On Wed, Jul 27, 2016 at 01:03:59PM +0200, Jakub Hrozek wrote:
On Wed, Jul 27, 2016 at 12:22:46PM +0200, Lukas Slebodnik wrote:
> On (27/07/16 12:08), Jakub Hrozek wrote:
> >On Wed, Jul 27, 2016 at 12:02:24PM +0200, Jakub Hrozek wrote:
> >> On Wed, Jul 27, 2016 at 11:54:16AM +0200, Lukas Slebodnik wrote:
> >> > ehlo,
> >> >
> >> > attached patch fixes acces denied after activating user in 389ds.
> >> > Jakub had some comments/ideas in ticket but I think it's better to
discuss
> >> > about virtual attributes and timestamp cache on mailing list.
> >>
> >> Yes, so the comment I have is that while this works, it might break some
> >> strange LDAP servers.
> >>
> >> We use modifyTimestamp as a 'positive' indicator that the entry has
not
> >> changed -- if the modifyTimestamp didn't change, we consider the
cached
> >> entry the same as what is on the server and only bump the timestamp
> >> cache. If the timestamp is different, we do a deep-comparison of cached
> >> attribute values with what is on the LDAP server and write the sysdb
> >> cache entry only if the attributes differ.
> >>
> >> I was wondering if we can use the modifyTimestamp at all, then, because
> >> even if it's the same, we might want to check the attributes to see if
> >> some of the values are different because some of the attributes might be
> >> this operational/virtual attribute..
> >
> >Sorry, sent too soon.
> >
> >I think the questions are -- 1) can we enumerate the virtual attributes?
> That might be a question for 389-ds developers.
> But it's very likely it will be different on other LDAP servers.
>
> >2) Would different LDAP servers have different virtual attributes.
> >
> >For 2) maybe a possible solution might be to set a non-existing
> >modifyTimestamp attribute value, but I would consider that only a
> >kludge, we shouldn't break existing setups..
>
> I am not satisfied with this POC solution either.
>
> So should we remove usage of modifyTimestamp for detecting changes?
I would prefer to ask the DS developers before removing it completely.
At least for large groups it might take a long time to compare all attribute
values and IIRC we don't depend on any virtual attributes for groups. Maybe
we could parametrize that part of the code and enable the fast way with
modifyTimestamps for 'known' server types, that is for setups with AD and
IPA providers.
For users, there is typically not as many attributes so we should be
fine deep-comparing all attributes.
I'm adding Thierry (so please reply-to-all to keep him in the thread).
Thierry, in the latest sssd version we tried to add a performance
improvement related to how we store SSSD entries in the cache. The short
version is that we store the modifyTimestamp attribute in the cache and
when we fetch an entry, we compare the entry modifyTimestamp with what
is on the server. When the two are the same, we say that the entry did
not change and don't update the cache.
This works fine for most attributes, but not for attributes like
nsAccountLock which do not change modifyTimestamp when they are
modified. So when an entry was already cached but then nsAccountLock
changed, we treated the entry as the same and never read the new
nsAccountLock value.
To fix this, I think we have several options:
1) special-case the nsAccountLock. This seems a bit dangerous,
because I'm not sure we can say that some other attribute we are
interested in behaves the same as nsAccountLock.
2) drop the modifyTimestamp optimization completely. Then we fall
back to comparing the attribute values, which might work, but for
huge objects like groups with thousands of members, this might be
too expensive.
3) only use the modifyTimestamp optimization for cases where we know
we don't read any virtual attributes.
And my question is -- can we, in general, know if the modifyTimestamp
way of detecting changes is realiable for all LDAP servers? Or do you
think it should only be used for cases where we know we are not
interested in any virtual attributes (that would mostly be storing
groups from servers where we know exactly what is on the server side,
like IPA or AD).