On 02/27/2014 12:52 PM, Jay Pipes wrote:
On Thu, 2014-02-27 at 16:13 +0000, Henry Nash wrote:
So a couple of things about this:
1) Today (and also true for Grizzly and Havana), the user can chose
what LDAP attribute should be returned as the user or group ID. So it
is NOT a safe assumption today (ignoring any support for
domain-specific LDAP support) that the format of a user or group ID is
a 32 char UUID. Quite often, I would think, that email address would
be chosen by a cloud provider as the LDAP id field, by default we use
the CN. Since we really don't want to ever change the user or group
ID we have given out from keystone for a particular entity, this means
we need to update nova (or anything else) that has made a 32 char
assumption.
I don't believe this is correct. Keystone is the service that deals with
authentication. As such, Keystone should be the one and only one service
that should have any need whatsoever to need to understand a non-UUID
value for a user ID. The only value that should ever be communicated
*from* Keystone should be the UUID value of the user.
We never said it was a UUID, its only been that in the "allocated by
Keystone" cases thus far.
userids are blobs....well, character blocks, and are not expected to be
interprested by the end services, but also not confined to be 32
characters long.
Most LDAP use cases do not allow us to write data back the Directory.
Federation has the same use cases. That means that we are going to end
up with a slew of ephemeral data in Keystone for no real value.
If the Keystone service uses LDAP or federation for alternative
authentication schemes, then Keystone should have a mapping table that
translates those elongated and non-UUID identifiers values (email
addresses, LDAP CNs, etc) into the UUID value that is then communicated
to all other OpenStack services.
The short of it is that we need to define a max userid length and stick
to it. Thus far, we've had 64 characters in Keystone, and I think we
can stick with that. Nova needs to expand its view to 64 chars.
The rest is Keystone internals to deal with.
Best,
-jay
2) In oder to support the ability for service providers to be able to
have the identity part of keystone be satisfied by a customer LDAP
(i.e. for a given domain, have a specific LDAP), then, as has been
stated, we need to subsequently, when handed an API call with just a
user or group ID, be able to "route" this call to the correct LDAP.
Trying to keep true to the openstack design principles, we had
planned to encode a domain identifier into the user or group ID - i.e.
distribute the data to where it is needed, in ortherwords, the user
and group ID provide all the info we need to route the call to the
right place. Two implementations come to mind:
2a) Simply concatenate the user/group ID with the domain_id, plus some
separator and make a composite public facing ID. e.g.
"user_entity_id@@UUID_of_domain". This would have a technical maximum
size of 64+2+64 (i.e. 130), although in reality since we control
domain_id and we know it is always 32 char UUID - in fact the max size
would be 98. This has the problem of increasing the size of the
public facing field beyond the existing 64. This is what we had
planned for IceHouse - and is currently in review.
2b) Use a similar concatenation idea as 2a), but limit the total size
to the existing 64. Since we control domain_id, we could (internally
and not visibly to the outside world), create a domain_index, that was
used in place of domain_id in the publicly visible field, to minimize
the number of chars it requires. So the public facing composite ID
might be something like <up to 54 chars of entity_id>@@<8 chars of
domain_index>. There is a chance, of course, that the 54 char
restriction might be problematic for LDAP users....but I doubt it. We
would make that a restriction and if it really became a problem, we
could consider a field size increase at a later release
3) The alternative to 2a and 2b is to have, as had been suggested, an
internal mapping table that maps external facing entity_ids to a
domain plus local entity ID. The problem with this idea is that:
- This could become a very big table (you will essentially have an
entry for every user in every corporate LDAP that has accessed a given
openstack)
- Since most LDAPs are RO, we will never see deletes...so we won't
know when (without some kind of garbage collection) to cull entries
- It obviously does not solve 1) - since existing LDAP support can
break the 32 char limit - and so it isn't true that this mapping table
causes all public facing entity IDs to be simple 32 char UUIDs
From a delivery into IceHouse point of view any of the above are
possible, since the actual mapping used is relatively small part of
the patch. I personally favor 2b), since it is simple, has "less
moving parts" and does not change any external facing requirements for
storage of user and group IDs (above and beyond what is true today).
Henry
On 27 Feb 2014, at 03:46, Adam Young <ayo...@redhat.com> wrote:
On 02/26/2014 08:25 AM, Dolph Mathews wrote:
On Tue, Feb 25, 2014 at 2:38 PM, Jay Pipes <jaypi...@gmail.com>
wrote:
On Tue, 2014-02-25 at 11:47 -0800, Morgan Fainberg wrote:
> For purposes of supporting multiple backends for
Identity (multiple
> LDAP, mix of LDAP and SQL, federation, etc) Keystone is
planning to
> increase the maximum size of the USER_ID field from an
upper limit of
> 64 to an upper limit of 255. This change would not
impact any
> currently assigned USER_IDs (they would remain in the
old simple UUID
> format), however, new USER_IDs would be increased to
include the IDP
> identifier (e.g. USER_ID@@IDP_IDENTIFIER).
-1
I think a better solution would be to have a simple
translation table
only in Keystone that would store this longer identifier
(for folks
using federation and/or LDAP) along with the Keystone user
UUID that is
used in foreign key relations and other mapping tables
through Keystone
and other projects.
Morgan and I talked this suggestion through last night and agreed
it's probably the best approach, and has the benefit of zero
impact on other services, which is something we're obviously
trying to avoid. I imagine it could be as simple as a user_id to
domain_id lookup table. All we really care about is "given a
globally unique user ID, which identity backend is the user from?"
On the downside, it would likely become bloated with unused
ephemeral user IDs, so we'll need enough metadata about the
mapping to implement a purging behavior down the line.
UUIDs are 32 chars long. Its really just uuid@@uuid that pushes us
over the 64 character limit.
If we can shorten up the IDP_ID we can fit everything in 64 chars
(which means only Nova needs to expand its column size)
What if we enumerated IDPs by index, from 10000000 to 99999999 or
something comparable, and then use the new domain_index (or prot
domain id to not be a uuid). Then the above scheme would work and
no migration would be required.
The only identifiers that would ever be communicated to
any non-Keystone
OpenStack endpoint would be the UUID user and tenant IDs.
> There is the obvious concern that projects are utilizing
(and storing)
> the user_id in a field that cannot accommodate the
increased upper
> limit. Before this change is merged in, it is important
for the
> Keystone team to understand if there are any places that
would be
> overflowed by the increased size.
I would go so far as to say the user_id and tenant_id
fields should be
*reduced* in size to a fixed 16-char BINARY or 32-char
CHAR field for
performance reasons. Lengthening commonly-used and
frequently-joined
identifier fields is not a good option, IMO.
Best,
-jay
> The review that would implement this change in size
> is https://review.openstack.org/#/c/74214 and is
actively being worked
> on/reviewed.
>
>
> I have already spoken with the Nova team, and a single
instance has
> been identified that would require a migration (that
will have a fix
> proposed for the I3 timeline).
>
>
> If there are any other known locations that would have
issues with an
> increased USER_ID size, or any concerns with this change
to USER_ID
> format, please respond so that the issues/concerns can
be addressed.
> Again, the plan is not to change current USER_IDs but
that new ones
> could be up to 255 characters in length.
>
>
> Cheers,
> Morgan Fainberg
> —
> Morgan Fainberg
> Principal Software Engineer
> Core Developer, Keystone
> m...@metacloud.com
>
>
> _______________________________________________
> OpenStack-dev mailing list
> OpenStack-dev@lists.openstack.org
>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
