Streaming replication in practice

Streaming replication in practice.
PgConf.Russia 2016, Moscow
Lesovsky Alexey
lesovsky@pgco.me

Working stuff: http://goo.gl/Yy4UzH
I. How streaming replication works.
II. Replication setup.
Practice.
III. Monitoring and maintenance.
IV. Possible problems.
V. Switchiver и Failover.
Practice.
Agenda.

What is replication and it kinds?
Write Ahead Log. REDO and REDO realization in PostgreSQL.
Common replication architecture in PostgreSQL.
Description of processes that involved in replication.
Part I. How replication works.

Synchronize objects
Objects changes are moving to each others.
Replication maybe physical and logical.
What is replication.

Pros:
● Works between different major versions and architectures.
● Allow to replicate tables and tables sets.
Cons:
● Hard to implement synchronous replication.
● CPU overhead (triggers, text conversions, etc).
Examples:
● Slony, Londiste (Skytools), Bucardo, Pglogical.
Logical replication.

Pros:
● Resource usage minimal overhead.
● Easy setup, usage and maintenace.
Cons:
● Standbys are only read-only.
● Can't work with different versions and architectures.
● Can't replicate tables and tables sets.
Physical replication.

Commit all changes in database (Durability in ACID).
Flush all data from REDO buffers at COMMIT.
REDO log has history of all changes in database.
Any changes in database, written into REDO.
REDO log usage:
● in crash recovery;
● backup and Point In Time Recovery;
● replication.
Write Ahead Log. REDO.

In PostgreSQL, REDO also known as Write Ahead Log (WAL).
WAL guaranties that changes are commited before data will changed.
How it guarantied:
LSN (log sequence number) — record location (position) inside WAL;
● Any page is marked with LSN of last record that touched the page;
● Before page is written to disk, bufmgr must check that the WAL flushed to
specified LSN.
Write Ahead Log. REDO implementation in PostgreSQL.

Common view.
backend backend backend autovac ...
WAL writer*
WAL
Startup/Recovery
SR/Archiving/Hot Standby
Point in Time
WAL Sender/WAL Receiver
* - может отсутствовать

Main startup process task is run the database.
In standby mode it initializes infinite replay loop.
Read recovery.conf at REDO start.
REDO:
● read segments from pg_xlog/archive;
● start wal receiver and reading XLOG from upstream server.
When consistency point reached (min recovery ending location) allow
connections and starts checkpointer/bgwriter.
Processing all others parameters from recovery.conf.
More details see in StartupXLOG() function.
Startup process.

For any client postmaster runs dedicated process — backend.
WAL sender is backend too (it has am_walsender flag).
This backend runs exec_replication_command().
exec_replication_command() can do various things:
● create/remove replication slots;
● start basebackup;
● start physical/logical replication.
In last case, backend sends XLOG segments to the client.
Or sleeps when no new XLOG.
WAL Sender process.

Startup process checks XLOG sources.
Startup process init startup of WAL receivers.
Need recovery.conf with primary_conninfo.
WAL receiver:
● check start location for transfer XLOG;
● connects to master and sends start position;
● receive XLOG and write it to disk;
● update variable in shared memory (WalRcv→receivedUpto);
● sends statistics + feedback.
Startup process uses the variable and replay WAL to this location.
WAL Receiver process.

Questions?
backend backend backend autovac ...
WAL writer
WAL
Startup/Recovery
SR/Archiving/Hot Standby
Point in Time
WAL Sender/WAL Receiver

Setup options.
Prepare master for replication.
Tools and utilities.
Start replication and verify results.
Specific options.
Part II. Replication setup.

Synchronous/Asynchronous replication.
Cascade configurations.
Uni-directional/Bi-directional.
Setup options.

Prepare the master.
Copy the DATADIR.
Prepare standbys.
Start standby.
Result verify.
Common logic.

Dedicated user for a replication.
Edit the postgresql.conf.
Edit the pg_hba.conf.
Create replication slots (if required).
Master setup.

● CREATE ROLE ... WITH LOGIN REPLICATION PASSWORD '…';
Master setup.

● wal_level = hot_standby
● max_wal_senders > 0
● Restart the PostgreSQL
Master setup.

● host replication username client_addr/mask authtype
● host replication replica 10.1.0.99/32 md5
● pg_reload_conf()
Master setup.

● max_replication_slots > 0
● pg_create_physical_replication_slot('slotname');
● primary_slot_name = 'slotname' (recovery.conf)
Master setup.

pg_basebackup (since 9.1)
-h, --host=...; -p, --port=...; -U, --username=...; -d, --dbname=...; -D, --pgdata=...
-c, --checkpoint=fast | spread
-X, --xlog-method=fetch | stream – stream с 9.2
-R, --write-recovery-conf – с 9.3
-r, --max-rate=… – с 9.4
--xlogdir=… – с 9.4
-T, --tablespace-mapping=olddir=newdir – с 9.4
-P, --progress
pg_basebackup -P -R -X stream -c fast -h 127.0.0.1 -U replica -D /pgdb
Copy the DATADIR.

