1. HAWQ
Architecture
Alexey Grishchenko

2. Who I am
Enterprise Architect @ Pivotal
• 7 years in data processing
• 5 years of experience with MPP
• 4 years with Hadoop
• Using HAWQ since the first internal Beta
• Responsible for designing most of the EMEA HAWQ
and Greenplum implementations
• Spark contributor
• http://0x0fff.com

3. Agenda
• What is HAWQ

4. Agenda
• What is HAWQ
• Why you need it

5. Agenda
• What is HAWQ
• Why you need it
• HAWQ Components

6. Agenda
• What is HAWQ
• Why you need it
• HAWQ Components
• HAWQ Design

7. Agenda
• What is HAWQ
• Why you need it
• HAWQ Components
• HAWQ Design
• Query execution example

8. Agenda
• What is HAWQ
• Why you need it
• HAWQ Components
• HAWQ Design
• Query execution example
• Competitive solutions

9. What is
• Analytical SQL-on-Hadoop engine

10. What is
• Analytical SQL-on-Hadoop engine
• HAdoop With Queries

11. What is
• Analytical SQL-on-Hadoop engine
• HAdoop With Queries
Postgres Greenplum HAWQ
2005
Fork
Postgres 8.0.2

12. What is
• Analytical SQL-on-Hadoop engine
• HAdoop With Queries
Postgres HAWQ
2005
Fork
Postgres 8.0.2
2009
Rebase
Postgres 8.2.15
Greenplum

13. What is
• Analytical SQL-on-Hadoop engine
• HAdoop With Queries
Postgres HAWQ
2005
Fork
Postgres 8.0.2
2009
Rebase
Postgres 8.2.15
2011 Fork
GPDB 4.2.0.0
Greenplum

14. What is
• Analytical SQL-on-Hadoop engine
• HAdoop With Queries
Postgres HAWQ
2005
Fork
Postgres 8.0.2
2009
Rebase
Postgres 8.2.15
2011 Fork
GPDB 4.2.0.0
2013
HAWQ 1.0.0.0
Greenplum

15. What is
• Analytical SQL-on-Hadoop engine
• HAdoop With Queries
Postgres HAWQ
2005
Fork
Postgres 8.0.2
2009
Rebase
Postgres 8.2.15
2011 Fork
GPDB 4.2.0.0
2013
HAWQ 1.0.0.0
HAWQ 2.0.0.0
Open Source
2015
Greenplum

16. HAWQ is …
• 1’500’000 C and C++ lines of code

17. HAWQ is …
• 1’500’000 C and C++ lines of code
– 200’000 of them in headers only

18. HAWQ is …
• 1’500’000 C and C++ lines of code
– 200’000 of them in headers only
• 180’000 Python LOC

19. HAWQ is …
• 1’500’000 C and C++ lines of code
– 200’000 of them in headers only
• 180’000 Python LOC
• 60’000 Java LOC

20. HAWQ is …
• 1’500’000 C and C++ lines of code
– 200’000 of them in headers only
• 180’000 Python LOC
• 60’000 Java LOC
• 23’000 Makefile LOC

21. HAWQ is …
• 1’500’000 C and C++ lines of code
– 200’000 of them in headers only
• 180’000 Python LOC
• 60’000 Java LOC
• 23’000 Makefile LOC
• 7’000 Shell scripts LOC

22. HAWQ is …
• 1’500’000 C and C++ lines of code
– 200’000 of them in headers only
• 180’000 Python LOC
• 60’000 Java LOC
• 23’000 Makefile LOC
• 7’000 Shell scripts LOC
• More than 50 enterprise customers

23. HAWQ is …
• 1’500’000 C and C++ lines of code
– 200’000 of them in headers only
• 180’000 Python LOC
• 60’000 Java LOC
• 23’000 Makefile LOC
• 7’000 Shell scripts LOC
• More than 50 enterprise customers
– More than 10 of them in EMEA

24. Apache HAWQ
• Apache HAWQ (incubating) from 09’2015
– http://hawq.incubator.apache.org
– https://github.com/apache/incubator-hawq
• What’s in Open Source
– Sources of HAWQ 2.0 alpha
– HAWQ 2.0 beta is planned for 2015’Q4
– HAWQ 2.0 GA is planned for 2016’Q1
• Community is yet young – come and join!

25. Why do we need it?

26. Why do we need it?
• SQL-interface for BI solutions to the Hadoop
data complaint with ANSI SQL-92, -99, -2003

27. Why do we need it?
• SQL-interface for BI solutions to the Hadoop
data complaint with ANSI SQL-92, -99, -2003
– Example - 5000-line query with a number of
window function generated by Cognos

28. Why do we need it?
• SQL-interface for BI solutions to the Hadoop
data complaint with ANSI SQL-92, -99, -2003
– Example - 5000-line query with a number of
window function generated by Cognos
• Universal tool for ad hoc analytics on top of
Hadoop data

29. Why do we need it?
• SQL-interface for BI solutions to the Hadoop
data complaint with ANSI SQL-92, -99, -2003
– Example - 5000-line query with a number of
window function generated by Cognos
• Universal tool for ad hoc analytics on top of
Hadoop data
– Example - parse URL to extract protocol, host
name, port, GET parameters

30. Why do we need it?
• SQL-interface for BI solutions to the Hadoop
data complaint with ANSI SQL-92, -99, -2003
– Example - 5000-line query with a number of
window function generated by Cognos
• Universal tool for ad hoc analytics on top of
Hadoop data
– Example - parse URL to extract protocol, host
name, port, GET parameters
• Good performance

31. Why do we need it?
• SQL-interface for BI solutions to the Hadoop
data complaint with ANSI SQL-92, -99, -2003
– Example - 5000-line query with a number of
window function generated by Cognos
• Universal tool for ad hoc analytics on top of
Hadoop data
– Example - parse URL to extract protocol, host
name, port, GET parameters
• Good performance
– How many times the data would hit the HDD during
a single Hive query?

32. HAWQ Cluster
Server 1
SNameNode
Server 4
ZK JM
NameNode
Server 3
ZK JM
Server 2
ZK JM
Server 6
Datanode
Server N
Datanode
Server 5
Datanode
interconnect
…

33. HAWQ Cluster
Server 1
SNameNode
Server 4
ZK JM
NameNode
Server 3
ZK JM
Server 2
ZK JM
Server 6
Datanode
Server N
Datanode
Server 5
Datanode
YARN NM YARN NM YARN NM
YARN RM
YARN App
Timeline
interconnect
…

34. HAWQ Cluster
HAWQ Master
Server 1
SNameNode
Server 4
ZK JM
NameNode
Server 3
ZK JM
HAWQ
Standby
Server 2
ZK JM
HAWQ Segment
Server 6
Datanode
HAWQ Segment
Server N
Datanode
HAWQ Segment
Server 5
Datanode
YARN NM YARN NM YARN NM
YARN RM
YARN App
Timeline
interconnect
…

35. Master Servers
Server 1
SNameNode
Server 4
ZK JM
NameNode
Server 3
ZK JM
Server 2
ZK JM
HAWQ Segment
Server 6
Datanode
HAWQ Segment
Server N
Datanode
HAWQ Segment
Server 5
Datanode
YARN NM YARN NM YARN NM
YARN RM
YARN App
Timeline
interconnect
…
HAWQ Master
HAWQ
Standby

36. Master Servers
HAWQ Master
Query Parser
Query
Optimizer
Global
Resource
Manager
Distributed
Transactions
Manager
Query Dispatch
Metadata
Catalog
HAWQ Standby Master
Query Parser
Query
Optimizer
Global Resource
Manager
Distributed
Transactions
Manager
Query Dispatch
Metadata
Catalog
WAL
repl.

