The document discusses PostgreSQL replication, detailing its features, historical evolution, and operational modes. It highlights replication's importance for high availability, load balancing, and reducing downtime, along with the distinctions between asynchronous and synchronous replication. Additionally, it addresses limitations, failover mechanisms, and monitoring capabilities within PostgreSQL's replication framework.

1. Copyright © 2013 NTT DATA Corporation
27 April 2013
Fujii Masao (twitter: @fujii_masao)
PostgreSQL replication
Inaugural Indian PostgreSQL User Group meetup!

2. 2Copyright © 2013NTT DATA Corporation
Who am I ?
 Database engineer in NTT DATA
• NTT DATA is global IT service and consulting
company
 PostgreSQL developer since 2008
 One of main authors of PostgreSQL replication

3. 3Copyright © 2013NTT DATA Corporation
Index
 What's replication?
 History
 PostgreSQL replication
1. Features
2. Synchronous vs. Asynchronous

4. Copyright © 2013 NTT DATA Corporation 4
What's replication?

5. 5Copyright © 2013 NTT DATA Corporation
What's replication?
ClientClient
ChangeChange
DB servers
Change
Replicate
DB servers
Change
Proxy server
Create a replica of the database on multiple servers

6. 6Copyright © 2013 NTT DATA Corporation
Why is replication required?
Replication is an essential feature to get the system to work properly 24/7!
High
Availability
Load
Balancing
If one server crashes, another can keep providing the service
We can easily reduce the downtime of the system
The load of query execution can be distributed to multiple servers
We can improve the performance of whole system
ClientClient
SQL SQLSQL
High Availability Load Balancing
DB servers DB servers

7. Copyright © 2013 NTT DATA Corporation 7
History

8. 8Copyright © 2013 NTT DATA Corporation
History of PostgreSQL replication
2007
2008
2009
2010
2012
2011
9.0(Sep 2010 release)
• Async reploication
9.1(Sep 2011 release)
• Synchronous replication
• Improve replication monitoring
• pg_basebackup
9.2(Sep 2012 release)
• Cascade replication
• Backup from standby
• New sync rep mode
Slony-I
Bucardo Londiste
Sequoia
PGCluster
PostgresForest
Postgres-R
Mammoth
PL/Proxy
pgpool-II
rubyrep
Postgres-XC
GridSQL
syncreplicator
Replication war !
Historically the community policy has been to avoid putting replication into core. But,
because of large user needs, replication has been introduced in 9.0.
Then, replication keeps evolving steady.

9. Copyright © 2013 NTT DATA Corporation 9
Features

10. 10Copyright © 2013 NTT DATA Corporation
Single master / Multiple standbys
Single master
Multiple standbys Multiple standbys
Replicate Replicate
Write SQL
Read SQL
Read SQL
Read SQL
 Replication from single master to multiple standbys
 Cascade replication
 Only master accepts write query, both accepts read query
Read scalable, not write scalable

11. 11Copyright © 2013 NTT DATA Corporation
Read-only query on standby
Online backup
– (Logical) pg_dump
– (Physical) pg_basebackup
Maintenance command
– VACUUM, ANALYZE
※No need to do VACUUM and
ANALYZE on standby because
the result of mantenance
execution on master is
automatically replicated to
standby
Allow
Query access
– SELECT
– PREPARE, EXECUTE
– CURSOR operation
Disallow
DML
– INSERT, UPDATE, DELETE
– SELECT FOR UPDATE
DDL
– CREATE, DROP, ALTER
Temporary table

12. 12Copyright © 2013 NTT DATA Corporation
Log-shipping
 Transaction log (WAL) is shipped from master to standby
 Standby is in recovery mode
 Standby keeps the database current by replaying shipped WAL
RecoveryMaster Standby
ClientWrite SQL
WAL writing
WAL shipping
WAL writing

13. 13Copyright © 2013 NTT DATA Corporation
Limitation by log-shipping
The followings must be the same between master and standby
 H/W and OS architecture
 PostgreSQL major version
Slony-I can be used for replication between different
architectures and major versions
Master
Standby
64bit OS
PostgreSQL9.2.1
PostgreSQL9.1.0
32bit OS
64bit OS
PostgreSQL9.2.0
NG
NG
OK
2
1

14. 14Copyright © 2013 NTT DATA Corporation
All database objects are replicated
 Per-table granularity is not allowed
Slony-I supports per-table granularity replication
Per database cluster granularity
Per database cluster Per table
Master Standby Master Standby

15. 15Copyright © 2013 NTT DATA Corporation
SQL distribution
PostgreSQL doesn't provide SQL distribution feature
 Implement SQL distribution logic into client application
 Use SQL distributor like pgpool-II
ClientClient
Read SQLWrite SQL
マスタ
Write SQL
Write/Read SQL
Distributor
スタンバイ
Read SQL
Master Standby
Implement logic Use SQL distributor

16. 16Copyright © 2013 NTT DATA Corporation
Failover
PostgreSQL doesn't provide automatic failover feature
 Standby can be promoted to master anytime(pg_ctl promote)
 Automatic error detection and faliover requires clusterware
ClientClient
Master
pgpool-II
StandbyMaster スタンバイ
Pacemaker pgpool-II
VIP
Pacemaker supports the
resource agent for HA cluster
using PostgreSQL replication!

17. 17Copyright © 2013 NTT DATA Corporation
Monitoring
=# SELECT * FROM pg_stat_replication;
-[ RECORD 1 ]----+------------------------------
procpid | 26531
usesysid | 10
usename | postgres
application_name | tokyo
client_addr | 192.168.1.2
client_hostname |
client_port | 39654
backend_start | 2012-02-01 18:54:49.429459+09
state | streaming
sent_location | 0/406E7CC
write_location | 0/406E7CC
flush_location | 0/406E7CC
replay_location | 0/406E1B0
sync_priority | 1
sync_state | sync
Replication progress
How far has master sent WAL?
How far has standby written, flushed or
replayed WAL?
Replication connection information
Standby IP address, Port number,
ROLE for replication、
Replication start time, etc
Replication status
Current synchronous mode,
Standby has already caught up with master?

18. Copyright © 2013 NTT DATA Corporation 18
Synchronous vs. Asynchronous

19. 19Copyright © 2013 NTT DATA Corporation
Replication mode
Replication mode is configurable
 Asynchronous
 Synchronous

20. 20Copyright © 2013 NTT DATA Corporation
Asynchronous replication
COMMIT doesn't wait for the completion of replication
 No guarantee that WAL has already arrived at standby at end of
COMMIT
Data loss window on failover
Query on standby may see outdated data
Low performance impact on master
RecoveryMaster Standby
ClientCOMMIT
WAL writing
WAL shipping
WAL writing
OK
1
2
3
4
5
6
Committed data gets
lost if failover happens
between and
operations
3 4




21. 21Copyright © 2013 NTT DATA Corporation
Synchronous replication
COMMIT waits for completion of replication
 WAL is guaranteed to be flushed in master and standby at end of COMMIT
No data loss on failover!
Relatively high performance impact on master
Query on standby may see outdated data
 「Synchronous ≠ Committed data is available on standby immediately」
Recovery
Master Standby
ClientCOMMIT
WAL writing
WAL shipping
WAL writing
OK
1
2
6
3
4
7
Reply5


WAL shipping is synchronous,
but recovery is asynchronous


22. Copyright © 2011 NTT DATA Corporation
Copyright © 2013 NTT DATA Corporation
(Please omit notations when unnecessary)
This document contains confidential Company information. Do not disclose it to third parties without permission from the Company.