Copy with cp, scp, tar, rsync...
Snapshots:
● ZFS send/receive;
● LVM + dd.
pg_start_backup() + pg_stop_backup().
Copy the DATADIR. pg_basebackup alternatives.

Configuration files should be the same.
Configuration files:
● postgresql.conf;
● recovery.conf.
Standby setup.

Configuration files should be the same:
● Why?
● How?
Configuration files:
● postgresql.conf;
● recovery.conf.
Standby setup.

Configuration files (postgresql.conf):
● hot_standby = on;
● max_standby_streaming_delay;
● wal_receiver_status_interval;
● hot_standby_feedback;
● wal_receiver_timeout;
Standby setup.

Configuration files (recovery.conf):
● primary_conninfo = 'host=… port=…'
● standby_mode = on
● primary_slot_name = 'slotname'
● trigger_file = '...'
● recovery_min_apply_delay.
Standby setup.

pg_ctl — PostgreSQL native utility.
pg_ctlcluster — pg_ctl perl wrapper in Debian/Ubuntu.
sysvinit, upstart, openrc, systemd…
Standby start.

wal sender and wal receiver processes.
Check postgres log.
Simple connection with psql.
pg_stat_replication view.
Check results.

DATADIR, configs and Debian-based vs. RHEL-based.
pg_ctlcluster and «unable to connect» errors.
To less processes in «ps» output.
Specific options.

Read scalability.
Full text search.
OLAP.
Standby is not a backup.
Resume. Practical purpose of the replication.

root password: pgconf2016
# su - postgres – working under postgres account.
$ ps auxf – what we have?
$ pwd – where we are?
$ ls -l 9.5
Replication setup. Practice.

$ vi 9.5/data/postgresql.conf
● listen_addresses = '*' – Listen on all interfaces.
● wal_level = hot_standby – Set WAL verbose level.
● max_wal_senders = 4 – Limiting walsenders.
● hot_standby = on – Allow read-only queries on standby.
Master setup.

$ psql – Creare dedicated user.
CREATE ROLE replica WITH LOGIN REPLICATION PASSWORD 'rep123';
$ vi .pgpass – Setup password file.
*:*:*:replica:rep123
$ chmod 600 .pgpass
$ vi 9.5/data/pg_hba.conf – Add auth rules for replication user.
host replication replica 127.0.0.1/32 md5
$ pg_ctl -D 9.5/data/ -m fast restart – Apply the changes.
Master setup.

$ pg_basebackup -P -R -c fast -X stream -h 127.0.0.1 -U replica -D 9.5/replica
● -c fast — do the force checkpoint.
● -X stream — copy new XLOG through dedicated connection.
● -R — create minimal recovery.conf
$ vi 9.5/replica/postgresql.conf – edit port number.
port = 5433
$ pg_ctl -D 9.5/replica/ start – start this standby.
Create the standby.

$ ps auxf – wal sender/receiver process.
$ psql -p 5433 – check the status on the standby.
select pg_is_in_recovery(); – standby nust be in recovery mode.
$ psql – check the status on the master.
select * from pg_stat_replication ; – check statistics from standby.
Check result.

Yes, replication is ready.
Questions.

«Setup and forget» - this is about PostgreSQL streaming replication.
Monitoring:
● Internal statistics;
● auxiliary functions;
● queries examples.
Maintenance:
● add or remove standbys;
● pause replication;
● add or remove slots.
Part III. Monitoring and maintenance.

System views:
● pg_stat_replication
● pg_stat_replication_slots
Monitoring.

pg_is_in_recovery()
pg_current_xlog_location()
pg_last_xact_replay_timestamp()
pg_last_xlog_receive_location()
pg_last_xlog_replay_location()
pg_xlog_location_diff()
Monitoring. Auxiliary functions.

On standby:
pg_current_xlog_location() → pg_last_xlog_receive_location()
WAL amount:
● SELECT pg_xlog_location_diff(pg_current_xlog_location, '0/0');
Lag in seconds:
● SELECT now() - pg_last_xact_replay_timestamp();
Monitoring.

Add new standbys or removing existing standbys.
Temporary pause replication.
Maintenance.

● max_wal_senders
● max_replication_slots
● pg_create_physical_replication_slot()
● pg_drop_replication_slot()
Temporary pause replication.
Maintenance.

Temporary pause replication:
● pg_is_xlog_replay_paused()
● pg_xlog_replay_pause()
● pg_xlog_replay_resume()
Maintenance.

Replication lag.
Replication stopping.
Disk and network problems.
100% disk usage.
Recovery conflicts.
Tables and indexes bloat.
pg_xlog/ bloat.
Part IV. Problems.

