hash
distributed and placed onto different physical nodes of the
cluster. Section 3 provides details on how partitioning and
distribution can be combined into one system.
Algorithm 1: PlacePartSelectors
Input : List inputPartSelectors, Expression expr
Output: Expression where all partition selection has been
enforced
1 List partSelectorsOnTop;
2 List childPartSelectors;
3 expr.operator.ComputePartSelectors(inputPartSelectors,
partSelectorsOnTop, childPartSelectors);
4 List newChildren;
5 foreach child in expr.children, childPartSelectorList in
childPartSelectors do
6 Expression newChild = PlacePartSelectors(child,
childPartSelectorList);
7 newChildren.Add(newChild);
8 end
9 return EnforcePartSelectors(partSelectorsOnTop,
expr.Operator, newChildren);
PartSelectorSpec
partScanID: int
partKey: ColRef
partPredicate: Expression
Figure 7: PartSelectorSpec
The main function is shown in Algorithm 1. It accepts an
input expression expr and a list of input PartSelectorSpecs,
and returns an expression where all PartitionSelectors have
been placed. Each PartSelectorSpec is a compact specifi-
cation of the PartitionSelector operator which needs to be
placed for each unresolved DynamicScan. Its structure is
shown in Figure 7 and contains the partScanId identify-
ing the DynamicScan, the partitioning key, and optionally a
predicate on the partitioning key that can be used for parti-
tion elimination. Initially, the partPredicate is NULL, but as

we push PartitionSelectors through the different operators,
it may get augmented, as shown in the following subsections.
The input PartSelectorSpecs for the PlacePartSelectors
function are initialized by traversing the tree and identi-
fying all DynamicScans that need corresponding Partition-
Selectors. This input list is first passed to the function
ComputePartSelectors, the main driver of the computa-
tion (line 3). This function is overloaded for each opera-
tor type, and computes which PartitionSelectors need to be
placed on top of the current operator (in the output list
partSelectorsOnTop), and which need to be pushed down
to the children (output list childPartSelectors). After the
two lists have been computed in Line 4 of Algorithm 1, the
algorithm recursively pushes the necessary PartitionSelec-
tors to the child nodes (Lines 6 and 7), and places the rest
of the PartitionSelectors on top (Line 9).
The following subsection show the implementation
of ComputePartSelectors for several of the partition-
eliminating operators, including Join (Section 2.3.3), Se-
lect (Section 2.3.2), and a default implementation for non-
partition eliminating operators (Section 2.3.1). Listed below
are the helper functions used in the algorithms presented
next:
• EnforcePartSelectors: Given an operator and a list
of PartitionSelectors which need to be enforced on top,
constructs a new expression tree where the Partition-
Selectors have been placed on top
• Operator::HasPartScanId: Checks whether the
DynamicScan with the given part scan id is in scope
of the expression tree rooted at the current operator
• FindPredOnKey: Given a scalar expression, ex-
tracts a predicate referring to the given column
• Conj: Construct a conjunction of the given predicates
2.3.1 Default PartitionSelector Placement
The default implementation of ComputePartSelectors is
presented in Algorithm 2 and is used for operators do not
have a partition-filtering predicate, such as GroupBy, Union,
Project, etc. It simply pushes the PartSelectorSpecs to
the subexpression which defines the DynamicScan with the
given partScanId (Line 8), or places it on top, if the Dynam-
icScan is not defined in the subtree rooted by the operator
(Line 3).
2.3.2 PartitionSelector Placement for Select
Algorithm 3 shows the implementation of
ComputePartSelectors for the Select operator. For
each input PartSelectorSpec, we check whether the cor-
responding part scan is defined in the subtree rooted by
the Select operator (Line 2). If not, it is enforced on top
of the Select operator. If it is defined below, we push the
PartSelectorSpec to the child node, but before that we
extract any partition-filtering predicates on the partitioning
key, and add them to the PartSelectorSpec (Line 6-13)
by first constructing a conjunction with any predicates
passed from the top. An example run of the function
for the example of Figure 8(a) is shown in Figure 8(c).
DynamicScan with id 1 is defined in the subtree of the
Select, so it is pushed further down to the child node. Also,
since the Select’s predicate involves the partitioning key
Algorithm 2: Operator::ComputePartSelectors
Input : List inputPartSelectors, List
partSelectorsOnTop, List childPartSelectors
Output: Compute default partition selectors for operators
1 foreach partSpec in inputPartSelectors do
2 if !this.HasPartScanId(partSpec.partScanId) then
3 partSelectorsOnTop.Add(partSpec);
4 end
5 else
6 foreach child operator op do
7 i = order of op among children;
8 if op.HasPartScanId(partSpec.partScanId)
then
9 childPartSelectors[i].Add(partSpec);
10 end
11 end
12 end
13 end
Algorithm 3: Select::ComputePartSelectors
Output: Compute partition selectors for Select operator
4 end
5 else
6 Expression partKeyPredicate =
FindPredOnKey(partSpec.partKey,
this.Predicate());
7 if partKeyPredicate found then
8 PartSelectorSpec newPartSpec = new
PartSelectorSpec(partSpec.partScanId,
partSpec.partKey, Conj(partKeyPredicate,
partSpec.partPredicate);
9 childPartSelectors[0].Add(newPartSpec);
10 end
11 else
12 childPartSelectors[0].Add(partSpec);
13 end
14 end
15 end
month, the predicate is added to the PartSelectorSpec for
the child node. The PartitionSelector for DynamicScan 2
needs to be resolved on top of the Select operator since it is
not defined in the subtree.
2.3.3 PartitionSelector Placement for Join
Algorithm 4 shows the implementation of
ComputePartSelectors for the Join operator. Just
like the default implementation of this method, and the one
for Select, we start by checking whether a given part scan
is in the scope of the subtree rooted at the Join operator
(Line 2). If not, it needs to be enforced on top of the Join
operator. Otherwise we probe to see if it is defined in the
outer side. Recall that the DynamicScan is the consumer,
and the PartitionSelector is the producer in our model. So

Select
(month
≥
10
and
month
≤
12)

DynamicScan(1,
date_dim)

HashJoin(id=date_id)

DynamicScan(2,
sales_fact)

Hash

Select
(state=‘CA’)

Scan(customer_dim)

HashJoin(id=cust_id)

Hash

(a) Hash join query tree before PartitionSelector placement
Select
(month
≥
10
and
month
≤
12)


DynamicScan(2,
sales_fact)

Hash

Select
(state=‘CA’)

Scan(customer_dim)

HashJoin(id=cust_id)

Hash

DynamicScan(1,
date_dim)

Sequence

ParFFonSelector(1,

month
≥
10
and
month
≤
12)

ParFFonSelector(2,
date_id=id)

Tuple
flow

OID
flow

(b) Hash join query tree after PartitionSelector placement
!"#"$%&'()*%+&,&-.&/*0&()*%+&1&-23&
45*/(6$!$/*'-7&0/%"806(3&
6*9:%;/<%!"#"$%)<!9"$=&&
>?-7&()*%+7&!@7&?27&0/%"8607&0/%"860A60@B&
&
9/<%!"#"$%)<CD*E)9=&&
>?27&0/%"8607&0/%"860A60@B&
&
$+6#0;/<%!"#"$%)<C=&
&>>?-7&()*%+7&()*%+&,&-.&/*0&()*%+&1&-2@BB&
6*9:%;/<%!"#"$%)<!9"$=&&
>?-7&()*%+7&()*%+&,&-.&/*0&()*%+&1&-2@B&
&
9/<%!"#"$%)<CD*E)9=&&
>?-7&()*%+7&()*%+&,&-.&/*0&()*%+&1&-2@B&
&
$+6#0;/<%!"#"$%)<C=&>B&
(c) Execution of ComputePartSelectors for Select
Select
(month
≥
10
and
month
≤
12)

DynamicScan(1,
date_dim)


…

inputPartSelectorSpec:

{<1,
month,
>,
<2,
date_id,
>}

partSelectorsOnTop:
{}

childPartSelectors:

{{<1,
month,
>,
<2,
date_id,
date_id=id>},
{}}

Hash

(d) Execution of ComputePartSelectors for HashJoin
Figure 8: Placing PartitionSelectors in an expression
if the DynamicScan is defined in the outer side of a Join
operator, we cannot place the PartitionSelector on the inner
side, as that destroys the order between the producer and
the consumer. This check is implemented in Lines 7-8. If
the DynamicScan is defined in the inner side, we check to
see if the join predicate contains a condition on the parti-
tioning key, which can be used for partition elimination. If
the answer is yes, we push the PartSelectorSpec onto the
outer side (Line 16), otherwise it needs to be resolved on
the inner side, close to where the DynamicScan is defined.
An example run of the function for the example of Fig-
ure 8(a) is shown in Figure 8(d). DynamicScan with id 1
is defined in the outer side of the HashJoin, so the PartSe-
lectorSpec is pushed further down to the outer child node.
DynamicScan with id 2 is defined in the inner child, and the
HashJoin’s predicate restricts the partitioning key date id.
This means that we can use it to do partition pruning by
pushing the PartSelectorSpec onto the outer side. The result
of the function ComputePartSelectors for the join node is
shown in the bottom part of the box in Figure 8(d). No Par-
titionSelectors are placed in the “on top” list, and none are
pushed to the inner child of the hashJoin (noted by the {}
in the childPartSelector list). Thus both PartitionSelectors
for scan ids 1 and 2 will be placed on the join’s outer side.
2.4 Multi-level Partitioned Tables
Most database systems also support hierarchical partition-
ing, in which a table is partitioned over multiple levels. Fig-
ure 9 depicts an example of such partitioning scheme. The
orders table uses a 2-level partitioning. The first level uses
the date column such that each partition contains the data
for one month. These partitions are further split into sub-
partitions using the region column. A query on this table
can specify particular date ranges, regions, or both. The
system can use these conditions to avoid scanning partitions
that will yield no results.
In order to support multi-level partitioning, the structures
and algorithms explained earlier need to be extended. The
PartitionSelector defined in Section 2.2 has to be extended so
that it is annotated with a list of optional predicates instead