37. HAWQ Master
HAWQ
Standby
Segments
Server 1
SNameNode
Server 4
ZK JM
NameNode
Server 3
ZK JM
Server 2
ZK JM
Server 6
Datanode
Server N
Datanode
Server 5
Datanode
YARN NM YARN NM YARN NM
YARN RM
YARN App
Timeline
interconnect
HAWQ Segment HAWQ SegmentHAWQ Segment …

38. Segments
HAWQ Segment
Query Executor
libhdfs3
PXF
HDFS Datanode
Local Filesystem
Temporary Data
Directory
Logs
YARN Node Manager

39. Metadata
• HAWQ metadata structure is similar to
Postgres catalog structure

40. Metadata
• HAWQ metadata structure is similar to
Postgres catalog structure
• Statistics
– Number of rows and pages in the table

41. Metadata
• HAWQ metadata structure is similar to
Postgres catalog structure
• Statistics
– Number of rows and pages in the table
– Most common values for each field

42. Metadata
• HAWQ metadata structure is similar to
Postgres catalog structure
• Statistics
– Number of rows and pages in the table
– Most common values for each field
– Histogram of values distribution for each field

43. Metadata
• HAWQ metadata structure is similar to
Postgres catalog structure
• Statistics
– Number of rows and pages in the table
– Most common values for each field
– Histogram of values distribution for each field
– Number of unique values in the field

44. Metadata
• HAWQ metadata structure is similar to
Postgres catalog structure
• Statistics
– Number of rows and pages in the table
– Most common values for each field
– Histogram of values distribution for each field
– Number of unique values in the field
– Number of null values in the field

45. Metadata
• HAWQ metadata structure is similar to
Postgres catalog structure
• Statistics
– Number of rows and pages in the table
– Most common values for each field
– Histogram of values distribution for each field
– Number of unique values in the field
– Number of null values in the field
– Average width of the field in bytes

46. Statistics
No Statistics
How many rows would produce the join of two
tables?

47. Statistics
No Statistics
How many rows would produce the join of two
tables?
 From 0 to infinity

48. Statistics
No Statistics
Row Count
How many rows would produce the join of two
tables?
 From 0 to infinity
How many rows would produce the join of two 1000-
row tables?

49. Statistics
No Statistics
Row Count
How many rows would produce the join of two
tables?
 From 0 to infinity
How many rows would produce the join of two 1000-
row tables?
 From 0 to 1’000’000

50. Statistics
No Statistics
Row Count
Histograms and MCV
How many rows would produce the join of two
tables?
 From 0 to infinity
How many rows would produce the join of two 1000-
row tables?
 From 0 to 1’000’000
How many rows would produce the join of two 1000-
row tables, with known field cardinality, values
distribution diagram, number of nulls, most common
values?

51. Statistics
No Statistics
Row Count
Histograms and MCV
How many rows would produce the join of two
tables?
 From 0 to infinity
How many rows would produce the join of two 1000-
row tables?
 From 0 to 1’000’000
How many rows would produce the join of two 1000-
row tables, with known field cardinality, values
distribution diagram, number of nulls, most common
values?
 ~ From 500 to 1’500

52. Metadata
• Table structure information
ID Name Num Price
1 Яблоко 10 50
2 Груша 20 80
3 Банан 40 40
4 Апельсин 25 50
5 Киви 5 120
6 Арбуз 20 30
7 Дыня 40 100
8 Ананас 35 90

53. Metadata
• Table structure information
– Distribution fields
ID Name Num Price
1 Яблоко 10 50
2 Груша 20 80
3 Банан 40 40
4 Апельсин 25 50
5 Киви 5 120
6 Арбуз 20 30
7 Дыня 40 100
8 Ананас 35 90
hash(ID)

