Oracle FAQ | Your Portal to the Oracle Knowledge Grid |
![]() |
![]() |
Home -> Community -> Usenet -> c.d.o.server -> Re: partiioning option not worth it?
It has been brought to my attention that much of what I've said is based on
theory not practice.
On reflection, much of what I argue about partitioning is based on theory.
My starting point however is that last week I saved my employer £14000 at a cost of £300. I also have many years experience in the industry which leads me to believe that the statement "that's not possible" is actually just "I don't know how to do that and I can't be bothered to find out" and "that's how is has to be done" means either "that's how everyone does it" or "I can't be bothered to think beyond what the salesman/rep/consultant said"
Now I could easily be wrong. But frankly, my own career is full of instances of doing things "can't be done". Additionally, I had the (mis)fortune in my first job to work for an organisation where "Can I do it cheaper" mattered.
On my recent posts, (with the exception of one/two posts that were merely abusive and one of which I aplogised for) I cannot find anything unjustifiable.
I will (and have) admitted that there are circumstances where the partitioning option is cost-effective - I just believe that in most circumstances, it is not, if you have an appropriate design. The big advantage of partitioning is that it allows you to architect the client side of a (very) big system as though it were a small(er) system. I believe I have explained the defficiences with the alternative approach I've proposed, and those defficiencies are not dependent on the size of the database, but are dependent on the complexity of the database.
I have made disparaging remarks about the general quality of DBA's but my
experience includes:
a DBA who deleted one of the database files and then couldn't restore it
because "the backup sometimes tries to go over more than one tape".
a DBA who, when I pointed it that having system/manager was poor security on
a test database which had a database link including password to a dba
account which was present on all production databases, asked how I'd found
out the password (it's not present in dba_db_links but is in user_db_links
this implies that you merely need to look at the view definition to see how
to find the password).
a DBA who said "we need to consider scalability" about a database that had
perfectly adequate performance and was only going to shrink as offices
closed - I said "actually, we need to consider shrinkability"
I was once asked to produce a backup/recovery plan and procedure for a
database server. I said that it had to fit in with the disaster recovery
plan for the site and who should I speak to. The respone was "what?" and a
look of incomprehension. When I told the story to a DBA, he proudly said
that they took two tape backups everynight and pointed to the cupboard
adjacent to the servers where the tapes were stored. I did ask him what
would happen if a fire broke out.
a DBA team who all took a coffee break at the same time.
a DBA team who introduced a six hour outage every weekend for backup with
telling the users, the developers, or the application support team
a DBA team who (somehow) managed to lose tables or views when asked to
update a test system with production system data and then managed to have a
batch job in a test system feed into a production system.
a DBA who defined RAID5 as the standard disk configuration who insisted on
database level redo log mirrors (ie to files per group, if I remember the
terminology correctly, which I may not).
Many, many DBAs who do not understand the difference between "that will be
slower" and "that will be too slow to be usable".
a DBA who said "a SAN is better performing than local disk". A statement
which he'd never tested in general never mind wrt the particular project. I
have asked for, and not seen, any evidence that this is true. If the cost
difference was slight, you argue "who cares?" but it is not. There is an
argument for the administrative advanges for a SAN, but certainly the
administrative savings quoted in HP documents I've seen are ludicrous. I'm
more than happy to change this view, but not without evidence.
DBAs who have said we *must* move to a (much more expensive) UNIX
environment because "it is more reliable". Seemingly without understanding
the concept of "yes, but is it worth it", and when they had made no effort
to address the causes of the unrelibiality in the NT environment ie "we
should move to a UNIX environment because I can be arsed to perform a root
cause analysis there". And, by the way, this is the same team that
introduced a six hour outage every weekend.
There is also the issue of motivation - DBAs are often motivated by purely technical issues e.g. "it could be faster if...", "my administrative effort is less if..." rather than "will it be fast enough" or "my extra 1 day's effort will save 2 months of developer effort". This is just human nature and is not the subject of blame, just understanding.
Am I wrong, therefore to be suspicious of DBAs in general? (I won't even mention what some sysadmins have done)
Other questions I've asked (without actually requesting any extra work, Howard!) are "has anyone actually tested the impact of changing redo log size - have I missed anything") - on that, Howard J. Rogers was most helpful in explaining something that I'd missed and I have in mind a test scenario. The question goes to the core of "don't believe what anyone tells you - test it yourself" and "never trust that what you tested yesterday is still true today"
I've asserted that it's cheaper to retain data than to archive it - frankly the cost of tuning a larger system is nothing compared to the cost of building a bullet-proof archive and restore system. I've not heard any costing to change that opinion.
I have stated that an inability to recover a database in noarchivelog mode because the backups hadn't actually worked was the result of incompetence/unprofessionalism wrt the backup/restore procedure, not the cost of archivelog storage. While it may be that the problem would not have happened if logs were available online, there would still have been a worse problem if the entire machine had been lost. This reflects my belief that an untested backup is the same as no backup, and the backup/restore procedure you tested last month will probably not work this month. Is this a misguided belief?
I have been wrong about the implications of checkpoint timeout, but I think had a reasonable point about the difference between loss of data by volume and loss of data by human memory span i.e. it may be more important to limit data loss by elapsed time rather than quantity of data changed.
So my question, am I misguided or not?
Have a laugh
Keith Received on Fri Mar 01 2002 - 16:49:07 CST
![]() |
![]() |