Algorithm 4: Join::ComputePartSelectors
Output: Compute partition selectors for join
4 end
5 else
6 Expression partKeyPredicate =
FindPredOnKey(partSpec.partKey,
this.Predicate());
7 bool definedInOuterChild =
children[0].HasPartScanId(partSpec.partScanId);
8 if definedInOuterChild then
10 end
11 else if partKeyPredicate not found then
13 end
14 else
15 PartSelectorSpec newPartSpec = new
PartSelectorSpec(partSpec.partScanId,
partSpec.partKey, Conj(partKeyPredicate,
partSpec.partPredicate);
16 childPartSelectors[0].Add(newPartSpec);
17 end
18 end
19 end
!"#$%&'%(
!"#$"%&
)*+,*-.+( "/)0#1( +"1,( *,2.)#(
3,4$%&'%(
5,6$%&'7(
8(
8(
9,2.)#('(
9,2.)#(%(
8(
9,2.)#('(
9,2.)#(%(
:,;,<('=(5"1,( :,;,<(%=(9,2.)#(
>'(
>%(
>%?(
>'@'(
>'@%(
>%@'(
>%@%(
8(
9,2.)#('(
9,2.)#(%(
>%?@'(
>%?@%(
Figure 9: Multilevel partitioning by date and region
of a single optional predicate. Based on the OID T and the
given predicates, it computes all leaf-level child partition
OIDs that satisfy the predicates. For example, using the
orders table in Figure 9, we show in Figure 10 a number of
possible predicates given to the PartitionSelector, and the
child partition OIDs it computes in each case.
The PartSelectorSpec from Figure 7 has to support a list
of partKeys and a list of predicates as shown in Figure 11.
The number of items in both lists have to be equal, and be
the same as the number of partitioning levels for the table
partPredicate Partition OIDs
date=’Jan-2012’ T1,1, T1,2, ..., T1,n
region=’Region 1’ T1,1, T2,1, ..., T24,1
date=’Jan-2012’ AND
region=’Region 1’
T1,1
φ all leaf part OIDs
Figure 10: Multi-level partition selection
PartSelectorSpec
partScanID: int
partKeys: List<ColRef>
partPredicates: List<Expression>
Figure 11: Extended PartSelectorSpec
on which this PartSelectorSpec is defined. Note that some
elements of the partPredicates list may be empty, indicating
the absence of a predicate on the corresponding partKey.
The only change to the algorithms is that the FindPre-
dOnKey function used in Algorithms 3 and 4 must be ex-
tended to take a list of partKeys, and return a list of pred-
icates corresponding to these keys. This function returns
NULL if no partition-filtering predicates are found.
3. IMPLEMENTATION
Greenplum Database (GPDB) is Pivotal’s massively par-
allel relational database system. HAWQ [12] is Piv-
otal’s SQL engine for Hadoop. Both Pivotal products use
Orca [15], a new query optimizer designed specifically to
support large scale analytic processing in different comput-
ing architectures. We describe the implementation of parti-
tion selection inside Orca in Section 3.1. We then describe
the runtime environment in Section 3.2.
3.1 Query Optimization
In an MPP system, data can be distributed to different
hosts or physical machines. During query execution, the
distribution of intermediate results can be enforced in mul-
tiple ways including hash distribution, where tuples are dis-
tributed to hosts based on some hash function, replicated
distribution, where a full copy of a table is stored at each
host and singleton distribution, where the whole distributed
table is gathered from multiple hosts to a single host. Data
distribution is orthogonal to partitioning. That is, a dis-
tributed table can also be partitioned on each host.
A query plan can enforce a particular data distribution
through special Motion operators. During query execution,
a Motion operator acts as the boundary between two ac-
tive processes sending/receiving data and potentially run-
ning in different hosts. This constrains the shape of valid
plans that can perform partition selection, since we rely
on shared memory to accomplish the required communica-
tion between PartitionSelector and DynamicScan operators
(cf. Section 2.2). Specifically, the optimizer must guarantee
that a pair of communicating PartitionSelector and Dynam-
icScan operators run within the same process. This means
that no Motion operator can exist between PartitionSelector,
DynamicScan and their lowest common ancestor.
Figure 12 shows examples of valid and invalid plans based
on the previous constraint. In general, it is not straight-
forward to constrain the shape of arbitrarily complex plans,

