Unlocking the Mysteries Behind Update Statistics

1
Unlocking the Mysteries Behind Update Statistics

• John F. Miller III

2
The Dice Problem

• Throw dice, how many will be 1?

3
Questions about the Dice

• How many dice are you throwing?
• How many sides does each dice have?
• Are all the dice the same?

The better the information, the more accurate the
estimate.
4
What does Update Statistics do?

• Collects information for the optimizer
• Statistics LOW
• Distributions MEDIUM
  HIGH
• Drop Distributions
• Compile stored procedures

5
Statistics Collected

• systables
• systables
• syscolumns
• syscolumns
• sysindexes
• sysindexes

• Number of Rows
• Number of pages to store the data
• Second largest value for a column
• Second smallest value for a column
• of unique values for the lead key
• How highly clustered the values for the lead key

6
Update Statistics LowBasic Algorithm

• Walk the leaf pages in each index
• Submit btree cleaner requests when deleted items
  are found causing re-balancing of indexes
• Collects the following information
• Number of unique items
• Number of leave pages
• How clustered the data is
• Second highest and lowest value

7
How to Read Distributions
of rows represented in this bin
--- DISTRIBUTION --- (
-1 1 ( 868317, 70,
75) 2 ( 868317, 24,
100) 3 ( 868317, 12, 116) 4
( 868317, 30, 147) 5 (
868317, 39, 194) 6 ( 868317,
28, 222) --- OVERFLOW
--- 1 ( 779848, 43) 2
( 462364, 45)
of unique values
Highest Value in this bin

• To get the range of values look at the highest
  value in the previous bin.

The value
of rows for this value
8
Example - Approximating a Value
--- DISTRIBUTION --- (
-1 1 ( 868317, 70,
75) 2 ( 868317, 24,
100) 3 ( 868317, 12, 116) 4
( 868317, 30, 147) 5 (
868317, 39, 194) 6 ( 868317,
28, 222) --- OVERFLOW
--- 1 ( 779848, 43) 2
( 462364, 45)

• There are 868317 rows containing a value between
  -1 and 75
• There are 70 unique values in this range
• The optimizer will deduce 868317 / 70 12,404
  records for each value between -1 and 75

9
Example - Dealing with Data Skew
--- DISTRIBUTION --- (
-1 1 ( 868317, 70,
75) 2 ( 868317, 24,
100) 3 ( 868317, 12, 116) 4
( 868317, 30, 147) 5 (
868317, 39, 194) 6 ( 868317,
28, 222) --- OVERFLOW
--- 1 ( 779848, 43) 2
( 462364, 45)

• Data skew
• For the value 43 how many records will the
  optimizer estimate will exist?
• Answer 779848 values
• Any value that exceeds 25 of the bin size will
  be placed in an overflow bin

10
Basic Algorithmfor Medium and High

• Develop scan plan based on available resources
• Scan table
• High All rows
• Medium Sample of rows
• Sort each column

• Build distributions
• Begin transaction
• Delete old columns distributions
• Insert new columns distributions
• Commit transaction

11
Scan

• The table is scanned in its entirety for update
  stats high, while it is only sampled for update
  stats medium (see Sample Size)
• The reading of rows is done in dirty read
  isolation, regardless of what the user has set
  for their transaction level.

12
Scan

• This scan of the table may occur several times
  depending on the amount of sort memory available
  and the number of columns to collect statistics
  about.
• The approximate number of table scans is defined
  by the (size of the data to sort) / (amount of
  sort memory)

13
Sort

• The rows processed by the scan phase are passed
  directly to the sort package.
• Each column in the row for which statistics are
  being generated is passed to a unique invocation
  of a sort.

14
Build

• After the sort is completed we read the sorted
  column data finding out the number of duplicates
  and unique values creating approximately 200
  range bins by default.
• Any count of a duplicates value that exceeds 25
  the size of a bin will be placed in an overflow
  bin.

15
Insert

• Now we have to delete the old distributions and
  insert the new distributions. As long as the user
  was not in a transaction this will be done as its
  own transaction. This transaction will last for
  less than 1 second and will hold NO locks on the
  tables, but locks on the system catalogs while
  the update occurs.

16
Sample Size

• HIGH
• The entire tables is scanned and all rows are
  used.
• Medium
• Misconception about the number of rows sampled is
  based on the number of rows in the table, this is
  incorrect.
• The number of samples depends on the Confidence
  and Resolution. See the following chart.

17
Update Statistics Medium Sample Size
18
Update Statistics Medium Memory Requirements
19
Update Statistics High Memory Requirements

• In memory sort
• Approximate Memory number of rows sum(column
  widths 2 sizeof(pointer) )

20
Memory Rules

• Estimated Update Stats memory is below 100MB
• Hard coded limit of 4MB
• Attempts to minimize the scans by fitting as many
  columns into 4MB
• Estimated Update Stats memory is above 100MB
• Memory is requested from MGM
• