jump to navigation

Friday Philosophy – Should the Software or the User be the Stupid One? August 7, 2009

Posted by mwidlake in internals, performance.
Tags: , ,
5 comments

Oracle’s performance engine is complex and copes with a lot of database situations automatically – or to be more precise, it tries to cope with lots of database situations automatically.

Over the last few versions, Oracle has added many, many things to allow the database to cope automatically with all sorts of different data volumes, spreads of data, relationships between tables, use of different oracle technologies (By this I mean bitmap tables, index tables, partitions, clusters, external tables). All of these things aim to allow the database to just “cope” with whatever you need it to do, with less and less effort by the users {by users, I mean technical users; DBAs and Developers}. Thus it allows for “stupid” users. {I mean no offence, maybe read “inexperienced” instead of stupid}.

As an example, you can now have some very large tables consisting of several partitions and some status look-ups. You query against them. Oracle’s CBO will automatically ignore partitions it can ignore, use indexes or full table scans to use the least amount of IO,use histograms to spot where clauses are on low-cardinality values, Hash joins rather then nested loops as appropriate depending on memory availability, use bitmap indexes when it thinks it can and merge the results from several bitmap indexes, use function based indexes to support functions in where clauses….
It even self-gathers the information to look after all this. Column usage and table modifications are tracked, statistics are gathered when needed and in ways to support data skew, PGA and SGA can be automonitoring and managing…

It all sounds great. In fact, most of the time, for most people, it is great. {I know, most people reading this post are probably people who have encountered the problem systems and so know it goes wrong and so you need more knowledge to cope – you are a biased set of people. In the nicest way, I should add :-) } The idea is, I believe, that you do not neet to be smart to look after Oracle.

If it is not great, if this highly complex system gets it wrong and tries to satisfy SQL statements in sub-optimal ways, then the User has to step in and fix things. ie You.

It is now horrifically complex for us technical users to understand what is going on. You have to not only be “not stupid”, but “not average” either. Sometimes you have to be “not great”, ie brilliant.

In my example, we need to look at if the SQL is constructed to allow the indexes to be used, are functions correctly laid out to use function indexes, are partitions being maintained correctly, when were stats last gathered, did it include histograms and do they help, has oracle missed the need for histograms, are the indexes analyzed at a high enough sample size, are the bitmaps greatly slowing down inserts, have hints been used on the code, are initialisation parameters set to override default fucntionality…

You get the idea, I won’t drone on further. I didn’t even mention memory considerations though {OK, I’ll shut up}.

My point is, the more complex the software, the more “intelligent” it is, the more it is designed to allow for “stupid” users, then the more super-intelligent the user has to be to cope when it breaks.

How about an alternative?

How would it be if we went back to the Rule Based Optimizer and no automatic management of complex situations?

Oracle would maybe need to add a few rules to the RBO for it to cope with later developments, so it would be slightly more complex than V6 but not a lot.
Everything else, the User decides. You only gather stats you decide to gather, on objects you decide need them. No you don’t, it’s a Rule Based Optimizer – no stats gathering! {But see below}.

No automatic memory management. No automatic anything.

The User {the technical user, the DBA and Developer} would have to be smart. Not brilliant, just smart. You would probably have to do more, but most of it would be easier as the levels of complexity and interdependence are reduced. All those tweaks and tricks in the CBO and all the monitoring to cope with “complex” would not exist to go wrong.

Plus it might solve another concern I have. I think there is a chasm growing as there is no need to solve simple problems as Oracle copes but then having to solve complex problems when Orcle does not cope. If you don’t develop skills and experience solving the simple problems, how do you solve the complex ones? I think this is why most Oracle performance and architecture experts are old {Sorry, pleasantly middle-aged}. Young people new to the arena have a massive learning mountain to climb.

So, if we have stupid software, maybe we can get away with more stupid “smart” expert users. ie ALL of us can cope. You cut your teeth on smaller, simpler systems and learn how to cope with the stupid software beast. As you learn more, you learn to cope with more complex situations and they never get that complex as the database is not so “clever”

I’d actually still argue that all the intelligence gathering the Oracle database does should still continue – stats gathered on objects, the ability to gather information on memory usage and thus advice on changes, tracking column usages and table changes. But We, the Stupid Users get to look at it and use it as we see fit for our systems.

I’m sure many systems would not work quite so fast in my senario, but I’d rather have a system working at 75% it’s theoretical fastest all the time rather than one working at 95% and breaking regularly, and in ways so complex it needs weeks to work out and fix.

I now await all the comments to tell me how stupid I am {I can be blindlingly stupid, especially on Fridays}.

Follow

Get every new post delivered to your Inbox.

Join 166 other followers