!"#$%&'()*+,(-.%
/0$-1#2,2-$&'.%
3-455"$,67628"4%
&9:%'()*:%'()*+,(-.%
,2-$&,.%
;"5"$%
!%
!"#$%&&'('
!"#$%&&')'
/0$-1#2,2-$&'.%
3-455"$,67628"4%
&9:%'()*:%'()*+,(-.%
,2-$&,.%
;"5"$%
!"#$%&&'('
!"#$%&&')'
"%
!"#$%&'()*+,(-.%
!"#$%&'()%$"(
*$%&'()%$"(
Figure 12: Interaction of Motion and PartSelector
where PartitionSelector and DynamicScan appear in distant
sub-trees. We next show how the query optimizer handles
this requirement in a principled way.
Orca is the new Pivotal query optimizer based on the Cas-
cades optimization framework [6]. Orca compactly encodes
optimizer’s search space in the Memo structure, which is
composed of a set of containers called groups, where each
group contains logically equivalent expressions called group
expressions. Each group expression is an operator that has
other groups as its children. This recursive Memo structure
allows compact encoding of a very large space of plan alter-
natives. Figure 13 shows an example Memo. The dependen-
cies between operators are encoded using group references.
For example, HashJoin[1,2] in Group 0 is a group expression
that has two child groups 1 and 2.
We describe optimization of queries on partitioned tables
in Orca using the following simple example:
SELECT * FROM R, S WHERE R.pk=S.a
where R and S are hash distributed tables, R is a partitioned
table with part key R.pk and S is a non-partitioned table.
Orca has an extensible property enforcement framework
that models requirements such as data distribution and par-
tition selection as physical properties. A given plan may
either satisfy a physical property on its own (e.g., a hash-
distributed table delivers hash-distributed data), or an en-
forcer operator (e.g., Motion) needs to be plugged in the
plan to deliver the required property. The PartitionSelector
operator is the enforcer of partition selection property.
Figure 13 shows a partial Memo structure for the previous
query. We distinguish enforcer operators using black boxes.
Enforcer’s child belongs to the same group containing the
enforcer itself. For example, ‘Redistribute(R.pk)[1]’ is an en-
forcer that delivers Hashed(R.pk) distribution using a child
operator in Group 1. Similarly, ‘Replicate[2]’ is an enforcer
that delivers Replicated data distribution using a child op-
erator in Group 2. Adding multiple enforcers to the same
group allows considering different permutations of enforced
properties, while discarding the invalid ones, as we show
next.
2:#Hash#Join#[2,1]#
2:#Replicate[1]#
## Opt.#Request# Best#GExpr#
1# Any,##<0,#R.pk,#Φ># 2"
GROUP#0#
2# Replicated,#<0,#R.pk,#Φ># 2"
3# Hashed(R.pk),#<0,#R.pk,#Φ># 3"
4# Any,#<0,#R.pk,#Φ># 4"
5# Any,<># 1"
1:#DynamicScan(R)[]#
GROUP#1#
2:#Replicate[2]#1:#Scan(S)[]#
GROUP#2#
4:#Par>>onSelector(0,#R.pk,#R.pk=S.a)##[2]#
3:#Redistribute(R.pk)#[1]#
6# Replicated,#<># 2"
7# Any,#<># 1"
8# Replicated,#<0,#R.pk,#R.pk=S.a># 4"
9# Hashed(S.a),#<0,#R.pk,#R.pk=S.a># 3"
3:#Redistribute(S.a)#[2]#
4:#Par>>onSelector(0,#R.pk,#Φ)##[1]#
1# 8,#5#
8# 6#
6# 7#
1:#Hash#Join#[1,2]#
Groups#Hash#Tables# Memo#
Figure 13: Partial Memo
Optimization starts by specifying required properties in
an initial optimization request submitted to the Memo group
that corresponds to the root of the query expression. An
optimization request r submitted to Memo group g is com-
puted by finding the plan satisfying r, rooted by an operator
in g and has the least estimated cost. We call the root op-
erator of such a plan the best group expression (GExpr) of
request r.
For each incoming optimization request to g, each group
expression in g creates corresponding requests to child
groups based on operator’s local requirements. Figure 13
shows groups hash tables, where each incoming optimization
request is cached and is associated with its best GExpr. The
small tables below some group expressions show the relevant
mappings between incoming requests and child requests, as
we describe next.
In our example query we assume the initial optimization
request is req. #1: {Any, <0, R.pk, φ>} in Group 0, which
specifies that results can have Any data distribution but
must select partitions for table 0, which is R. When optimiz-
ing the HashJoin operators in Group 0 for this request, we
need to consider additional local requirements to co-locate
tuples to be joined based on the condition R.pk=S.a. This
can be accomplished in multiple ways including: (1) re-
questing one child to be Replicated and the other child to
have Any distribution, and (2) requesting S distribution to
be Hashed(S.a) and R distribution to be Hashed(R.pk). For
partition selection, HashJoin requests using the interesting
partition selection condition R.pk=S.a in its left child, be-
cause of the implicit execution order of join children (left
to right). These different optimization alternatives are enu-
merated and encoded using different optimization requests,
as shown in Figure 13.
Valid plans are encoded in the Memo by maintaining, for
each operator, a mapping of computed optimization requests
to the corresponding child optimization requests. For exam-
ple, in Figure 13 the small table below PartitionSelector in
Group 2 indicates that the best plan rooted by Partition-
Selector for req. #8 requires a child plan in Group 2 that
satisfies req. #6. This plan is given by Replicate operator,
which in turn requires a child plan satisfying req. #7 which
is given by Scan(S).