54. Metadata
• Table structure information
– Distribution fields
– Number of hash buckets
ID Name Num Price
1 Яблоко 10 50
2 Груша 20 80
3 Банан 40 40
4 Апельсин 25 50
5 Киви 5 120
6 Арбуз 20 30
7 Дыня 40 100
8 Ананас 35 90
hash(ID)
ID Name Num Price
1 Яблоко 10 50
2 Груша 20 80
3 Банан 40 40
4 Апельсин 25 50
5 Киви 5 120
6 Арбуз 20 30
7 Дыня 40 100
8 Ананас 35 90

55. Metadata
• Table structure information
– Distribution fields
– Number of hash buckets
– Partitioning (hash, list, range)
ID Name Num Price
1 Яблоко 10 50
2 Груша 20 80
3 Банан 40 40
4 Апельсин 25 50
5 Киви 5 120
6 Арбуз 20 30
7 Дыня 40 100
8 Ананас 35 90
hash(ID)
ID Name Num Price
1 Яблоко 10 50
2 Груша 20 80
3 Банан 40 40
4 Апельсин 25 50
5 Киви 5 120
6 Арбуз 20 30
7 Дыня 40 100
8 Ананас 35 90

56. Metadata
• Table structure information
– Distribution fields
– Number of hash buckets
– Partitioning (hash, list, range)
• General metadata
– Users and groups

57. Metadata
• Table structure information
– Distribution fields
– Number of hash buckets
– Partitioning (hash, list, range)
• General metadata
– Users and groups
– Access privileges

58. Metadata
• Table structure information
– Distribution fields
– Number of hash buckets
– Partitioning (hash, list, range)
• General metadata
– Users and groups
– Access privileges
• Stored procedures
– PL/pgSQL, PL/Java, PL/Python, PL/Perl, PL/R

59. Query Optimizer
• HAWQ uses cost-based query optimizers

60. Query Optimizer
• HAWQ uses cost-based query optimizers
• You have two options
– Planner – evolved from the Postgres query
optimizer
– ORCA (Pivotal Query Optimizer) – developed
specifically for HAWQ

61. Query Optimizer
• HAWQ uses cost-based query optimizers
• You have two options
– Planner – evolved from the Postgres query
optimizer
– ORCA (Pivotal Query Optimizer) – developed
specifically for HAWQ
• Optimizer hints work just like in Postgres
– Enable/disable specific operation
– Change the cost estimations for basic actions

62. Storage Formats
Which storage format is the most optimal?

63. Storage Formats
Which storage format is the most optimal?
 It depends on what you mean by “optimal”

64. Storage Formats
Which storage format is the most optimal?
 It depends on what you mean by “optimal”
– Minimal CPU usage for reading and writing the data

65. Storage Formats
Which storage format is the most optimal?
 It depends on what you mean by “optimal”
– Minimal CPU usage for reading and writing the data
– Minimal disk space usage

66. Storage Formats
Which storage format is the most optimal?
 It depends on what you mean by “optimal”
– Minimal CPU usage for reading and writing the data
– Minimal disk space usage
– Minimal time to retrieve record by key

67. Storage Formats
Which storage format is the most optimal?
 It depends on what you mean by “optimal”
– Minimal CPU usage for reading and writing the data
– Minimal disk space usage
– Minimal time to retrieve record by key
– Minimal time to retrieve subset of columns
– etc.

68. Storage Formats
• Row-based storage format
– Similar to Postgres heap storage
• No toast
• No ctid, xmin, xmax, cmin, cmax

69. Storage Formats
• Row-based storage format
– Similar to Postgres heap storage
• No toast
• No ctid, xmin, xmax, cmin, cmax
– Compression
• No compression
• Quicklz
• Zlib levels 1 - 9

70. Storage Formats
• Apache Parquet
– Mixed row-columnar table store, the data is split
into “row groups” stored in columnar format

71. Storage Formats
• Apache Parquet
– Mixed row-columnar table store, the data is split
into “row groups” stored in columnar format
– Compression
• No compression
• Snappy
• Gzip levels 1 – 9

