Ed Wilts wrote:

>On Mon, Jul 12, 2004 at 05:34:16PM -0400, Craig Steinberger wrote:
>  
>
>>Using the stream numbers for the bpend process will not work, since
>>you can't predict which streams will end first. Say you have 5
>>streams, and stream #2 takes the longest. How do you know which one
>>should restart the database?
>>
>>Typically what is done is have the scripts keep a counter of some
>>sort, and when each stream ends it decreases the counter. You have to
>>be careful of race conditions on the counter file (if that's how you
>>implement it), and of automatic reruns in case of errors.
>>    
>>
>
>You also have to be careful of what happens if your last stream fails
>so you don't leave your database in hot backup mode.  In the rerun case,
>you could put your database in hot backup mode and never take it out
>because there is only 1 stream.  Alternatively, we've cases where the
>database never went into hot backup mode in the first place.  Add to all
>that the complexity of making sure you really do have all your archive
>logs and you begin to see how hard the problem just might be.
>
>Frankly, if you're trying to do your own version of a database agent,
>give up.  The aggravation simply isn't worth it.  No matter how hard you
>try, you will have days when you won't be able to recover.  Something
>always seems to happen and it will be the event you haven't anticipated
>(today for us was a master server crash in the middle of a vault eject 
>run).
>
>        .../Ed
>
>  
>
I wrote a system for a customer that demanded certain custom 
functionality due to
poor database architecture, politics and  an over priced storage 
solution.  The software
works but the customer has done little to both maintain it and more 
importantly,
mandate recovery testing.  There are strange nuances caused by other 
complex parts
of NetBackup that can cause unknowing havoc on an approach, such as the 
media
manager and java GUI's.  Veritas has pr oven solutions that can be less 
expensive
in the long run.  I am certainly not an advocate of dishing out the cash 
to Veritas, but
they have made this custom enterprise capability just one step away from 
a viable
solution by limiting the arguments returned by bpbkar and bpcd.  I tried 
to get the
real story during Vision 2004 in Vegas but could not hook up with the 
developer closest
to the issue.  I thought that checkpoints might lead to that last final 
step, but that does
not look hopeful.  If you don't need multi-streaming, things can be 
fine, otherwise, it is
quite difficult to guarantee restores unless you are willing to put the 
cash up for frequent
recovery tests and ongoing tweaks to keep up with MPs and new versions.  
Oh, and don't
forget that you generally need a test bed environment to get this going, 
nothing better than
doing testing on a multi-terabyte production database no matter how many 
wavers are
signed.


Thanks
Peter DrakeUnderkoffler