Hash%Join%
Scan(S)%
Par$$onSelector,
(0,,R.pk,,Φ),
DynamicScan(R)%
Replicate,
Hash%Join%
Scan(S)%
Par$$onSelector,
(0,,R.pk,,Φ),
DynamicScan(R)%
Redistribute(R.pk),
Hash%Join%
Scan(S)%
Par$$onSelector,
(0,,R.pk,,Φ),
DynamicScan(R)%
Replicate,
Hash%Join%
DynamicScan(R)%
Par$$onSelector,
(0,,R.pk,,R.pk=S.a),
Scan(S)%
Replicate,
Plan,1, Plan,2,
Plan,4,Plan,3,
Redistribute(S.a),
Figure 14: Partial plan space
To illustrate, consider HashJoin[2,1] operator in Group 0.
A pair of optimization requests are generated for the child
groups: req. #8: {Replicated, <0, R.pk, R.pk=S.a>} for
Group 2, and req. #5: {Any, <>} for Group 1. In order
to satisfy req. #8 using plans rooted by Replicate in Group
2, we cannot consider PartitionSelector as a child. The reason
is that such a plan would make the communication between
partition selection producer and consumer impossible, as we
discussed earlier in the beginning of this section. On the
other hand, when considering plans rooted by PartitionSe-
lector in Group 2, we can consider Replicate as a child. This
capability is implemented in Orca by allowing each operator
to prohibit enforcing some properties on top of it. When-
ever a plan alternative is considered, operator-specific logic
is executed to guarantee that enforcers are plugged in the
right order.
Figure 14 shows the partial plan space corresponding to
the partial Memo in Figure 13. Plans 1, 2, and 3 are
generated through optimization requests originating from
HashJoin[1,2], while Plan 4 is generated through optimiza-
tion requests originating from HashJoin[2,1]. The only plan
that performs partition selection is Plan 4. This is possi-
ble at the cost of replicating S. Orca’s cost model considers
these different alternatives in order to pick the best plan.
3.2 Runtime Environment
GPDB and HAWQ allow users to partition a table by
specifying a range or categorical constraints for each parti-
tion. On disk, partitions are represented as separate physi-
cal tables, with associated check constraint to represent the
corresponding constraint on the partitioning key. Note that
each constraint can be written in the form pk ∈ ∪i(ai1 , aik ),
where (ai1 , aik ) is an open or closed interval, possibly open-
ended. This representation also covers categorical partition-
ing, where the interval’s start and end overlap.
To implement PartitionSelector operators, we use a com-
bination of special-purpose built-in functions, and existing
query operators to invoke these functions. The built-in func-
tions retrieve metadata information about the partitioned
table during query execution, and the query plan may have
filters on top to select only those partitions which need to
be scanned. Table 1 lists the built-in functions which imple-
!"#$%&'('$#)*+,-.,
/0&#)12345267.,
890:;'<)123+,,
,,,,!"#$$%&'!#%!"("$%&)-+,
,,,,,,!"#$$%&')*+*,$%&)520&=+,123...,
5$>?;('$#)1.,
(a) Implementation of equality-based partition selec-
tion
!"#$%&'('$#)*+,-.,
(/01/#'/,
2345/'6)7$3884#97347$:$84#)-+,4&;..,
(/</'6,)7=>?@.,
!"#$$%&'(%&)*#"+&*))*A4&;.,
(/</'6)%&#>?@.,
(b) Implementation of range-based partition selec-
tion
Figure 15: Implementation of partition selectors in
runtime environments
ment partition selection functionality. The first three func-
tions perform retrieval of the partitioning metadata in dif-
ferent forms, and the last one, partition propagation, pushes
a selected partition OID from the PartitionSelector to the
DynamicScan with the given id.
Figures 15(a) and 15(b) show the implementation of the
PartitionSelectors for equality and range-based partition se-
lection from Figures 5(d) and 5(c), respectively. To imple-
ment equality-based partition selection (Figure 15(a)), we
call the function partition selection with the join parame-
ter for the partitioning key (R.A), and pass the resulting
OID to partition propagation, which pushes it to the Dy-
namicScan. To implement range-based partition selection
(Figure 15(b)), we invoke the function partition constraints
which returns all child partitions together with their range
constraints. The selection on top of the function call selects
only those partitions whose range starts before the constant
specified in the query, and the Project on top propagates the
OIDs of tuples which pass the filter to the DynamicScan.
Note that in this approach “static” and “dynamic” parti-
tion selection are implemented in a uniform way. For ex-
ample, to implement the “static” partition elimination from
Figure 5(b), we need to invoke partition selection with the
constant from the query, whereas in Figure 15(a) we pass
values from a table.
4. EXPERIMENTS
In this section, we present the experimental evaluation of
our approach. The goal of our experiments is to show that:
(1) our partitioning model does not introduce overhead in
table scans, (2) our partition elimination algorithm is ef-
fective, and (3) our model guarantees the compactness and
scalability of query plan sizes.