Symptoms:
● Data between standby and master are differ.
Causes:
● Long queries on standby, much writes on master;
● Hardware issues.
Solutions:
● Application optimizations.
Replication lag.

Network lag:
● full_page_writes = off;
● ssh tunnels with compression.
Storage lag:
● full_page_writes =off;
● filesystem barriers;
● writethrough/writeback;
● RAID BBU learning;
● ionice (only for cfq elevator).
Networking and Storage.

Symptoms:
● Recovery process uses 100% CPU;
● Lag increasing.
Causes:
● Heavy update/delete, too many autovacuums.
Solutions:
● Increasing wal_keep_segments;
● Temporary disabling of full_page_writes;
● Set priorities with ionice and renice.
Replication stopping.

Causes:
● Replication slots and stopped standby → save XLOG segments;
Solutions:
● Remove the slot and use wal_keep_segments.
Dirty hack:
● Filesystem's reserved blocks percentage and tune2fs.
100% disk usage.

Why conflicts occurs:
● Autovacuum;
● XLOG replay.
Solutions:
● hot_standby_feedback = on;
● Increasing max_standby_streaming_delay.
Recovery conflicts.

Causes:
● Long transactions on a standby.
Solutions:
● pgstattuple;
● VACUUM FULL, pgcompacttable, pg_reorg...;
Tables and indexes bloat.

Symptoms:
● Different size pg_xlog/ and amount of XLOG segments.
Solutions:
● Decreasing checkpoint_timeout;
● Decreasing checkpoint_completion_target.
pg_xlog/ bloat on a standby.

What is it?
For what is needed?
How to do it?
Part V. Switchover and Failover.

Switchover and Failover.
Purposes:
● Updates of software, operating system, or hardware.
● Hardware failures.
Prerequisites.

Run chekpoint on master.
Check replication lag.
Shutdown the master.
Remove recovery.conf and restart a standby.
Switchover.

Pros:
● Old master fast reuse;
● No lost transactions.
Cons:
● Warm cache after restart;
● pg_prewarm extension (since 9.4).
Switchover.

Create trigger file
● recovery.conf: trigger_file = '…'
● Need restart after recovery.conf changes.
With pg_ctl:
● pg_ctl -D ... promote
Failover.

Pros:
● It's fast;
● Don't need a restart;
● Don't need a cache warm.
Cons:
● Lost transactions risk;
● Old master should be reinitialized (until 9.5).
Failover.

Switchover:
● create recovery.conf and start.
Failover:
● reinit as standby (until 9.5);
● pg_rewind (since 9.5).
● timeline must be differs between master and standby.
● old master shut be shutdowned correctly.
● but sometimes issues occurs.
● pg_rewind --target-pgdata=9.5/main --source-server="host=10.0.0.1"
Old master reuse.

$ vi 9.5/replica/postgresql.conf – edit configuration before restart.
port = 5432
$ mv 9.5/replica/recovery.conf /tmp/ – remove recovery.conf from DATADIR
$ psql
> CHECKPOINT; – reduce restart time.
$ pg_ctl -D 9.5/data -m fast stop – shutdown the master.
$ pg_ctl -D 9.5/replica -m fast restart – promote new master.
$ tail -f 9.5/replica/pg_log/postgresql-Wed.log
$ ps auxf
Switchover. Practice.

$ vi 9.5/data/postgresql.conf – edit config.
port = 5433
$ mv /tmp/recovery.conf 9.5/data/ – create recovery.conf.
$ pg_ctl -D 9.5/data start – start.
$ ps auxf – check.
Switchover. Reuse old master.

$ vi 9.5/data/postgresql.conf
$ vi 9.5/replica/postgresql.conf
● wal_log_hints = on – this options required for pg_rewind
● wal_keep_segments = 32
$ pg_ctl -D 9.5/data -m fast restart
$ pg_ctl -D 9.5/replica -m fast restart
Failover. Prepare.

$ pg_ctl -D 9.5/replica -m immediate stop – «crash» a master.
$ pg_ctl -D 9.5/data promote – promote a standby.
$ psql -p 5433 – «doing the changes».
create database test;
Failover.

$ pg_ctl -D 9.5/replica start
$ pg_ctl -D 9.5/replica stop
$ pg_rewind -D 9.5/replica --source-server="host=127.0.0.1 port=5433"
$ vi 9.5/replica/postgresql.conf
port = 5432
$ mv 9.5/replica/recovery.done 9.5/replica/recovery.conf
$ vi 9.5/replica/recovery.conf
port = 5433
$ pg_ctl -D 9.5/replica start
$ ps auxf
Failover. Reuse the old master.

Thanks.
Alexey Lesovsky, PostgreSQL Consulting.
lesovsky@pgco.me
Questions.

Streaming replication in practice

Recommended

More Related Content

What's hot (20)

Similar to Streaming replication in practice (20)

More from Alexey Lesovsky (20)

Recently uploaded (20)

Streaming replication in practice