Local or public NTP servers?

For a relatively large network (thousands of hosts) - what are the arguments for and against running a locally managed (pool of) NTP server(s) (perhaps periodically set via some public NTP server) and having all other hosts on the network use that (pool of) NTP server(s) versus having all hosts simply use public NTP servers directly, say via ntp.pool.org?

Aside from the pros and cons, What is typical best practice today?

The best practice is to run your own pool of NTP servers set to sync from public NTP servers. In the event that your organization was to lose internet access, you would not want your clocks to become skewed. Further, it is rude to set thousands of hosts to public servers when you could (and should) operate a mirror.

Finally, if you have a secure computing requirement, then you should operate your own independent NTP hosts. You would require special hardware for these systems to operate.

EDIT: Since there was discussion of it, here is some hardware:

Any hardware supporting PPS seems to work on a modern ntpd. This includes some GPS units, although this seems to be rare, at least as rare as serial GPS units are these days. There are hardware devices sold explicitly for this function, however, including one product called TSync-PCIe. According to the manufacturer's site:

The TSync-PCIe offers several configurations of a synchronized timecode reader/generator package offering flexibility and easy integration of precise timing into an embedded computing application. Choose from synchronization to IRIG (and other similar timecodes), GPS (internal or external receivers), or Precise Time Protocol (PTP/IEEE-1588v2). - Site Link: http://i564f.6o.to

Even on a small network I use a local NTP service, which itself updates from an external one. One reason is purely historical, dating back to when the only connection to the Internet was via dial-up modems. The other is that if the NTP service is wrong for any reason I would prefer all the machines to still be consistent, which is more likely to be the case if they all update from a single source.

Best practice, setup 2 (or more) NTP hosts at your location, peer them. Have them sync against at least 4 (preferably, up to 8) external servers from 0.pool.ntp.org to 3.pool.ntp.org. If you use more than 4 you should adjust the frequency that they poll the pool members.

Here's an edited version of my ntp.conf:

server 0.us.pool.ntp.org minpoll 8 maxpoll 14
server 1.us.pool.ntp.org minpoll 8 maxpoll 14
server 2.us.pool.ntp.org minpoll 8 maxpoll 14
server 3.us.pool.ntp.org minpoll 8 maxpoll 14

peer ntp2.example.com

driftfile /var/db/drift.ntp
logfile /var/log/ntp.log
logconfig +sysall +syncall

You can omit the minpoll and maxpoll arguments, I add them so I'm a bit lighter on those servers. The values are 2^n seconds, where n is the argument; those values are higher than the defaults (6 & 10) because I already poll 12 different servers between my three NTP hosts.

If you're very concerned with accuracy you might add the following as well:

server tick.usno.navy.mil prefer minpoll 10 maxpoll 16

This will poll the navy's atomic clock. Note the high poll times as they're fairly heavily loaded and have requested people take it easy on their server (actually a 3 node cluster).