function return type description
partition expansion(rootOid) setof(OID) set of all child partition OIDs for the
given root OID
partition selection(rootOid, value) OID OID of child partition containing the
given value for the partitioning key
partition constraints(rootOid) (OID,min,minincl,max,maxincl) set of child partition OIDs with their
constraints
partition propagation(partScanId, oid) void pushes the given partition oid to the
DynamicScan with given id
Table 1: Built-in partition selection functions
#parts Description Overhead
42 each part represents 2 months 3%
84 partitioned monthly 3%
169 partitioned bi-weekly 1%
361 partitioned weekly 2%
Table 2: Partitioning lineitem
4.1 Setup
For our experiments, we use a cluster with four nodes
connected with 10Gbps Ethernet. Each node has dual In-
tel Xeon eight-core Sandy Bridge processors at 2.7GHz,
64GB of RAM and twelve 600GB SAS drives in two RAID-5
groups. The operating system is Red Hat Enterprise Linux
5.5.
For partition elimination experiments, we use 256GB
TPC-DS benchmarks [1] with partitioned tables. TPC-DS
is an industry standard decision-supporting benchmark that
consists of a set of complex business analytic queries. Its 25
tables, 429 columns and 99 query templates can well repre-
sent a modern decision-support system and is an excellent
benchmark for testing query optimizers.
4.2 Overhead of Partitioning
In this section, we show the scalability of our approach
with respect to the number of partitions. We accomplish
that by using the simple query:
SELECT * FROM lineitem;
using the lineitem table from the TPC-H benchmark [2],
with 7 years worth of data. We vary the number of parti-
tions, using some common partitioning scenarios, as illus-
trated in Table 2. The table also shows the overhead intro-
duced by each of these scenarios versus the case of using an
unpartitioned lineitem table. We can see from the results
that partitioning does not introduce significant overhead,
and that the performance of the full scan query is stable
regardless of the number of partitions. This makes parti-
tioning a very scalable approach to improve performance.
4.3 Partition Elimination Effectiveness
In this section we use a subset of the TPC-DS bench-
mark [1] to demonstrate the effectiveness of our partition
elimination approach. We compare Orca to the legacy query
optimizer of GPDB (a.k.a. Planner). The workload used in
this experiment contains queries from TPC-DS that refer-
ence partitioned tables: store sales, web sales, catalog sales,
store returns, web returns, catalog returns and inventory.
We ran the entire workload using both Planner and Orca.
Table 3 shows a high-level classification of the workload
based on partition elimination. In 80% of the queries, Orca’s
partition elimination algorithm is just is good as Planner.
In 14% of the workload, Orca succesfully eliminates more
partitions than Planner, resulting in reduced scanning cost.
In 6% of the workload (the last 2 categories), Orca pro-
duced sub-optimal plans that do not eliminate partitions or
eliminate fewer partitions than Planner. These sub-optimal
plans are partly due to cardinality estimation errors or sub-
optimal cost model parameters that need further tuning.
We are actively investigating these issues and constantly im-
proving Orca.
Figure 16 shows the number of scanned parts from each
table, aggregated across the whole workload. As can be seen
in the figure, the number of partitions scanned from each ta-
ble using Orca is less than the number of partitions scanned
using Planner. Orca’s partition elimination approach elim-
inates up to 80% of the partitions, as is the case with the
web returns table.
Note that in the above experiment we do not report abso-
lute running times, as Orca contains multiple improvements
over the legacy Planner, which makes it hard to quantify the
runtime benefits brought solely by partition selection across
the two systems. A comprehensive performance study of
Orca, including comparison with Planner and competitor
systems is outside of the scope of this paper, and is pre-
sented in [15].
For the purposes of this paper, we conducted another ex-
periment with the TPC-DS benchmark from above, in which
we ran the same workload in Orca for two different configura-
tions; in one case partition selection was enabled, and in the
other case it was disabled, keeping all other parameters the
same. Figure 17 shows the relative improvement for each
query obtained as a result of enabling partition selection.
The improvement is depicted as a percentage of the running
time without partition selection, so that an improvement
of 50% means the query ran in half the time, and the big-
ger the percentage, the higher the savings. One can see that
across the board partition selection speeds up execution time
in various degrees, both for short-running and long-running
queries. More than half of the queries improved above 50%,
and for over 25% of the queries the improvement was above
70%. Figure 17 also shows one large and two small out-
liers in the short-running, and the medium-to-long running
query blocks, respectively, for which performance degraded
when partition selection was turned on. We investigated
these cases and found out that the outliers are caused by
Orca picking a suboptimal plan for the partition-selection
case, which we contribute to imperfect tuning of cost model
parameters.