72. Storage Formats
• Apache Parquet
– Mixed row-columnar table store, the data is split
into “row groups” stored in columnar format
– Compression
• No compression
• Snappy
• Gzip levels 1 – 9
– The size of “row group” and page size can be set
for each table separately

73. Resource Management
• Two main options
– Static resource split – HAWQ and YARN does not
know about each other

74. Resource Management
• Two main options
– Static resource split – HAWQ and YARN does not
know about each other
– YARN – HAWQ asks YARN Resource Manager for
query execution resources

75. Resource Management
• Two main options
– Static resource split – HAWQ and YARN does not
know about each other
– YARN – HAWQ asks YARN Resource Manager for
query execution resources
• Flexible cluster utilization
– Query might run on a subset of nodes if it is small

76. Resource Management
• Two main options
– Static resource split – HAWQ and YARN does not
know about each other
– YARN – HAWQ asks YARN Resource Manager for
query execution resources
• Flexible cluster utilization
– Query might run on a subset of nodes if it is small
– Query might have many executors on each cluster
node to make it run faster

77. Resource Management
• Two main options
– Static resource split – HAWQ and YARN does not
know about each other
– YARN – HAWQ asks YARN Resource Manager for
query execution resources
• Flexible cluster utilization
– Query might run on a subset of nodes if it is small
– Query might have many executors on each cluster
node to make it run faster
– You can control the parallelism of each query

78. Resource Management
• Resource Queue can be set with
– Maximum number of parallel queries

79. Resource Management
• Resource Queue can be set with
– Maximum number of parallel queries
– CPU usage priority

80. Resource Management
• Resource Queue can be set with
– Maximum number of parallel queries
– CPU usage priority
– Memory usage limits

81. Resource Management
• Resource Queue can be set with
– Maximum number of parallel queries
– CPU usage priority
– Memory usage limits
– CPU cores usage limit

82. Resource Management
• Resource Queue can be set with
– Maximum number of parallel queries
– CPU usage priority
– Memory usage limits
– CPU cores usage limit
– MIN/MAX number of executors across the system

83. Resource Management
• Resource Queue can be set with
– Maximum number of parallel queries
– CPU usage priority
– Memory usage limits
– CPU cores usage limit
– MIN/MAX number of executors across the system
– MIN/MAX number of executors on each node

84. Resource Management
• Resource Queue can be set with
– Maximum number of parallel queries
– CPU usage priority
– Memory usage limits
– CPU cores usage limit
– MIN/MAX number of executors across the system
– MIN/MAX number of executors on each node
• Can be set up for user or group

85. External Data
• PXF
– Framework for external data access
– Easy to extend, many public plugins available
– Official plugins: CSV, SequenceFile, Avro, Hive,
HBase
– Open Source plugins: JSON, Accumulo,
Cassandra, JDBC, Redis, Pipe

86. External Data
• PXF
– Framework for external data access
– Easy to extend, many public plugins available
– Official plugins: CSV, SequenceFile, Avro, Hive,
HBase
– Open Source plugins: JSON, Accumulo,
Cassandra, JDBC, Redis, Pipe
• HCatalog
– HAWQ can query tables from HCatalog the same
way as HAWQ native tables

87. Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Resource Prepare Execute Result CleanupPlan

88. Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Resource Prepare Execute Result CleanupPlan
QE

89. Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Resource Prepare Execute Result CleanupPlan
QE

90. Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Resource Prepare Execute Result CleanupPlan
QE

91. Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Resource Prepare Execute Result CleanupPlan
QE

92. Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Resource Prepare Execute Result CleanupPlan
QE ScanBars
b
HashJoinb.name =s.bar
ScanSells
s
Filterb.city ='SanFrancisco'
Projects.beer, s.price
MotionGather
MotionRedist(b.name)

