Occasionally I come across complaints that dbms_stats is not obeying the estimate_percent when sampling data and is therefore taking more time than it “should” when gathering stats. The complaint, when I have seen it, always seems to be about the sample size Oracle chose for indexes.
There is a simple but (I believe) undocumented reason for this: because indexes are designed to collate similar data values they are capable of accentuating any skew in the data distribution, which means a sample taken from a small number of leaf blocks can be highly misleading as a guide to the whole index – so Oracle aims for a minimum sample size for gathering index stats.
I’ve found remnants of a note I wrote on comp.databases.oracle.server in December 2004 which claims that this limit (as of Oracle 9.2) was 919 leaf blocks – and I have a faint memory of discovering this figure in an official Metalink (MOS) note. I can’t find the note any more, but it’s easy enough to set up a test to see if the requirement still exists and if the limit is still the same. Here’s a test I ran recently on 126.96.36.199 using an 8KB block size:
This really could be published in the Errata and Addenda of “Cost Based Oracle – Fundamentals”, but it’s more convenient to publish the whole thing here and just add a link to the Errata pages.
In chapter 7, on page 156, I gave an example of the type of SQL that Oracle runs (in the dbms_stats package) to generate a histogram. A sample of the code, and the plan from the 188.8.131.52 tkprof output, is listed below:
As in – how come a unique (or primary key) index is predicted to return more than one row using a unique scan. Here’s and example (running on 10.2.0.3 – but the same type of thing happens on newer versions):
I’m very keen on the 11g extended stats feature, but I’ve just discovered the first of a list of critical weaknesses in the implementation details that could lead to some surprising instability in execution plans [See end of article for links to further problems]. It’s a combination of “column group” statistics and “out of range” predicates. Let’s start with some sample data. (Note: I have run this test on 184.108.40.206 and 220.127.116.11):
A recent comment on a note I wrote some time ago about faking histograms asked about the calculations of selectivity in the latest versions of Oracle. As I read the question, I realised that I had originally supplied a formula for calculating cardinality, rather than selectivity, so I thought I’d supply a proper example.
We’ll start with a script to create some data and stats – and I’m going to start with a script I wrote in Jan 2001 (which is why it happens to use the analyze command rather than dbms_stats.gather_table_stats, even though this example comes from an instance of 18.104.22.168).
A couple of days ago I wrote about some things I’d like to see in index statistics, which means changing dbms_stats.gather_index_stats(). Here’s an idea for dbms_stats.gather_table_stats(). I’d like to see the number of chained and migrated rows counted separately in dba_tables when we run the packaged procedure. The optimizer will use information about chained or migrated rows, but the information is only collected if you use the analyze command (and even then the two figures are summed into the single value chain_cnt) .
Here are a few thoughts on dbms_stats – in particular the procedure gather_index_stats.
The procedure counts the number of used leaf blocks and the number of distinct keys using a count distinct operation, which means you get an expensive aggregation operation when you gather stats on a large index. It would be nice efficiency feature if Oracle changed the code to use the new Approximate NDV mechanism for these counts.
A quick collation – and warning – for 11.2
Bottom line – be careful about what you do with system stats on 11.2
Footnote: the MOS link is a search string producing a list of references. I set it up like that because one of the articles referencing the bug is called “Things to consider before upgrade to 22.214.171.124” and it’s worth reading.
Addendum: one of the people on the two-day course I’ve just run in Berlin sent me a link for a quick note on how to set your own values for the system stats if you hit this bug. It’s actually quite a reasonable thing to do whether or not you hit the bug given the way that gathering the stats can produce unsuitable figures anyway: setting system stats. (I’ve also added their company blog to the links on the right, they have a number interesting items and post fairly regularly.)
Here’s one of those quick answers I give sometimes on forums or newsgroups. I forget where I wrote this, and when, and what the specific question was – but it was something to do with rebuilding an index on a small table where data was constantly being deleted and inserted.
Another problem with high insert/delete rates appears with very small indexes.
If you have a table that is small but constantly recycles its space you may also find you have an index where the number of leaf blocks puts you close to the borderline between having blevel = 1 and blevel = 2. If the size crosses that border occasionally and the statistics are updated to reflect the change – which is quite likely for a table subject to lots of updates and deletes if you have automatic stats collection enabled – then execution plans could change, resulting in dramatic changes in performance.
