[Freeipa-devel] Suggestion for the A part of IPA

[Martin Kosek]

On 06/02/2015 10:29 AM, Innes, Duncan wrote:
> Just a bit of a head's up and a refresh of this with perhaps some new
> data.
>
>>
>> Good to hear :-) We recently also started investigating the Audit
>> capabilities for (notice I write "for" and not "in") IPA. You can
>> check my initial nudge to the freeipa-users list, which was
>> unfortunately with no reply:
>>
>> https://www.redhat.com/archives/freeipa-users/2015-March/msg00940.html
>>
>
> First up, just got round to reading this Martin.  Not sure how I missed
> it
> when it first came out as it's a strong area of interest for me.

We have more now! I created

https://www.freeipa.org/page/Centralized_Logging

where you can read more about this POC project and even see demo showing what 
the current POC ELK server can do (with link to Docker server container and 
sources of course).

> The main part of what this message is about is a big change I made to
> our
> logging recently.  I added in 4 of our main production IPA servers
> (there
> are 8 in total, but 4 sit beyond firewalls that take more scrutiny for
> changes than I wanted for now).  The 4 I've added, though, serve more
> clients I figure.  The amount of log traffic to the pair of Logstash
> servers has now jumped from around 50k records/hour to around 250k.
>
> Doubtless this still doesn't push any of the parts to their limits, but
> there has been a barely noticeably increase in CPU usage on the 2
> Logstash
> servers.  We've gone from around 2% CPU usage to 4%.
>
> Since the CPU usage on our 'loudest' IPA server rarely peaks above 10%,
> this doesn't present nearly as much load as I had anticipated.  I have
> run
> Logstash parsers on my DEV IPA boxes, but will now investigate running
> them on my Prod servers too.

Cool!

> What I'm getting at is that perhaps clients sending logs back to the IPA
> servers for parsing, then being sent on to a central DB for storage,
> isn't
> going to break the bank performance-wise.
>
> All of the systems in question here are 2vCPU with 4Gb vRAM running on
> ESXi
> hosts, so nothing special in the performance arena.
>
> It strikes me as a reasonably elegant solution to pair the
> authentication
> and log parsing services on the same set of servers.  This would allow
> each
> client to use the same servers/failover etc for SSSD as for rsyslog.
>
> There may, of course, be other considerations, but I'm suggesting that
> system load isn't necessarily one of them.  Much as projects such as
> Katello
> can run with everything on the same server, or split out Postgres and
> the
> like onto separate servers when there are performance considerations.
>
> Thoughts?  I'm not saying they should always be paired, but that if a
> user
> designs a system with enough horse power, this piggy-backing could work
> well.

Ah, interesting idea and measurement. We have not thought about this kind of 
architecture yet. What we did in our POC is to configure FreeIPA clients and 
servers to send the logs directly to the logging server which was on completely 
different machine (container) than rest of the infrastructure.

It may be an alternative scheme, to have FreeIPA server containing the log 
processing and then forwarding further to other REK/ELK server and clients 
simply forwarding the logs to the same server as where they are authenticating. 
If all the log configuration is baked in ipa-{server,replica,client}-install, 
it would be extremely easy to integrate.

I am not sure if the authentication+logging binding would be that easy, nor 
that it belong together that much. SSSD would need to dynamically export the 
address of the FreeIPA server it communicates with, maybe similarly as it does 
with Kerberos (http://linux.die.net/man/8/sssd_krb5_locator_plugin) - but that 
does not seem as a good fit either.

In any case, CCing Jakub for reference.

Thanks,
Martin