This is not a straightforward task. A couple of things I can think of:
- Are the primary keys numeric? And what's the expected normalized form?
- What will happen when there is a collision with the sub-hash keys during collation?
- Database updates would be trivial once the first two are resolved.

I have written a little script below to demonstrate how to do collision in the ideal world, ie., numeric primary keys, no sub-hash key collision.

use strict;
use warnings;
use Data::Dumper;

my $href = {
    '1' => { 'key1' => 'value1' },
    '1.0' => { 'key2' => 'value2' },
    '0001' => { 'key3' => 'value3' },
    '2' => { 'key1' => 'value1' },
    '0002' => { 'key2' => 'value2' },
};

print Dumper($href);

# Build a cross-reference table of primary key remapping
my %key_xref = map { $_ => normalize_key($_) } keys %$href;

# Remap primary keys / merge data
while (my ($from, $to) = each %key_xref) {
    next if $from eq $to;

    foreach (keys %{$href->{$from}}) {
        $href->{$to}{$_} = $href->{$from}{$_};
    }

    delete $href->{$from};
}

print Dumper($href);

# Normalize primary keys
sub normalize_key
{
    my $key = shift;
    return int $key;
}
[download]

And the output -

# before
$VAR1 = {
          '0002' => {
                      'key2' => 'value2'
                    },
          '0001' => {
                      'key3' => 'value3'
                    },
          '1' => {
                   'key1' => 'value1'
                 },
          '1.0' => {
                     'key2' => 'value2'
                   },
          '2' => {
                   'key1' => 'value1'
                 }
        };

# after
$VAR1 = {
          '1' => {
                   'key2' => 'value2',
                   'key1' => 'value1',
                   'key3' => 'value3'
                 },
          '2' => {
                   'key2' => 'value2',
                   'key1' => 'value1'
                 }
        };
[download]