OSDC 2012 | Taking hot backups with XtraBackup by Alexey Kopytov

Taking hot backups with
XtraBackup
Alexey.Kopytov@percona.com
Principal Software Engineer
April 2012

Taking hot backups with XtraBackup
Supported storage engines
● InnoDB/XtraDB
– hot backup
● MyISAM,Archive, CSV
– with read lock
● Your favorite exotic engine
– may work if supports FLUSH TABLES WITH READ LOCK

Supported platforms
● Linux
– RedHat 5; RedHat 6
– CentOS, Oracle Linux
– Debian 6
– Ubuntu LTS
● Windows
– experimental releases
● Solaris, Mac OS X
– binaries available soon
● FreeBSD
– build from source code

Backup idea
● backup logic is identical to InnoDB recovery
procedure
● 2nd
stage reuses code from InnoDB recovery

Distribution structure
● xtrabackup
– Percona Server 5.1 with XtraDB
– MySQL 5.1 with InnoDB plugin
● xtrabackup_51
– MySQL 5.0
– Percona Server 5.0
– MySQL 5.1 with built-in InnoDB
● xtrabackup_55
– MySQL 5.5
– Percona Server 5.5
● innobackupex
– Perl script
– wrapper around xtrabackup* binaries
● xbstream

FLUSH TABLES WITH READ LOCK
● set the global read lock - after this step,
insert/update/delete/replace/alter statements cannot run
● close open tables - this step will block until all statements
started previously have stopped
● set a lag to block commits

Why? Consider:
● Copy table1.frm
● Copy table2.frm
● Copy table3.frm
● Copy table4.frm
– .......... ALTER TABLE table1 starts.......
● Copy table5.frm
● Copy table6.frm

With FTWRL:
● FLUSH TABLES WITH READ LOCK
● Copy table1.frm
● Copy table2.frm
● Copy table3.frm
● Copy table4.frm
– .......... ALTER TABLE table1 -- LOCKED.......
● Copy table5.frm
● Copy table6.frm
● UNLOCK TABLES
● .......... ALTER TABLE table1 starts.............

● same problem with MyISAM:
– non-transactional storage engine
– no REDO logs

Basic usage
(taking a backup)
innobackupex [options…] /data/backup
Options:
● --defaults-file=/path/to/my.cnf
– datadir (/var/lib/mysql by default)
● --user
● --password
● --host
● --socket

Streaming backups
● Local backups:
– local ilesystem
– NFS mounts
● Streaming backups:
– innobackupex | gzip | ssh ...

Streaming backups
Basic command:
innobackupex --stream=tar /tmpdir

Streaming backups
To extract:
innobackupex --stream=tar /tmpdir |
tar -xvif - -C /data/backup
-i is important!
● “ignore blocks of zeros in archive (normally mean
EOF)”

Streaming backups
● Usage:
– compression:
● innobackupex --stream=tar . | gzip - >
/data/backup/backup.tar.gz
– encryption:
● innobackupex --stream=tar . | openssl des3 -salt
-k “password” > backup.tar.des3
– remote backup:
● innobackupex --stream=tar . | ssh user@host “tar
-xif - -C /data/backup”
– compressed + encrypted + remote backup:
● innobackupex --stream=tar . | gzip - | openssl
des3 -salt -k “password” | ssh @user@host “cat -
> /data/backup.tar.gz.des3”

Incremental backups
● handles incremental changes to InnoDB
● does NOT handle MyISAM or other engines
– makes a full copy instead

Incremental backups
Basic usage:
$ innobackupex --incremental
--incremental-basedir=/previous/full/or/incremental/
/data/backup/inc
LSN of the full backup is read from
xtrabackup_checkpoints:
backup_type = full-backuped
from_lsn = 0
to_lsn = 1597945
last_lsn = 1597945

Incremental backups
Merging full + incremental:
● innobackupex --apply-log
--redo-only /data/backup/full
● innobackupex --applog-log
--incremental-dir=/data/backup/inc

