[Date Prev][Date Next] [Chronological] [Thread] [Top]

Re: resynchronizing partial replica after ACL change

Ferenc Wagner wrote:

Howard Chu <hyc@symas.com> writes:


Ferenc Wagner wrote:


I'm running a couple of partial replicas, where the replicated content
is narrowed by provider ACLs.  All is fine.  Now, I need to extend the
replicas with a new attribute, so I extend the ACL to allow reading of
this attribute as well.  How can I get the replicas synchronize this new
attribute without downtime?  Is forcing a full reload by slapd -c rid=1
a good solution?


That should work, for a standard provider/consumer replication.


Pretty standard, I guess.  But for the future: where's the catch?  For what
setups would this fail to work?
In MMR, updates whose entryCSN match the entry's existing entryCSN will be ignored, so a resync attempt like this won't change anything. 


Also, this slapd restart mean at least some downtime.  Is there a way to
trigger such resync online, without slapd being stopped at all?  If not,
is there a fundemental problem stopping the implementation of such a
feature?
Submit an enhancement request to the ITS, find someone interested in coding it. 


Does that incur the slapd restart downtime only, or would the replica
database be in a crippled state until the full resync finishes?


No crippled state. Just inconsistent, since some entries will have the
new attributes and others won't have updated yet.


That's fine, the new attributes are not used on the replicas yet.  Going
further: how do I know that the resync finished, thus the above
inconsistencies are already washed out?
Poll cn=monitor, watch the runqueue information, wait for the syncrepl thread to finish. 


In general: are matching contextCSNs of a given suffix on the provider
and the consumer equivalent to the replication being idle?


In general, yes.


(In my case
the ACL change does not modify the contextCSN of the provider, so this
does not matter, but I'd like to devise a sound and efficient way of
monitoring the replication.  Also, some measure of the replication delay
would be useful for trending, does slapd collect such info?)


No. Patches welcome. But first, define exactly "delay" relative to what?

--
  -- Howard Chu
  CTO, Symas Corp.           http://www.symas.com
  Director, Highland Sun     http://highlandsun.com/hyc/
  Chief Architect, OpenLDAP  http://www.openldap.org/project/