Category Percentage
Orca eliminates parts, Planner does not 11%
Orca eliminates more parts than Planner 3%
Orca and Planner eliminate parts equally 80%
Orca eliminates fewer parts than Planner 3%
Orca does not eliminate parts, Planner does 3%
Table 3: Workload classification
!"
#!"
$!"
%!"
&!"
'!!"
'#!"
'$!"
()*+,-(./,(" 0,1-(./,(" 2.)./*3-(./,(" ()*+,-+,)4+5(" 0,1-+,)4+5(" 2.)./*3-+,)4+5(" 657,5)*+8"
!"#$"%$&'())*+$,(-.&$
/(01*&$
9/.55,+" :+2."
Figure 16: Partition elimination
4.4 Plan Size Comparison
We performed experiments to show the scalability of our
approach with respect to plan size. We compared the re-
sults of Orca against plans produced by Planner, where par-
titioning is based on PostgreSQL inheritance mechanism,
and query plans explicitly list all partitions that need to be
scanned. Note that the latter is also the approach presented
in [7].
4.4.1 Queries with a Constant Partition-eliminating
Predicate
In this experiment we start with a partitioned lineitem ta-
ble from the TPC-H benchmark, and run a simple selection
query, which limits the number of partitions that need to be
scanned with a predicate of the form l shipdate < X. By
varying X we produce queries that select 1%, 25%,50%,75%,
and 100% of the partitions. Figure 18(a) shows that Orca’s
query plan size remains constant, while Planner’s plan grows
linearly with the number of partitions that need to be
scanned (and thus explicitly enumerated in the plan).
4.4.2 Queries with a Join Partition-eliminating
Predicate
In this experiment, we consider synthetically generated
tables R(a, b), S(a, b) partitioned on R.b and S.b, respec-
tively. We vary the number of partitions in each table, and
execute the following join query:
select * from R, S where R.b=S.b and S.a<100;
The planner supports dynamic partition elimination,
where the necessary partition OIDs are computed at run-
time and stored in a parameter, which is then passed to the
actual query plan for the join, which uses the parameter to
determine whether a given partition needs to be scanned or
not. The query plan however needs to list all partitions, as
it is not known at optimization time which partitions will
be eliminated. Thus the query plan size is a function of
the number of partitions, as can be seen in Figure 18(b).
As seen in the figure, the measured plan size for Orca also
shows some dependence on the number of partitions. This is
due to a limitation of the way metadata is replicated on the
segment nodes, which requires some part of the metadata
required by the partitioning functions in Section 3.2 to be
embedded in the query plan structure and shipped together
!"#$%
!&#$%
!'#$%
#$%
'#$%
&#$%
"#$%
(#$%
)##$%
!"#$%&'"'()*
+,'-./%(*01"'*
*+,-.!-/00102% 3451/3% 6,02!-/00102%
Figure 17: Relative improvement in execution time
when partition selection is enabled
to the segment nodes. The actual query plan size is indeed
independent of the number of partitions and has the shape
of the plan shown in Figure 8(b).
4.4.3 DML Queries over Partitioned Tables
In this experiment, we use tables R and S from Sec-
tion 4.4.2. We execute the following DML statement which
updates table R with values from S:
update R set b=S.b from S where R.a=S.a;
Figure 18(c) shows that the plan size for Planner-
produced plans grows quadratically with the number of par-
titions in the tables, as the plan needs to enumerate all join
combinations between the individual parts. In Orca, this
enumeration is not necessary, so the plan size remains nearly
the same. Again, the small variations in Orca’s plan size are
due to inefficiencies of the metadata catalog as pointed out
in the previous experiment, and not to the actual plan shape.
5. RELATED WORK
In this section, we review related efforts in exploiting pos-
sible query optimization opportunities when querying parti-
tioned tables. Most DBMSs (IBM DB2 [9], Oracle [11], Mi-
crosoft SQL Server [16]) support both single-level and multi-
level partitioning. In most of these systems, static partition
elimination takes place during query optimization, where se-
lection predicates are used to determine which parts to scan,
and only these parts are referenced in the execution plan.
Dynamic partition elimination works by binding the join
variables at run time, and computing qualifying partitions
for scanning. Additionally, Oracle [11] supports partition-
wise joins for cases where two partitioned tables are joined
on their corresponding partitioning key. In this case, the join
is broken down to smaller joins, where every partition from
the first table is only joined with the matching partitions
from the other table. While some form of dynamic partition
elimination is implemented in those systems, we could only
find a handful of simple examples of single-level equality
joins, with no description of how the query optimizer selects
these plans, and whether the techniques work for complex
queries or handle only simple join patterns. To the best of
our knowledge, our paper is the first comprehensive study of
how to implement dynamic partition elimination techniques