93. Plan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Prepare Execute Result Cleanup
QE
Resource

94. Plan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Prepare Execute Result Cleanup
QE
Resource
I need 5 containers
Each with 1 CPU core
and 256 MB RAM

95. Plan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Prepare Execute Result Cleanup
QE
Resource
I need 5 containers
Each with 1 CPU core
and 256 MB RAM
Server 1: 2 containers
Server 2: 1 container
Server N: 2 containers

96. Plan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Prepare Execute Result Cleanup
QE
Resource
I need 5 containers
Each with 1 CPU core
and 256 MB RAM
Server 1: 2 containers
Server 2: 1 container
Server N: 2 containers

97. Plan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Prepare Execute Result Cleanup
QE
Resource
I need 5 containers
Each with 1 CPU core
and 256 MB RAM
Server 1: 2 containers
Server 2: 1 container
Server N: 2 containers

98. Plan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Prepare Execute Result Cleanup
QE
Resource
I need 5 containers
Each with 1 CPU core
and 256 MB RAM
Server 1: 2 containers
Server 2: 1 container
Server N: 2 containers
QE QE QE QE QE

99. ResourcePlan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Execute Result Cleanup
QE
QE QE QE QE QE
Prepare

100. ResourcePlan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Execute Result Cleanup
QE
QE QE QE QE QE
Prepare
ScanBars
b
HashJoinb.name =s.bar
ScanSells
s
Filterb.city ='SanFrancisco'
Projects.beer, s.price
MotionGather
MotionRedist(b.name)

101. ResourcePlan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Execute Result Cleanup
QE
QE QE QE QE QE
Prepare
ScanBars
b
HashJoinb.name =s.bar
ScanSells
s
Filterb.city ='SanFrancisco'
Projects.beer, s.price
MotionGather
MotionRedist(b.name)

102. ResourcePlan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Result Cleanup
QE
QE QE QE QE QE
Prepare Execute
ScanBars
b
HashJoinb.name =s.bar
ScanSells
s
Filterb.city ='SanFrancisco'
Projects.beer, s.price
MotionGather
MotionRedist(b.name)

103. ResourcePlan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Result Cleanup
QE
QE QE QE QE QE
Prepare Execute
ScanBars
b
HashJoinb.name =s.bar
ScanSells
s
Filterb.city ='SanFrancisco'
Projects.beer, s.price
MotionGather
MotionRedist(b.name)

104. ResourcePlan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Result Cleanup
QE
QE QE QE QE QE
Prepare Execute
ScanBars
b
HashJoinb.name =s.bar
ScanSells
s
Filterb.city ='SanFrancisco'
Projects.beer, s.price
MotionGather
MotionRedist(b.name)

105. ResourcePlan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Cleanup
QE
QE QE QE QE QE
Prepare Execute Result

106. ResourcePlan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Cleanup
QE
QE QE QE QE QE
Prepare Execute Result

107. ResourcePlan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Cleanup
QE
QE QE QE QE QE
Prepare Execute Result

108. ResourcePlan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
QE
QE QE QE QE QE
Prepare Execute Result Cleanup

109. ResourcePlan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
QE
QE QE QE QE QE
Prepare Execute Result Cleanup
Free query resources
Server 1: 2 containers
Server 2: 1 container
Server N: 2 containers

110. ResourcePlan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
QE
QE QE QE QE QE
Prepare Execute Result Cleanup
Free query resources
Server 1: 2 containers
Server 2: 1 container
Server N: 2 containers
OK

111. ResourcePlan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
QE
QE QE QE QE QE
Prepare Execute Result Cleanup
Free query resources
Server 1: 2 containers
Server 2: 1 container
Server N: 2 containers
OK

112. ResourcePlan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
QE
QE QE QE QE QE
Prepare Execute Result Cleanup
Free query resources
Server 1: 2 containers
Server 2: 1 container
Server N: 2 containers
OK

