krb5 1.13 Replication enhancements

[Richard Basch]

I'll send a revised patch later today. I introduced a problem into kproplog... it affects the first patch.


Richard Basch

(Sent via my Samsung GALAXY S™4)

-------- Original message --------
From: Richard Basch <basch at alum.mit.edu> 
Date: 10/02/2014  20:24  (GMT-05:00) 
To: basch at alum.mit.edu,Greg Hudson <ghudson at mit.edu>,Tom Yu <tlyu at mit.edu>,krbdev at mit.edu 
Cc: kayla.c.harrison at gmail.com 
Subject: RE: krb5 1.13 Replication enhancements 
 
Revised commits (in particular, the second one only differs in the syslog output generated, not in the core functionality). I was concerned about excessive logging by kadmind’s servicing lots of slaves, so instead of indicating each slave being notified (which was useful during debugging), merely indicate the number of slaves being notified (being notified does not mean they “understand” the notification or will check-in).
 
https://github.com/rbasch/krb5/commit/ce238c4a19e117caedc7ce53f4780270b86d521c      first_sno fix
https://github.com/rbasch/krb5/commit/1bf5b212392ff2679250fcb3d3f197fee3ced3e1      notify slaves
 
I have also documented an alternate approach to implementing the feature in my Wiki, but it is far too involved for me to pursue. MIT should decide if they want to keep the separation of programs as it stands today, in which case this patch is probably among the better implementations possible.
 
 
From: Richard Basch [mailto:basch at alum.mit.edu] 
Sent: Wednesday, October 1, 2014 9:22 PM
To: 'Greg Hudson'; 'tlyu at mit.edu'; 'krbdev at mit.edu'
Cc: 'kayla.c.harrison at gmail.com'; 'Richard Basch'
Subject: krb5 1.13 Replication enhancements
 
I have uploaded my commits to github (I am still testing, but I am reasonably in my git commit)
See https://github.com/rbasch/krb5/wiki/1.13-Replication-enhancements
 
Basically, this uplifts the 1.13-beta1 code base to the same level as I previously provided as an enhancement for 1.12, specifically:
- Do not require the DB first_sno to be present in the ulog (this avoids extraneous db resyncs to downstream slaves)
- Implement downstream slave notification when ulog is updated.
 
The following are my GIT commits, but it may be better to reference the commits via the wiki in case I have to apply another fix and rebase.
 
Allow kdb_first_* not to be present in the ulog
https://github.com/rbasch/krb5/commit/ce238c4a19e117caedc7ce53f4780270b86d521c     
 
Notify downstream slaves of pending ulog updates
https://github.com/rbasch/krb5/commit/db0c450f6c09f4623bff3c3aff16794ba4e2c763     
 
 
-----Original Message-----
From: Richard Basch [mailto:basch at alum.mit.edu]
Sent: Tuesday, September 30, 2014 3:26 PM
To: 'Greg Hudson'; 'tlyu at mit.edu'
Cc: 'kayla.c.harrison at gmail.com'
Subject: RE: krb5-1.13-beta1 iprop
 
Yeah, re #2... I originally went through all the code which I could find which pertained to the ulog parsing... the only other change I had was in kproplog at the time so that it wouldn't overrun a missing entry. I had added some defensive checks in my original code to only allow for 1 missing entry, but upon a ulog wrap, it would reset the number.
 
I'll try to re-incorporate all of this...
 
 
-----Original Message-----
From: Greg Hudson [mailto:ghudson at mit.edu]
Sent: Tuesday, September 30, 2014 2:58 PM
To: Richard Basch; tlyu at mit.edu
Cc: kayla.c.harrison at gmail.com
Subject: Re: krb5-1.13-beta1 iprop
 
On 09/30/2014 01:30 PM, Richard Basch wrote:
> If you want, I believe I can have patches for 1.13 beta in the next
> couple days and publish them to github. In essence, there should be no
> changes to command-line arguments at this point, since you have iprop
> tree-based propagation already available.
>
> Will that help?
 
Yes, that will provide a better starting point.
 
> Preliminary testing also suggests that it may be causing an extraneous
> full resync (I had this problem in one of my early implementations
> too). I suspect you are resetting the first_sno when you get your
> first ulog entry (my 1.12 patches avoided resetting the first_sno
> until it wrapped so that the first "dump/restore" would not be
> followed by a second one if there were no additional ulog entries to
> apply).  I am not certain if this is what is going on, but some
> preliminary testing suggested such (again, I haven't looked at the
> 1.13
code yet).
 
See item #2 in
http://k5wiki.kerberos.org/wiki/Projects/Hierarchical_iprop#Related_problems
 
Your implementation had a workaround for this issue, but it broke (at
least) the detection of ulog size changes.  I ran out of time to design a better workaround.