Postgresql – How to take pg_dump of a very large postgres database

While using select datname, xact_commit from pg_stat_database; as suggested by @Jack Douglas doesn't quite work (apparently due to autovacuum), select datname, tup_inserted, tup_updated, tup_deleted from pg_stat_database does appear to work. Both DML and DDL changes will change the values of tup_* columns while a vacuum does not (vacuum analyze on the other hand...).

In the off chance that this may be useful for others, I'm including the backup script that I've put in place. This works for Pg 8.4.x but not for 8.2.x-- YMMV depending on the version of Pg used.

#!/usr/bin/env perl
=head1 Synopsis

pg_backup -- selectively backup a postgresql database cluster

=head1 Description

Perform backups (pg_dump*) of postgresql databases in a cluster on an
as needed basis.

For some database clusters, there may be databases that are:

 a. rarely updated/changed and therefore shouldn't require dumping as 
    often as those databases that are frequently changed/updated.

 b. are large enough that dumping them without need is undesirable.

The global data is always dumped without regard to whether any 
individual databses need backing up or not.

=head1 Usage

pg_backup [OPTION]...

General options:

  -F, --format=c|t|p    output file format for data dumps 
                          (custom, tar, plain text) (default is custom)
  -a, --all             backup (pg_dump) all databases in the cluster 
                          (default is to only pg_dump databases that have
                          changed since the last backup)
  --backup-dir          directory to place backup files in 
                          (default is ./backups)
  -v, --verbose         verbose mode
  --help                show this help, then exit

Connection options:

  -h, --host=HOSTNAME   database server host or socket directory
  -p, --port=PORT       database server port number
  -U, --username=NAME   connect as specified database user
  -d, --database=NAME   connect to database name for global data

=head1 Notes

This utility has been developed against PostgreSQL version 8.4.x. Older 
versions of PostgreSQL may not work.

`vacuum` does not appear to trigger a backup unless there is actually 
something to vacuum whereas `vacuum analyze` appears to always trigger a 
backup.

=head1 Copyright and License

Copyright (C) 2011 by Gregory Siems

This library is free software; you can redistribute it and/or modify it 
under the same terms as PostgreSQL itself, either PostgreSQL version 
8.4 or, at your option, any later version of PostgreSQL you may have 
available.

=cut

use strict;
use warnings;
use Getopt::Long;
use Data::Dumper;
use POSIX qw(strftime);

my %opts = get_options();

my $connect_options = '';
$connect_options .= "--$_=$opts{$_} " for (qw(username host port));

my $shared_dump_args = ($opts{verbose})
    ? $connect_options . ' --verbose '
    : $connect_options;

my $backup_prefix = (exists $opts{host} && $opts{host} ne 'localhost')
    ? $opts{backup_dir} . '/' . $opts{host} . '-'
    : $opts{backup_dir} . '/';

do_main();


########################################################################
sub do_main {
    backup_globals();

    my $last_stats_file = $backup_prefix . 'last_stats';

    # get the previous pg_stat_database data
    my %last_stats;
    if ( -f $last_stats_file) {
        %last_stats = parse_stats (split "\n", slurp_file ($last_stats_file));
    }

    # get the current pg_stat_database data
    my $cmd = 'psql ' . $connect_options;
    $cmd .= " $opts{database} " if (exists $opts{database});
    $cmd .= "-Atc \"
        select date_trunc('minute', now()), datid, datname, 
            xact_commit, tup_inserted, tup_updated, tup_deleted 
        from pg_stat_database 
        where datname not in ('template0','template1','postgres'); \"";
    $cmd =~ s/\ns+/ /g;
    my @stats = `$cmd`;
    my %curr_stats = parse_stats (@stats);

    # do a backup if needed
    foreach my $datname (sort keys %curr_stats) {
        my $needs_backup = 0;
        if ($opts{all}) {
            $needs_backup = 1;
        }
        elsif ( ! exists $last_stats{$datname} ) {
            $needs_backup = 1;
            warn "no last stats for $datname\n" if ($opts{debug});
        }
        else {
            for (qw (tup_inserted tup_updated tup_deleted)) {
                if ($last_stats{$datname}{$_} != $curr_stats{$datname}{$_}) {
                    $needs_backup = 1;
                    warn "$_ stats do not match for $datname\n" if ($opts{debug});
                }
            }
        }
        if ($needs_backup) {
            backup_db ($datname);
        }
        else {
            chitchat ("Database \"$datname\" does not currently require backing up.");
        }
    }

