Re: Please review: ticket #48766 Replication changelog can incorrectly skip over updates

[thierry bordaz <tbordaz@xxxxxxxxxx>]

Hi Ludwig,

Thanks for your explanation. The design looks very good. I think it would be good to put into the code (especially clcache_adjust_anchorcsn) the reference to the related design paragraph.

There is something I do not understand in clcache_skip_change.
My understanding is that this is the only place where the consumer_maxcsn is updated.
But there are several conditions, if we decide to skip the update that the consumer_maxcsn is not updated.
One of them is 'rid == buf->buf_consumer_rid'.
Does that mean that the consumer_maxcsn remains unchanged for the RID of the consumer ?

thanks
thierry

On 05/24/2016 09:22 AM, Ludwig Krispenz wrote:

Hi,

On 05/23/2016 06:29 PM, thierry bordaz wrote:

On 05/23/2016 03:06 PM, Ludwig Krispenz wrote:

This is the latest version of the "changelog buffer processing" fixes.


https://fedorahosted.org/389/ticket/48766

https://fedorahosted.org/389/attachment/ticket/48766/0001-reworked-clcach-buffer-code-following-design-at-http.patch

The background for the fix is here, I would like to get feedback on this as well to clarify what is unclear
http://www.port389.org/docs/389ds/design/changelog-processing-in-repl-state-sending-updates.html

Hello Ludwig,

I have not yet reviewed the patch. I was looking at the design.

Regarding your note: http://www.port389.org/docs/389ds/design/changelog-processing-in-repl-state-sending-updates.html#special-case-rid-of-the-consumer-in-the-current-replication-session.

If you refer to this part:

Special case: RID of the consumer in the current replication session

If the consumer in the replication session is also a master its RID will be contained at least in the consumerRUV. If it is also in the supplier RUV the question is if it should be considered in the decision if updates should be sent. Normally a master has the latest changes applied to itself, so there would be no need to check and send updates for its RID. But there can be scenarios where this is not the case: if the consumer has been restored from an older backup the latest csn for its own RID might be older than changes available on other servers.

NOTE: The current implementation ignores anchorCSNs based on the consumer RID. If, by chance, the anchor csn used is older than this csn, the changes will be sent, but they also ca nbe lost.

this referres to the "current" implementation before the fix, the doc started as a post-design doc, and it shoul dbe correctedd
with the fix the if the supplier has newer changes for the consumerRID than the consumer it will be reflected in the anchor csn calculation.

It is said that the anchorCSN will not be the from the consumerRID. What is the mechanism that guaranty that the consumer will receive all the updates it was the originator ?

thanks
thierry
--
389-devel mailing list
389-devel@xxxxxxxxxxxxxxxxxxxxxxx
http://lists.fedoraproject.org/admin/lists/389-devel@xxxxxxxxxxxxxxxxxxxxxxx

-- 
Red Hat GmbH, http://www.de.redhat.com/, Registered seat: Grasbrunn, 
Commercial register: Amtsgericht Muenchen, HRB 153243,
Managing Directors: Paul Argiry, Charles Cachera, Michael Cunningham, Michael O'Neill

--
389-devel mailing list
389-devel@xxxxxxxxxxxxxxxxxxxxxxx
http://lists.fedoraproject.org/admin/lists/389-devel@xxxxxxxxxxxxxxxxxxxxxxx

--
389-devel mailing list
389-devel@xxxxxxxxxxxxxxxxxxxxxxx
http://lists.fedoraproject.org/admin/lists/389-devel@xxxxxxxxxxxxxxxxxxxxxxx