Application of postgre sql to large social infrastructure

Copyright © 2016 NTT DATA Corporation
December 2, 2016
NTT Data Corporation
Ayumi Ishii
Application of PostgreSQL to large social
infrastructure
PGCONF.ASIA 2016

Copyright © 2016 NTT DATA Corporation 2
How to use PostgreSQL in social infrastructure

3Copyright © 2016 NTT DATA Corporation
Positioning of smart meter management system
aggregation
device
SM
SM
SM
smart meter
management
system
SM
Data Center
SM
SM
SM
aggregation device
wheeling
management system
fee calculation for
new menu
other
power
companies
billing
processing
member management
system
reward points system
switching support
system
Organization
for Cross-
regional
Coordination
of
Transmission
Operators
★

Main processing and mission of the system
main processing
5 million datasets
per 30 min
validate
save
data
save
calculated datacalculation
within 10minutes
• 240 million additional tuples per
day
• must be saved for 24 months
5 million
tuple
INSERT
Mission 1
Mission 2
large scale
SELECT
Mission 35 million
tuple
INSERT

Mission
1. Load 10 million datasets within 10 minutes !
2. Must save data for 24 months !
3. Stabilize large scale SELECT performance !

(1) Load 10 million datasets within 10 minutes !
★
main processing
5 million datasets
per 30 min
validate
save
data
save
within 10minutes
day
5 million
tuple
INSERT
Mission 2
large scale
SELECT
Mission 35 million
tuple
INSERT
Mission 1

Data model
data : [Device ID] [Date] [Electricity Usage]
ex) ID: 1 used 500 at 1:00 August 1st.
Method 1 ：UPDATE model
UPDATE new data for each device, daily
Device
ID
Day 0:00 0:30 1:00 1:30 …
1 8/1 100 300 500
2 8/1 200 400
Frequent UPADATEs are unfavorable for
PostgreSQL in terms of performance

Data model
Device
ID
Date Value
1 8/1 0:00 100
1 8/1 0:30 300
1 8/1 1:00 500
… … …
○ performance
× data size
Method 2 : INSERT model
INSERT new data for each device, every 30 mins
Device
ID
Day 0:00 0:30 1:00 1:30 …
1 8/1 100 300 500
2 8/1 200 400

Data model
Device
ID
Date Value
1 8/1 0:00 100
1 8/1 0:30 300
1 8/1 1:00 500
… … …
○ performance
× data size
Method 2 : INSERT model
INSERT new data for each device, every 30 mins
Device
ID
Day 0:00 0:30 1:00 1:30 …
1 8/1 100 300 500
2 8/1 200 400
Selected based on performance

Performance factors
number of tuples
in one transaction ?
multiplicity？ parameters?
data type？
restrictions?
index?
version?
pre research regarding performance factors
how to load to
partition table？

Performance factors
number of tuples
in one transaction
10000multiplicity
8
parameter
wal_bugffers=1GB
data type
minimumrestriction
minimum
index
minimum
version
9.4
direct load to
partition child table
DB design
performance tuning

Performance factors
number of tuples
in one transaction
10000multiplicity
8
parameter
wal_bugffers=1GB
data type
minimumrestriction
minimum
index
minimum
version
9.4
direct load to
partition child table

Bottleneck Analysis with perf
19.83% postgres postgres [.] XLogInsert ★
6.45% postgres postgres [.] LWLockRelease
4.41% postgres postgres [.] PinBuffer
3.03% postgres postgres [.] LWLockAcquire
WAL is the
bottleneck ！
perf
WAL
WAL
file
Disk
I/O
memory
WAL buffer
write
・commit
・buffer is full

wal_buffers parameter
“The auto-tuning selected by the default
setting of -1 should give reasonable results
in most cases.”
by PostgreSQL Document

wal_buffers
※INSERT only
（except SELECT）
0:00:00
0:01:00
0:02:00
0:03:00
0:04:00
0:05:00
0:06:00
0:07:00
0:08:00
0:09:00
16MB 1GB
Time
Impact of WAL_buffers

PostgreSQL version
・WAL performance improved
・JSONB
・GIN performance improved
・CONCURRENTLY option
9.3 9.4

Version up
• We had originally planned to use 9.3, but changed to 9.4.
0:00:00
0:01:00
0:02:00
0:03:00
0:04:00
0:05:00
0:06:00
0:07:00
0:08:00
9.3 9.4
time
impact of version up
※INSERT only
（except SELECT）

0:07:57
0:06:59
0:05:49
0:03:29
0:03:29
0:03:29
0:00:00
0:02:00
0:04:00
0:06:00
0:08:00
0:10:00
0:12:01
9.3, 16MB 9.3, 1GB 9.4, 1GB
time
Result
target
accomplished!!
other processes
are already
tuned.
■INSERT
■others