    # update the pg_stat_database data
    open my $fh, '>', $last_stats_file || die "Could not open $last_stats_file for output. !$\n";
    print $fh @stats;
    close $fh;
}

sub parse_stats {
    my @in = @_;
    my %stats;
    chomp @in;
    foreach my $line (@in) {
        my @ary = split /\|/, $line;
        my $datname = $ary[2];
        next unless ($datname);
        foreach my $key (qw(tmsp datid datname xact_commit tup_inserted tup_updated tup_deleted)) {
            my $val = shift @ary;
            $stats{$datname}{$key} = $val;
        }
    }
    return %stats;
}

sub backup_globals {
    chitchat ("Backing up the global data.");

    my $backup_file = $backup_prefix . 'globals-only.backup.gz';
    my $cmd = 'pg_dumpall --globals-only ' . $shared_dump_args;
    $cmd .= " --database=$opts{database} " if (exists $opts{database});

    do_dump ($backup_file, "$cmd | gzip");
}

sub backup_db {
    my $database = shift;
    chitchat ("Backing up database \"$database\".");

    my $backup_file = $backup_prefix . $database . '-schema-only.backup.gz';
    do_dump ($backup_file, "pg_dump --schema-only --create --format=plain $shared_dump_args $database | gzip");

    $backup_file = $backup_prefix . $database . '.backup';
    do_dump ($backup_file, "pg_dump --format=". $opts{format} . " $shared_dump_args $database");
}

sub do_dump {
    my ($backup_file, $cmd) = @_;

    my $temp_file = $backup_file . '.new';
    warn "Command is: $cmd > $temp_file" if ($opts{debug});

    chitchat (`$cmd > $temp_file`);
    if ( -f $temp_file ) {
        chitchat (`mv $temp_file $backup_file`);
    }
}

sub chitchat {
    my @ary = @_;
    return unless (@ary);
    chomp @ary;
    my $first   = shift @ary;
    my $now     = strftime "%Y%m%d-%H:%M:%S", localtime;
    print +(join "\n                  ", "$now $first", @ary), "\n";
}

sub get_options {
    Getopt::Long::Configure('bundling');

    my %opts = ();
    GetOptions(
        "a"             => \$opts{all},
        "all"           => \$opts{all},
        "p=s"           => \$opts{port},
        "port=s"        => \$opts{port},
        "U=s"           => \$opts{username},
        "username=s"    => \$opts{username},
        "h=s"           => \$opts{host},
        "host=s"        => \$opts{host},
        "F=s"           => \$opts{format},
        "format=s"      => \$opts{format},
        "d=s"           => \$opts{database},
        "database=s"    => \$opts{database},
        "backup-dir=s"  => \$opts{backup_dir},
        "help"          => \$opts{help},
        "v"             => \$opts{verbose},
        "verbose"       => \$opts{verbose},
        "debug"         => \$opts{debug},
        );

    # Does the user need help?
    if ($opts{help}) {
        show_help();
    }

    $opts{host}         ||= $ENV{PGHOSTADDR} || $ENV{PGHOST}     || 'localhost';
    $opts{port}         ||= $ENV{PGPORT}     || '5432';
    $opts{host}         ||= $ENV{PGHOST}     || 'localhost';
    $opts{username}     ||= $ENV{PGUSER}     || $ENV{USER}       || 'postgres';
    $opts{database}     ||= $ENV{PGDATABASE} || $opts{username};
    $opts{backup_dir}   ||= './backups';

    my %formats = (
        c       => 'custom',
        custom  => 'custom',
        t       => 'tar',
        tar     => 'tar',
        p       => 'plain',
        plain   => 'plain',
    );
    $opts{format} = (defined $opts{format})
        ? $formats{$opts{format}} || 'custom'
        : 'custom';

    warn Dumper \%opts if ($opts{debug});
    return %opts;
}

sub show_help {
    print `perldoc -F $0`;
    exit;
}

sub slurp_file { local (*ARGV, $/); @ARGV = shift; <> }

__END__

Update: the script has been put on github here.

To begin with, storing ibdata on a separate volume from the datadir is a bad idea because of

• MetaData Issues : I wrote about this as well
  • Why does InnoDB store all databases in one file? (See WARNING)
  • How to Recover an InnoDB table whose files were moved around
  • Error 'Unknown table engine 'InnoDB'' on query. after restarting mysql
• Performance Issues : I wrote about that back in August 2011
  • MySQL InnoDB data file directly on a disk slice (fixed-size)?

Getting back to your problem, the reason LVM for ibdata1 is not good has to do with the metadata in ibdata1 at any given point-in-time. If you were to LVM Snapshot ibdata1, then create new tables, and drop some tables, somne of the corresponding .frm files within datadir would not match. The time needed to coalesce that large an ibdata would be astronomical (unless 21 hours is acceptable to you). I tried using rsync as opposed to LVM and I fared no better.

The safest and most sensible way to clear the space with innodb_file_per_table disabled is to do the following on the Slave

• Step 01) mysqldump all databases
• Step 02) shutdown mysql
• Step 03) delete ibdata1
• Step 04) startup mysql
• Step 05) reload the mysqldump

Without this method, ibdata1 can never be shrunk otherwise. Even in my post in StackOverflow on InnoDB Cleanup (using innodb_file_per_table), you have to do those same 5 steps.