Re: syncrepl questions

["Kurt D. Zeilenga" <Kurt@OpenLDAP.org>]

At 10:22 PM 9/23/2003, Quanah Gibson-Mount wrote:
>--On Wednesday, September 24, 2003 12:16 AM -0400 Jonghyuk Choi <jongchoi@us.ibm.com> wrote:
>
>>>So, if I have a database with 5 million entries, it is going to send
>>>along  1 change + 4,999,999 ID's of unchanged entries?
>>
>>Yes.
>>The revised (soon too be published) LDAP Sync protocol has two modes.
>>One is state-based (sending add+present) and the other is log-based
>>(sending add+delete).
>>If OpenLDAP becomes capable of maintaining history store
>>then the Syncrepl engine will be enabled to switch between the two modes
>>depending on the consumer's state. If the consumer state is recent enough
>>to be covered by the history store, then only changes would be sent if it
>>is more efficient.
>>If the consumer state cannot be covered by the history store, then the
>>additive changes
>>and IDs of present entries will be sent, which is far more efficient than
>>a full reload.
>>On the other hand, the LCUP proposal only supports the log-based mode.
>>So, without the history store, full reload would be a frequent event.

Actually, its not a "full" reload, but a update+present refresh
that will be common event.  They occurs only during refresh and
then only when something has changed.  So, avoid refresh (use
persist mode) and keep the retry interval short (to reduce the
chance something changes while disconnected).

As Jong noted, history information is needed to use the
updates+delete mode (except when there are no deletes, context
CSNs can detect that).  That should make syncrepl good enough
for *most* directory deployments.   Syncrepl, even with optimal
selection of transfer mode, is likely not good enough for larger
directories and/or frequently updated directories.  For this,
one needs a protocol which transfers attribute level (or even
value-level) changes.  Syncrel transfers entry-level changes.

>Okay, then I understand what is happening with my replica/servers -- It is sending all the blank pieces, which eats up a huge amount of resources and causes my systems to start swapping horribly.  I'll start poking into what in syncrepl is causing the resource issues.  The fact that it always has to send everything in the current methodology, makes, IMHO, syncrepl useless for anyone with a medium or lager database, since the servers are going to be spending all their time sending the changes + empty entries.   For our systems, the dump is happening less quickly than changes are occuring, with means changes would be stacking up, at least if it is left in persistent mode.  Probably the same thing if it is put in periodic updates, unless I have it only do those once every couple of ours, which doesn't work for us, because we need updates to be as close to real time as possible.

Use persist mode.

Kurt