!"
#!"
$!!"
$#!"
%!!"
%#!"
$&" %&" #!&" '#&" $!!&"
!"#$%&'()%*+,-%
!)./)$0#1)%23%4#.552$&%&/#$$)6%
()*++,-" .-/*"
(a) Static partition elimination
!"
#!!"
$!!"
%!!"
&!!"
'!!!"
'#!!"
'$!!"
'%!!"
'&!!"
(!" '!!" '(!" #!!" #(!" )!!"
!"#$%&'()%*+,-%
./01)2%34%5#2663$&%'$%78)%&98)0#%
*+,--./" 0/1,"
(b) Dynamic partition elimination
!"
#!!!"
$!!!!"
$#!!!"
%!!!!"
%#!!!"
#!" $!!" $#!" %!!" %#!" &!!"
!"#$%&'()%*+,-%
./01)2%34%5#2663$&%'$%78)%&98)0#%
'()**+," -,.)"
(c) DML queries
Figure 18: Plan size comparison between Planner and Orca for different query patterns
for complex queries and integrate them in a general purpose
query optimizer for an MPP system.
A recent research effort [7] added sophisticated optimiza-
tion techniques for querying partitioned tables to the Post-
greSQL optimizer. The authors assume a star/snowflake
schema where the dimension and the fact tables are par-
titioned on the same column(s). Their approach works on
partition-wise joins and eliminates joins that cannot pro-
duce results based on subsumption of the partitioning con-
straints. Although the approach works well for the assumed
schema, it does not decouple the plan size from the number
of partitions. This is a significant issue when dealing with
thousands of partitions. In addition, it does not directly
apply to massively parallel processing systems.
As Hadoop quickly becomes a popular ecosystem for
big data analytics, many SQL on Hadoop solutions have
been proposed. Open-source approaches, such as Cloud-
era’s Impala [10], Facebook’s Presto [5] and Hortonworks’
Stinger [8] only support static partition elimination inher-
ited from Hive [4]. Although dynamic partition elimination
is proposed [13], it is not implemented yet, while Pivotal’s
HAWQ [12] utilizes Orca and supports both static and dy-
namic partition elimination in a unified way.
Finding good partitioning schemes is part of database
physical design tuning and is outside the scope of this pa-
per. However, various techniques to find a good partitioning
scheme automatically have been proposed [3, 14].
6. SUMMARY
The distributed nature of MPP database systems poses
unique challenges when dealing with partitioned tables. The
key if finding an abstraction that meshes well with both the
elementary building blocks of modern query optimizers and
yet meets performance-oriented scalability requirements.
In this paper, we presented a methodology that represents
partitioned tables and all relevant methods to access them
as elements of an algebra. The resulting framework enables
the query optimizer to explore a variety of plan alternatives.
Most importantly, our design emphasizes dynamic plans in
which desicions to access a partition—at potentially consid-
erable I/O cost—are deferred to query execution in order to
take advantage of specific characteristics of the underlying
data, which are unknown at optimization time.
The system presented is implemented in Pivotal Green-
plum Database as part of the Orca optimizer initiative and
has proven highly effective in production deployments.
For future work, we plan to address a number of advanced
subjects including indexing, better modeling of costs, and
completeness of interoperability with other optimizations.
7. REFERENCES
[1] TPC-DS. http://www.tpc.org/tpcds, 2005.
[2] TPC-H. http://www.tpc.org/tpch, 2009.
[3] S. Agrawal, V. Narasayya, and B. Yang. Integrating
Vertical and Horizontal Partitioning into Automated
Physical Database Design. In SIGMOD, 2004.
[4] Apache. Hive. http://hive.apache.org/, 2013.
[5] L. Chan. Presto: Interacting with petabytes of data at
Facebook. http://prestodb.io, 2013.
[6] G. Graefe. The Cascades Framework for Query
Optimization. IEEE Data Eng. Bul., 18(3), 1995.
[7] H. Herodotou, N. Borisov, and S. Babu. Query
Optimization Techniques for Partitioned Tables. In
SIGMOD, 2011.
[8] Hortonworks. Stinger, Interactive query for Apache
Hive. http://hortonworks.com/labs/stinger/, 2013.
[9] IBM. DB2 Partitioned Tables.
http://publib.boulder.ibm.com/infocenter/
db2luw/v9r7/topic/com.ibm.db2.luw.admin.
partition.doc/doc/c0021560.html, 2007.
[10] M. Kornacker and J. Erickson. Cloudera Impala:
Real-Time Queries in Apache Hadoop, for Real.
http://www.cloudera.com/content/cloudera/en/
products-and-services/cdh/impala.html, 2012.
[11] T. Morales. Oracle Database VLDB and Partitioning
Guide 11g Release 1 (11.1). 2007.
[12] Pivotal. HD: HAWQ. http://www.gopivotal.com/
sites/default/files/Hawq_WP_042313_FINAL.pdf,
2013.
[13] L. J. Pullokkaran and L. Leverenz. MapJoin and
Partition Pruning.
https://cwiki.apache.org/confluence/display/
Hive/MapJoin+and+Partition+Pruning, 2013.
[14] J. Rao, C. Zhang, N. Megiddo, and G. M. Lohman.
Automating Physical Database Design in a Parallel
Database. In SIGMOD, 2002.
[15] M. A. Soliman, L. Antova, V. Raghavan, A. El-Helw,
Z. Gu, E. Shen, M. Petropoulos, F. Waas,
S. Narayanan, K. Krikellas, and R. Baldwin. Orca: A
Modular Query Optimizer Architecture for Big Data.
In SIGMOD, 2014.
[16] R. Talmage. Partitioned Table and Index Strategies
Using SQL Server 2008. 2009.

Optimizing Queries over Partitioned Tables in MPP Systems

More Related Content

What's hot (19)

Viewers also liked (20)

Similar to Optimizing Queries over Partitioned Tables in MPP Systems (20)

More from EMC (20)

Recently uploaded (20)

Optimizing Queries over Partitioned Tables in MPP Systems