Re: multimaster replication and index corruption

Rich Megginson <rmeggins@xxxxxxxxxx> · Tue, 10 Nov 2015 12:16:18 -0700

On 11/10/2015 12:12 PM, ghiureai wrote:

Rich, thank you for all support  for last day , unfortunately there is 
a strong wave in developers team:" the multimaster replication is 
creating issues with UI"  ( I do not  totally agree since  can not  be 
reproduce+ full describe the issues).
Is been decided to moved down to master slave,  please I need to know 
if I still need to  exclude member of plugin from replication in this 
case ?

Adding 389-users to discussion.

I would say yes, you should still exclude memberOf from replication.  In 
any case, it is better to reduce the amount of replication traffic and 
replication processing, and let the slave calculate the memberOf values.

As far as the original issue - if we can't get enough information to 
diagnose/reproduce the problem, then we can be of little help.

Thanks a lot
Isabella
 On 11/10/2015 09:23 AM, Rich Megginson wrote:
On 11/10/2015 10:14 AM, Adrian Damian wrote:
Rich,

Thanks for your help. Let me jump in with more details.

We've seen index corruption on a number of occasions. It seems to
affect searchable attributes for which there are indexes. Queries on
an attribute in LDAP that used to work suddenly stopped working. They
would return incomplete results and no results at all, although the
data on the server was the same. The fix on those situations was to
drop the index corresponding to the attribute and re-create it.
So in this case, you have some sort of LDAP search client, and you are
doing a search for '(indexed_attribute=known_value)' and you are not
seeing a result, and this is what you mean by "index corruption"?

Are you aware of the dbscan tool?
https://access.redhat.com/documentation/en-US/Red_Hat_Directory_Server/10/html/Configuration_Command_and_File_Reference/dbscan.html 

This tool allows you to examine the index file in the database directly.

dbscan -f
/var/lib/dirsrv/slapd-instance_name/db/userRoot/indexed_attribute.db4 -k
known_value

This will allow you to look at the indexed_attribute index directly for
the value "known_value".

We've run the db fix script that the LDAP distribution comes with
What db fix script?  Do you have a link to it, or a link to the product
documentation for the script?

and there are no reports of corruption when this problem occurs. That
makes it very hard to detect. We don't know what else to look for when
we run into this again and more importantly, we don't know what
triggers it and how to prevent it.

Mind you we are currently doing active development changing both the
software clients that access the LDAP servers as well as the
configurations of the servers. It is possible to had been written to
both masters in the master replication configuration when the problem
occurred but because there were multiple clients concurrently
accessing the servers it is hard to figure out what triggered the 
issue.

Adrian

On 11/09/2015 05:06 PM, Rich Megginson wrote:
On 11/09/2015 05:47 PM, Ghiurea, Isabella wrote:
Hi Rich,
Thank you for your feedback , as always greatly appreciate when
comes from  389-DS RH support.
We are  not using vm just plain hardware, here is the description
I  got from developers team related to the issues they are seeing
when running   integration tests with multimaster replication :
"index corruption: put content, run tests: OK, do more stuff (reads,
writes, etc), ru tests: FAIL, notice "missing attributes", rebuild
index(ices), run tests: OK. "
What does this mean?  What program is printing these index corruption
messages?  Is it some tool provided by Red Hat?

Unfortunately, I understood  this cases/issue can not be reproduce
on regular basis,  no mode details can be provide at this time

All reads and writes are going to  only the master replication DS,
not slave .
     I totally agree with your this is the way to cfg and maintain
Directory  Server in a operation critical  env: multmaster
replication only one master for writes.
    Here is the DS version:
    rpm -qa | grep 389-ds
389-ds-console-doc-1.2.6-1.el6.noarch
389-ds-base-libs-1.2.11.15-34.el6_5.x86_64
389-ds-1.2.2-1.el6.noarch
389-ds-base-1.2.11.15-34.el6_5.x86_64
This is quite an old version of 389-ds-base.  I suggest upgrading to
RHEL 6.7 with latest patches.

389-ds-console-1.2.6-1.el6.noarch

Thank you
Isabella

FWD:

We have cfg multimaster replication /fractional replication memberof
plugging excluded , we are seeing from time to time index corruption
with some indexes , there is a strong feeling from developers this
are related to DS multimaster replication internal settings.
What version of 389?  rpm -q 389-ds-base
I'm assuming you are not using IPA.
What does "index corruption" mean?  What exactly do you see?

Are you running in virtual machines? If so, what kind? vmware? kvm?
Are you using virtual disks or dedicated physical devices/paravirt?

We are writing to only one DS , same server at all time but reading
from all DS 's cfg for mutlmaster.
Are you seeing "index corruption" on the write master or on all
servers?

Are other people seen this kind of issues with multimaster rep cfg ,
should we start avoiding this replication cfg at all ?

This is the recommended way to deploy. If this is not working for
you, either you have a configuration problem, or there is some sort
of vm or hardware problem, or there is a serious bug that requires
fixing ASAP.

We choose the multimaster for the fast and reliable option to switch
between master DS's , moving one step down to master/slave may
require some down time when switching DS's back.
Isabella

Hi Rich,
Thank you for your feedback , as always greatly appreciate when
comes from  389-DS RH support.
We are  not using vm just plain hardware, here is the description
I  got from developers team related to the issues they are seeing
when running  tests with multimaster replication  :index corruption:
put content, run tests: OK, do more stuff (reads, writes, etc), ru
tests: FAIL, notice "missing attributes", rebuild index(ices), run
tests: OK.

I belive we the reads and writes right now are only the master
replication DS , not slave .
I totally agree with your this is the way to cfg and maint DS in a
operation env: multmaster replication with one master for writes.
More comments , imput I appreciate
    rpm -qa | grep 389-ds
389-ds-console-doc-1.2.6-1.el6.noarch
389-ds-base-libs-1.2.11.15-34.el6_5.x86_64
389-ds-1.2.2-1.el6.noarch
389-ds-base-1.2.11.15-34.el6_5.x86_64
389-ds-console-1.2.6-1.el6.noarch
389-dsgw-1.1.11-1.el6.x86_64

________________________________________
From: ghiureai [isabella.ghiurea@xxxxxxxxxxxxxx]
Sent: Monday, November 09, 2015 1:05 PM
To: 389-users@xxxxxxxxxxxxxxxxxxxxxxx
Subject: multimaster replication and index corruption

Hi List,
We have cfg   multimaster replication /fractional replication 
memberof
plugging excluded ,    we are seeing from time to time index 
corruption
with some indexes , there is a  strong feeling from developers 
this are
related to DS  multimaster replication internal settings.
We are writing to only one DS  , same server at all time but reading
from all DS 's cfg for mutlmaster.
Are other  people seen this kind of issues with multimaster rep cfg ,
should we start avoiding this replication  cfg  at all ?
We choose the multimaster for the fast and reliable option to switch
between  master DS's , moving one step down to master/slave may 
require
some down time  when switching DS's back.
Isabella
--
389 users mailing list
389-users@xxxxxxxxxxxxxxxxxxxxxxx
https://admin.fedoraproject.org/mailman/listinfo/389-users

--
389 users mailing list
389-users@xxxxxxxxxxxxxxxxxxxxxxx
https://admin.fedoraproject.org/mailman/listinfo/389-users