Re: Authentication Methods for LDAP - last call

[Chris Newman <Chris.Newman@INNOSOFT.COM>]

On Sat, 1 Aug 1998, Steve Kille wrote:
> I am totally opposed to mandatory support of CRAM-MD5 in 
> LDAP.   CRAM-MD5 requires a shared secret between client 
> and server.  In a large scale distributed system, where a 
> given client might bind to many servers, this is totally 
> unmagageable.   I think that the policy documents should 
> NOT be requiring this.  I cannot overstate how BAD I think 
> this choice is?

I've been studying the mandatory-to-implement authentication mechanism
problem for over a year (see draft-newman-auth-mandatory-00.txt for a 
problem statement), so I have a strong opinion in this area.  Let me
start by observing the following:

* The IESG requires that if a random update-capable client and a random
read/write capable server are selected, there will be a way to configure
them to authenticate using something better than unencrypted clear text
passwords.  While this is a harsh requirement, it is not unreasonable.
Experience demonstrates that if no mandatory-to-implement mechanism is
defined, then the real-world mandatory-to-implement mechanism is
unencrypted clear text passwords.  This is true of POP3, LDAP and IMAP 
today.

* Scalability comes in two forms -- many users on one server or many
servers with distributed rules

* CRAM-MD5 is several orders of magnitude faster than X.509.

* CRAM-MD5 scales better for many users on one server than X.509

* X.509 scales better for a distributed system than CRAM-MD5

* CRAM-MD5 is a small burden on an implementor, X.509 is a huge burden

CRAM-MD5 is the right choice today for the mandatory-to-implement
mechanism in LDAP.  It has good enough properties and scalability for
single-server deployments.  It is not hard to implement.  There is, of
course, no requirement to use the mandatory-to-implement mechanism or even
to have it on by default as long as it can be turned on. 

I happen to think that TLS-protected clear text passwords make a good
SHOULD implement feature -- as they provide backwards compatibility with
legacy authentication sources such as NTLM or /etc/passwd -- an issue that
Mark Smith pointed out.

Implementations SHOULD NOT use unprotected clear text passwords.  That's a
recommendation most will ignore for practical reasons, but it causes
sufficient harm that the recommendation is important.

I believe the current draft is realistic and pragmatic.  Unless someone
else can make a compelling argument for a different mandatory-to-implement
standards-track mechanism, CRAM-MD5 is the best choice today.

		- Chris