On Tue, Aug 27, 2013 at 01:11:14PM +0200, Jakub Hrozek wrote:
> On Tue, Aug 27, 2013 at 12:34:27PM +0200, Sumit Bose wrote:
> > On Mon, Aug 26, 2013 at 10:16:34PM +0200, Jakub Hrozek wrote:
> > > On Mon, Aug 26, 2013 at 04:02:59PM -0400, Simo Sorce wrote:
> > > > On Mon, 2013-08-26 at 21:41 +0200, Jakub Hrozek wrote:
> > > > > On Mon, Aug 26, 2013 at 03:35:09PM -0400, Simo Sorce wrote:
> > > > > > On Mon, 2013-08-26 at 21:17 +0200, Jakub Hrozek wrote:
> > > > > > > On Mon, Aug 26, 2013 at 06:18:05PM +0200, Sumit Bose
wrote:
> > > > > > > > On Mon, Aug 26, 2013 at 05:20:21PM +0200, Jakub
Hrozek wrote:
> > > > > > > > > On Fri, Aug 23, 2013 at 03:44:09PM +0200,
Sumit Bose wrote:
> > > > > > > > > > Hi,
> > > > > > > > > >
> > > > > > > > > > currently in ipa-server-mode only the
AD groups memberships are
> > > > > > > > > > available. This patch adds the IPA
group memberships to trusted AD
> > > > > > > > > > users.
> > > > > > > > > >
> > > > > > > > > > This patch is missing some unit tests
for some of the helper functions.
> > > > > > > > > > I will send them later, but I
didn't want to delay the next release.
> > > > > > > > > >
> > > > > > > > > > bye,
> > > > > > > > > > Sumit
> > > > > > > > >
> > > > > > > > > I haven't done any testing yet but do we
need the timeout? Since the
> > > > > > > > > initgroups is a rare operation and on logins
we generally want to have
> > > > > > > > > the correct memberships, can we just rely on
responder caching?
> > > > > > > >
> > > > > > > > I was thinking of situations where multiple
logins happen in a short
> > > > > > > > time. Additionally I think even if group
memberships of a user might
> > > > > > > > change often the mapping of AD to IPA group
memberships via the external
> > > > > > > > groups will only change rarely.
> > > > > > > >
> > > > > > > > Maybe we can a cache time option to make it more
flexible?
> > > > > > > >
> > > > > > > > bye,
> > > > > > > > Sumit
> > > > > > >
> > > > > > > I was thinking about this more on my way home and I
think you're
> > > > > > > right we need to optimize the ipa_server_mode. This
could cause the "8AM
> > > > > > > login rush" to be a real bottleneck.
> > > > > > >
> > > > > > > But I think we can exploit the fact that we know the
server well during
> > > > > > > the ipa_server_mode. What about this approach?
> > > > > > > 1. on startup we download all external groups
> > > > > >
> > > > > > it could be a huge set, and would cause a huge load if
someone runs a
> > > > > > puppet script to reconfigure and restart a few 1000
machines with sssd.
> > > > > >
> > > > >
> > > > > A huge set of external groups?
> > > >
> > > > Why not ?
> > >
> > > No technical reason. I just don't see that as a common use-case
that's
> > > all.
> > >
> > > >
> > > > > > > 2. store the largest lastUSN to the server mode
context
> > > > > >
> > > > > > we already have this afaik
> > > > >
> > > > > No we don't. This is new code just for the server mode.
> > > >
> > > > Ahh, that explains some things, sorry.
> > >
> > > No problem, lot of patches are flailing around, I know.
> > >
> > > >
> > > > > >
> > > > > > > 3. on subsequent lookups, only download and store
groups with higher
> > > > > > > lastUSN
> > > > > >
> > > > > > if you change server you are back to square zero though
> > > > > >
> > > > >
> > > > > Which is never during in the server mode. (except for slapd
outages etc)
> > > > >
> > > > > > > 4. perform the lookup always. It's on the
server after all so
> > > > > > > network LDAP search is quite cheap.
> > > > > >
> > > > > > not sure I understand what this means ?
> > > > >
> > > > > The current patch re-downloads external groups every 600 seconds
and if
> > > > > 600 seconds hasn't passed, uses cache. I'm proposing we
download
> > > > > (&(objectClass=externalGroup)(lastUSN>=stored_last_usn))
on every
> > > > > request because in the server mode the latency is not an issue.
> > > >
> > > > 600 is a lot :(
> > >
> > > yes, that's why I was trying to come up with something more
optimized..
> > >
> > > >
> > > > > >
> > > > > >
> > > > > > The point of checking initgroups at login is to assure the
right
> > > > > > memberships are in place, both for security reasons and to
allow a user
> > > > > > to logoff and login back again and make sure eh gets new
group
> > > > > > memberships if he has been granted any new ones.
> > > > > >
> > > > > > Loggin off then back on is the only way to gain system-wide
the new
> > > > > > memberships so that's what an administrator will tell a
user to do if
> > > > > > the user complains he can't access something.
> > > > > >
> > > > > > This means that caching can't last long as the side
effects are severe.
> > > > > > So anything more than a few seconds would probably be bad.
A few seconds
> > > > > > make total sense for load issues when someone is abusing
pam atuh (for
> > > > > > example someon eusing basic auth wired to pam auth for a a
web server
> > > > > > that will receive potentially many tens of authentications
for the same
> > > > > > user within a fraction of a second as each image and file
is loaded in a
> > > > > > new connection).
> > > > > >
> > > > > > In this case you really want to completely cache the whole
operation and
> > > > > > not touch the network for any reason, otherwise latency
will make things
> > > > > > unbearable. But that's the extent to which you want to
go, a few seconds
> > > > > > for auth bursts, nothing more.
> > > > > >
> > > > > > Simo.
> > > > >
> > > > > The latency in the server mode is really small, we should always
be
> > > > > talking to the local server.
> > > >
> > > > I think we shouold defer this until the synrepl control is available
in
> > > > 389ds, then we can simply have a persistent search open using
syncrepl
> > > > and we'll get changes as needed and no more that that (including
> > > > deletions).
> > > >
> > > > Simo.
> > >
> > > Yes, that sounds like a very good solution, I just think we also need to
> > > have a way to handle the external groups in time for 1.11/3.3. If we can
> > > leverage a 389DS control later, even better.
> >
> > What about using a shorter timeout, e.g. 10s, for 1.11.0 and
> > optimizations in 1.11.x?
> >
> > bye,
> > Sumit
>
> That sounds good to me.
Here is a diff I would squash before pushing the patches:
--- a/src/providers/ipa/ipa_subdomains_ext_groups.c
+++ b/src/providers/ipa/ipa_subdomains_ext_groups.c
@@ -606,7 +606,7 @@ static void ipa_get_ext_groups_done(struct tevent_req *subreq)
state->server_mode->ext_groups->ext_groups = ext_group_hash;
/* Do we have to make the update timeout configurable? */
- state->server_mode->ext_groups->next_update = time(NULL) + 600;
+ state->server_mode->ext_groups->next_update = time(NULL) + 10;
ret = ipa_add_ext_groups_step(req);
if (ret == EOK) {
Also I filed
https://fedorahosted.org/sssd/ticket/2062 to track using
either persistent search or syncrepl
_______________________________________________
sssd-devel mailing list
sssd-devel(a)lists.fedorahosted.org
https://lists.fedorahosted.org/mailman/listinfo/sssd-devel