(2) Must save data for 24 months !
★
main processing
5 million datasets
per 30 min
validate
save
data
save
within 10minutes
day
5 million
tuple
INSERT
large scale
SELECT
Mission 35 million
tuple
INSERT
Mission 1
Mission 2

Reduce data size by selecting the best data type
• Integer
 Use the smallest data type that can cover the range and precision
• Boolean
 Use BOOLEAN instead of CHAR(1)
Type precision Size
SMALLINT 4 digit 2 byte
INTEGER 9 digit 4 byte
BIGINT 18 digit 8 byte
NUMERIC 1000 digit 3 or 6 or 8 + ceiling(digit / 4) * 2
Type available data Size
CHAR(1) string (length is 1) 5 byte
BOOLEAN true or false 1 byte

Reduce the data size by changing column order
• alignment
• PostgreSQL does not store data across the alignment
1 2 3 4 5 6 7 8
column_1(4byte) ***PADDING***
column_2(8byte)
8 byte

Column Type
column_1 integer
column_2 timestamp without time zone
column_3 integer
column_4 smallint
column_6 smallint
1 2 3 4 5 6 7 8
column_1 ***PADDING***
column_2
column_3 column_4 *PADDING*
column_5
column_6 ********PADDING*********
column_7
1 2 3 4 5 6 7 8
column_2
column_5
column_7
column_1 column_3
column_4 column_6
72 60
ex)
12 type / 1 tuple
 2.8GB /day！

Change data model
num data select
frequency
update
frequency
policy model
1 1st day
～65th day
high high performance is the
priority
INSERT
2 66th day
～24 months
low low data size is the
priority
UPDATE
We adopted INSERT model considering the performance
• However, data size is large making it difficult to store long term
convert model for old data

Change data model
ID date 0:00 0:30 1:00 … 22:30 23:00 23:30
1 8/1 100 300 500 … 1000 1100 1200
2 8/1 100 200 300 … 800 900 1000
ID timestamp value
1 8/1 0:00 100
2 8/1 0:00 100
1 8/1 0:30 300
2 8/1 0:30 200
1 8/1 1:00 500
2 8/1 1:00 300
… … …
1 8/1 22:30 1000
2 8/1 22:30 800
1 8/1 23:00 1100
2 8/1 23:00 900
1 8/1 23:30 1200
2 8/1 23:30 1000
INSERT model UPDATE model
remove duplicated data (ID, timestamp)
num of tuples/day： 240 million →5 million
size： 22GB→3GB

result
108
11
0
20
40
60
80
100
120
datasize（TB)
reduce data size
before after

(3) Stabilize large scale SELECT performance !
★
main processing
5 million datasets
per 30 min
validate
save
data
save
within 10minutes
day
5 million
tuple
INSERT
large scale
SELECT
5 million
tuple
INSERT
Mission 1
Mission 2
Mission 3

Stabilize the performance of 10 million SELECT statements!
“stable performance” is important
• Performance degradation is caused by sudden changes in
execution plan is problem
control
execution plans
pg_hint_plan
lock statistical
information
pg_dbms_stats
stable performance

Before using pg_hint_plan & pg_dbms_stats
In most cases, optimizer generates the best execution plan
fixing execution plan does not always bring good result
• The best execution plan at this time may not be best in the future.
However, it is necessary to reduce the risk.
If execution plan suddenly changed during operation, and
performance maybe reduced.
→Understand the demerits and use these extensions
• SELECT immediately after batch, before
ANALYZE
• SELECT from a lot of tables (JOIN)
• …

pg_dbms_stats
Planner
pg_dbms_stats
PostgreSQL
Original
statistics
Plan
generate
Lock
“Locked”
statistics

pg_dbms_stats in this system
usage
data
day
table
locked
statistics
day
table
locked
statistics
day
table
locked
statistics
day partition
set locked statistics with new table
COPY some statistics are different
depending on each child table
We can certainly get best plan even without
using ANALYZE.
• table’s OID, table name
• partition key, date

Replacing statistics that should be changed according to table
• Create assumed dummy data
• ANALYZE dummy data
Column statistic
partition key Most Common Value
Date Histogram
Ex) “ 8/1 0:00” , “8/1 0:30”, “8/1 1:00”
48 pattern per day. Uniform distribution.

1. Load 10 million datasets within 10 minutes !
2. Must save data for 24 months !
3. Stabilize large scale SELECT performance !
Mission
COMPLETE

conclusion
The 20th anniversary of PostgreSQL
PostgreSQL finally evolved to be adopted in large scale social infrastructure.
Both PostgreSQL technical knowledge and business application knowledge are necessary
to be successful in difficult and large scale projects.
Pre research and know-how are important to get the full out of PostgreSQL.

Application of postgre sql to large social infrastructure

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Application of postgre sql to large social infrastructure (20)

More from NTT DATA OSS Professional Services (15)

Recently uploaded (20)

Application of postgre sql to large social infrastructure