Restoring individual tables:
export
● problem: restore individual InnoDB table(s) from a full backup to another server
● use --export to prepare
● use improved table import feature in Percona Server to restore
● innodb_file_per_table=1
Full backupFull backup
ibdataibdata
actor.ibdactor.ibd
customer.ibdcustomer.ibd
film.ibdfilm.ibd
export
Server B
ibdataibdata
actor.ibdactor.ibd
film.ibdfilm.ibd
Server A
ibdataibdata
actor.ibdactor.ibd
film.ibdfilm.ibd
backup

export
Why not just copy the .ibd ile?
● metadata:
– InnoDB data dictionary (space ID, index IDs, pointers to
root index pages)
– .ibd page ields (space ID, LSNs, transaction IDs, index
ID, etc.)
● xtrabackup --export dumps index metadata to
.exp iles on prepare
● Percona Server uses .exp iles to update both data
dictionary and .ibd on import

export
$ xtrabackup --prepare --export --innodb-file-per-table=1
--target-dir=/data/backup
...
xtrabackup: export metadata of table 'sakila/customer' to
file `./sakila/customer.exp` (4 indexes)
xtrabackup: name=PRIMARY, id.low=23, page=3
xtrabackup: name=idx_fk_store_id, id.low=24, page=4
xtrabackup: name=idx_fk_address_id, id.low=25, page=5
xtrabackup: name=idx_last_name, id.low=26, page=6
...

import
● improved import only available in Percona Server
● can be either the same or a different server instance:
– (on different server to create .frm)
CREATE TABLE customer(...);
– SET FOREIGN_KEY_CHECKS=0;
– ALTER TABLE customer DISCARD TABLESPACE;
– <copy customer.ibd to the database directory>
– SET GLOBAL innodb_import_table_from_xtrabackup=1;
(Percona Server 5.5)
or
SET GLOBAL innodb_expand_import=1;
(Percona Server 5.1)
– ALTER TABLE customer IMPORT TABLESPACE;
– SET FOREIGN_KEY_CHECKS=1;

import
● Improved table import is only available in Percona
Server
● tables can only be imported to the same server
with MySQL (with limitations):
– there must be no DROP/CREATE/TRUNCATE/ALTER
between taking backup and importing the table
mysql> ALTER TABLE customer DISCARD TABLESPACE;
<copy customer.ibd to the database directory>
mysql> ALTER TABLE customer IMPORT TABLESPACE;

Partial backups
● backup individual tables/schemas rather than the
entire dataset
● InnoDB tables:
– require innodb_file_per_table=1
– restored in the same way as individual tables from a full
backup
– same limitations with the standard MySQL server (same
server, no DDL)
– no limitations with Percona Server when
innodb_import_table_from_xtrabackup is
enabled

Partial backups:
selecting what to backup
innobackupex:
● streaming backups:
● --databases=”database1[.table1] ...”,
e.g.: --databases=”employees sales.orders”
● local backups:
● --tables-file=filename, ile contains database.table, one per line
● --include=regexp,
e.g.: --include='^database(1|2).reports.*'
xtrabackup:
● --tables-file=filename (same syntax as with innobackupex)
● --tables=regexp (equivalent to --include in innobackupex)

Partial backups:
preparing
$ xtrabackup --prepare --export --target-dir=./
...
120407 18:04:57 InnoDB: Error: table 'sakila/store'
InnoDB: in InnoDB data dictionary has tablespace id 24,
InnoDB: but tablespace with that id or name does not exist. It will be
removed from data dictionary.
...
xtrabackup: export option is specified.
xtrabackup: export metadata of table 'sakila/customer' to file
`./sakila/customer.exp` (4 indexes)
xtrabackup: name=PRIMARY, id.low=62, page=3
xtrabackup: name=idx_fk_store_id, id.low=63, page=4
xtrabackup: name=idx_fk_address_id, id.low=64, page=5
xtrabackup: name=idx_last_name, id.low=65, page=6
...

Partial backups:
restoring
● Non-InnoDB tables
– just copy iles to the database directory
● InnoDB (MySQL):
– ALTER TABLE ... DISCARD/IMPORT TABLESPACE
– same limitations on import (must be same server, no
ALTER/DROP/TRUNCATE after backup)
● XtraDB (Percona Server):
– xtrabackup --export on prepare
– innodb_import_table_from_xtrabackup=1;
– ALTER TABLE ... DISCARD/IMPORT TABLESPACE
– no limitations

Minimizing footprint
● I/O throttling
● ilesystem cache optimizations
● parallel ile copying

Minimizing footprint: I/O throttling
--throttle=N
Limit the number of I/O operations per second in 1 MB units
xtrabackup --throttle=1 ...
readread writewrite readread writewrite
Second 1
Second 2
Second 3
readread writewrite waitwait
Second 1
Second 2
Second 3

Minimizing footprint:
FS cache optimizations
● XtraBackup on Linux:
– posix_fadvise(POSIX_FADV_DONTNEED)
– hints the kernel the application will not need the
speciied bytes again
– works automatically, no option to enable
● Didn't really work in XtraBackup 1.6, ixed in 2.0

Parallel ile copying
●
creates N threads, each thread copying one ile at a time
● utilizes disk hardware by copying multiple iles in parallel
– best for SSDs
– less seeks on HDDs due to more merged reqs by I/O scheduler
– YMMV, benchmarking before using is recommended
Data directoryData directory
ibdata1ibdata1
actor.ibdactor.ibd
film.ibdfilm.ibd
Backup locationBackup location--parallel=4
ibdata1
film.ibd
customer.ibd
actor.ibd

New features
● XtraBackup 2.0
– streaming incremental backups
– parallel compression
– xbstream
– LRU dump backups

Streaming incremental backups
Problem: send an incremental backup to a remote host:
innobackupex --stream=tar
--incremental ... | ssh ...
● didn't work in XtraBackup 1.6
– innobackupex used external utilities to generate TAR
streams, didn't invoke xtrabackup binary
– xtrabackup binary must be used for incremental
backups to scan data iles and generate deltas
● in XtraBackup 2.0:
– xtrabackup binary can produce TAR or XBSTREAM
streams on its own

Compression
● disk space is often a problem
● compression with external utilities has some
serious limitations
● built-in parallel compression in XtraBackup 2.0

Compression:
external utilities
● local backups:
– create a local uncompressed backup irst, then gzip iles
– must have suficient disk space on the same machine
– data is read and written twice
● streaming backups:
– innobackupex --stream=tar ./ | gzip - >
/data/backup.tar.gz
– innobackupex --stream=tar ./ | gzip - |
ssh user@host “cat - > /data/backup.tar.gz”
– gzip is single-threaded
– pigz (parallel gzip) can do parallel compression, but decompression is
still single-threaded
– have to uncompress the entire .tar.gz even to restore a single table

Compression:
XtraBackup 2.0
● new --compress option in both innobackupex and xtrabackup
● QuickLZ compression algorithm: http://www.quicklz.com/
– “the world's fastest compression library, reaching 308 Mbyte/s per core”
– combines excellent speed with decent compression
(8x in tests)
– more algorithms (gzip, bzip2) will be added later
● qpress archive format (the native QuickLZ ile format)
● each data ile becomes a single-threaded .qp archive
– no need to uncompress entire backup to restore a single table as with .tar.gz

Compression:
XtraBackup 2.0
● parallel! --compress-threads=N
● can be used together with parallel ile copying:
xtrabackup --backup --parallel=4
--compress --compress-threads=8
Data directoryData directory
ibdata1ibdata1
actor.ibdactor.ibd
film.ibdfilm.ibd
I/O threads Compression
threads
1
2
3
4
1
8
7
6
5
4
3
2

LRU dump backup
(XtraBackup 2.0)
LRU dumps in Percona Server:
page1page1 page2page2 page3page3 ......
InnoDB buffer pool
most recently used least recently used
ib_lru_dump
id id id id
●
Reduced warmup time by restoring buffer pool state from ib_lru_dump after restart

LRU dump backup
(XtraBackup 2.0)
● XtraBackup 2.0 discovers ib_lru_dump and backs it
up automatically
– buffer pool is in the warm state after restoring
from a backup!
– make sure to enable buffer pool restore in my.cnf
after restoring on a different server
●
innodb_auto_lru_dump=1 (PS 5.1)
●
innodb_buffer_pool_restore_at_startup=1 (PS 5.5)

Resources, further reading & feedback
● XtraBackup documentation:
http://www.percona.com/doc/percona-xtrabackup/
● Downloads:
http://www.percona.com/software/percona-xtrabackup/downloads/
● Google Group:
http://groups.google.com/group/percona-discussion
● #percona IRC channel on Freenode
● Launchpad project:
https://launchpad.net/percona-xtrabackup
●
Bug reports:
https://bugs.launchpad.net/percona-xtrabackup

We are hiring!
http://www.percona.com/about-us/careers

Percona Live: New York 2012
Oct 1 & 2, 2012
Call for papers open!

Questions?
alexey.kopytov@percona.com

OSDC 2012 | Taking hot backups with XtraBackup by Alexey Kopytov

More Related Content

What's hot (19)

Similar to OSDC 2012 | Taking hot backups with XtraBackup by Alexey Kopytov (20)

Recently uploaded (20)

OSDC 2012 | Taking hot backups with XtraBackup by Alexey Kopytov