113. ResourcePlan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
QE
QE QE QE QE QE
Prepare Execute Result Cleanup
Free query resources
Server 1: 2 containers
Server 2: 1 container
Server N: 2 containers
OK

114. ResourcePlan
Query Example
HAWQ Master
Metadata
Transaction Mgr.
Query Parser Query Optimizer
Query Dispatch
Resource Mgr.
NameNode
Server 1
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server 2
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
Server N
Local directory
HAWQ Segment
Postmaster
HDFS Datanode
YARN RMPostmaster
Prepare Execute Result Cleanup

115. Query Performance
• Data does not hit the disk unless this cannot be
avoided

116. Query Performance
• Data does not hit the disk unless this cannot be
avoided
• Data is not buffered on the segments unless
this cannot be avoided

117. Query Performance
• Data does not hit the disk unless this cannot be
avoided
• Data is not buffered on the segments unless
this cannot be avoided
• Data is transferred between the nodes by UDP

118. Query Performance
• Data does not hit the disk unless this cannot be
avoided
• Data is not buffered on the segments unless
this cannot be avoided
• Data is transferred between the nodes by UDP
• HAWQ has a good cost-based query optimizer

119. Query Performance
• Data does not hit the disk unless this cannot be
avoided
• Data is not buffered on the segments unless
this cannot be avoided
• Data is transferred between the nodes by UDP
• HAWQ has a good cost-based query optimizer
• C/C++ implementation is more efficient than
Java implementation of competitive solutions

120. Query Performance
• Data does not hit the disk unless this cannot be
avoided
• Data is not buffered on the segments unless
this cannot be avoided
• Data is transferred between the nodes by UDP
• HAWQ has a good cost-based query optimizer
• C/C++ implementation is more efficient than
Java implementation of competitive solutions
• Query parallelism can be easily tuned

121. Competitive Solutions
Hive SparkSQL Impala HAWQ
Query Optimizer

122. Competitive Solutions
Hive SparkSQL Impala HAWQ
Query Optimizer
ANSI SQL

123. Competitive Solutions
Hive SparkSQL Impala HAWQ
Query Optimizer
ANSI SQL
Built-in Languages

124. Competitive Solutions
Hive SparkSQL Impala HAWQ
Query Optimizer
ANSI SQL
Built-in Languages
Disk IO

125. Competitive Solutions
Hive SparkSQL Impala HAWQ
Query Optimizer
ANSI SQL
Built-in Languages
Disk IO
Parallelism

126. Competitive Solutions
Hive SparkSQL Impala HAWQ
Query Optimizer
ANSI SQL
Built-in Languages
Disk IO
Parallelism
Distributions

127. Competitive Solutions
Hive SparkSQL Impala HAWQ
Query Optimizer
ANSI SQL
Built-in Languages
Disk IO
Parallelism
Distributions
Stability

128. Competitive Solutions
Hive SparkSQL Impala HAWQ
Query Optimizer
ANSI SQL
Built-in Languages
Disk IO
Parallelism
Distributions
Stability
Community

129. Roadmap
• AWS and S3 integration

130. Roadmap
• AWS and S3 integration
• Mesos integration

131. Roadmap
• AWS and S3 integration
• Mesos integration
• Better Ambari integration

132. Roadmap
• AWS and S3 integration
• Mesos integration
• Better Ambari integration
• Cloudera, MapR and IBM Hadoop distributions
native support

133. Roadmap
• AWS and S3 integration
• Mesos integration
• Better Ambari integration
• Cloudera, MapR and IBM Hadoop distributions
native support
• Make the SQL-on-Hadoop engine ever!

134. Summary
• Modern SQL-on-Hadoop engine
• For structured data processing and analysis
• Combines the best techniques of competitive
solutions
• Just released to the open source
• Community is very young
Join our community and contribute!

135. Questions
Apache HAWQ
http://hawq.incubator.apache.org
dev@hawq.incubator.apache.org
user@hawq.incubator.apache.org
Reach me on http://0x0fff.com