Generating them isn't a huge cost, and it's just a one-time cost. And if you use a string rather than an array to store it, it's much more efficient, both in time and size.

benchmark for [[:alpha:]] on Windows:

__END__
C:\Users\peter.jones\Downloads\TempData\perl>posix_bench.pl
mem before generate arrays: 11,584 K
mem after generate arrays: 25,688 K
mem after delete arrays: 17,600 K
mem before generate strings: 17,600 K
mem after generate strings: 18,264 K
mem after delete strings: 18,264 K
COMPARE GENERATING:
         Rate genArr genStr
genArr 1.27/s     --   -92%
genStr 16.5/s  1198%     --
COMPARE ACCESSING:
            Rate getStr10k getArr10k
getStr10k 7.40/s        --      -98%
getArr10k  417/s     5539%        --
[download]

And practical code for using multiple posix character sets, and generating random strings from those sets:

#!perl

use 5.012; # strict, //
use warnings;
use open ':std', ':encoding(UTF-8)';

my (%posix, %lengths);
for my $class (qw/alpha digit punct/)
{
    $posix{$class} .= $_ for grep { /[[:${class}:]]/ } map {chr} 0 .. 
+0xEFFFF;
    $lengths{$class} = length($posix{$class});
}

sub get_random_char_from_posix
{
    my $class = shift;
    die "no such class" unless exists $posix{$class} and exists $lengt
+hs{$class};
    substr $posix{$class}, rand($lengths{$class}), 1;
}

use Data::Dump;
my $alpha_str; $alpha_str .= get_random_char_from_posix('alpha') for 1
+ .. 10;
my $digit_str; $digit_str .= get_random_char_from_posix('digit') for 1
+ .. 10;
my $punct_str; $punct_str .= get_random_char_from_posix('punct') for 1
+ .. 10;
dd $_ for $alpha_str, $digit_str, $punct_str;

__END__

C:\Users\peter.jones\Downloads\TempData\perl>posix_use.pl
"\x{288EC}\x{1309F}\x{88F8}\x{29B5D}\x{85A3}\x{209E2}\x{9EE1}\x{23015}
+\x{168AB}\x{2A691}"
"\x{B6D}\x{17E5}\x{1D7F6}\x{A627}\x{F21}\x{6F1}\x{1043}\x{A8D7}\x{118E
+4}\x{116C2}"
"\x{A673}\x{FE16}\x{110BC}\x{2CFF}\x{1BFE}\x{1804}\x{11642}\x{1AA0}\x{
+2051}\x{1183B}"
[download]