The workaround is fairly obvious – don’t let Oracle collect stats automatically on that table, instead create a stats-collection strategy for eliminating the change in blevel. For example, keep the stats locked except when you run your own code to deal with the stats, making sure that you overwrite the index blevel with 1 even if it has just crossed the boundary to 2.
Footnote: the reason why a change from 1 to 2 is dramatic is because Oracle ignores the blevel in the optimizer arithmetic when it is set to 1; so the change from 1 to 2 actually has the impact of a change from zero to 2. Then the cost of a nested loop access is “cost of single access multiplied by number of times you do it” – so the sudden appearance of a 2 in the formula gives an increment in cost of “2 * number of times you visit the table” if your small table is the second table in a nested loop join – and suddenly a nested loop becomes much more expensive without a real change in the data size.
Footnote 2: it should be obvious that you don’t need to rebuild the index once you know what the problem is; but since we’re talking about a small index with a blevel that is usually 1 it probably won’t take more than a fraction of a second to rebuild the index and there’s a fair chance you can find a safe moment to do it. In terms of complexity the solution is just as simple as the stats solution – so you might as well consider it. The only thing you need to be careful about is that you don’t happen to rebuild the index at a time when the blevel is likely to be 2.
Footnote 3: For an example of the type of code that will adjust the blevel of an index see this URL. (Note, the example talks about copying stats from one place to another – but the principle is the same.)
I’ve said in the past that one of the best new features, in my view, in 11g was the appearance of proper virtual columns; and I’ve also been very keen on the new “approximate NDV” that makes it viable to collect stats with the “auto_sample_size”.
Who’d have guessed that if you put them both together, then ran a parallel stats collection it would break :(
The bug number Karen quotes (10013177.8) doesn’t (appear to) mention extended stats – but since virtual columns, function-based indexes, and extended stats share a number of implementation details I’d guess that they might be affected as well.
A recent question on the OTN database forum:
Can any one please point to me a document or a way to calculate the average number of rows per block in oralce 10.2.0.3
One answer would be to collect stats and then approximate as block / avg_row_len – although you have to worry about things like row overheads, the row directory, and block overheads before you can be sure you’ve got it right. On top of this, the average might not be too helpful anyway. So here’s another (not necessarily fast) option that gives you more information about the blocks that have any rows in them (I picked the source$ table from a 10g system because source$ is often good for some extreme behaviour).
From time to time I see people asking how they can check how far their system has got in it’s call to collect stats. The question is often a little ambiguous – they know which call to dbms_stats they’ve used, so they know whether they’re trying to monitor stats collection against a single table (taking advantage of v$session_longops springs to mind) or against an entire schema.
Here’s one simple-minded approach that I whipped up a few years ago – it’s possible you could do better with the latest versions of dbms_stats. Its purpose is simply to give you some way of checking what work dbms_stats needs to do (in this example, for a schema), so that you can check how much has been done and how much remains. The basic method can be modified in various ways to match your tastes.
I wrote an article about system statistics / CPU Costing for Oracle magazine a few years ago – and last week I realised that I’ve never supplied a link to it in the notes and comments I’ve made about system statistics. So I’ve just run a search through the Oracle website trying to find it – and discovered that it’s no longer available. Apparently the editors have decided that any technical articles over a certain age should be withdrawn in case they are out of date and misleading. (Clearly they’ve read my blog on trust – I wish the people maintaining Metalink would do the same as the magazine editors – but they probably have a much larger volume to worry about).
However, I have discovered translations of the article in Russian, Korean and Chinese – so if you can read any of these languages, you might want to take a look at them before they disappear too.
If you want an original English version – dated April 2004, which is when I sent it in to Oracle Magazine, and before it underwent some editing – I’ve posted it as a pdf file.
[More on System Statistics]
In an earlier post on frequency histograms I described how Oracle creates an approximate histogram when dealing with character columns, and warned you that the strategy could lead to a couple of anomalies if you were unlucky. I’ve already published a note about one such anomaly that can occur with fairly long character strings, this note describes another anomaly that could appear in less extreme cases. Again, we start by constructing a data set.
In an earlier post on frequency histograms I described how Oracle creates an approximate histogram when dealing with character columns, and warned you that the strategy could lead to a couple of anomalies if you were unlucky. This note describes one such anomaly. We start with a slightly curious data set: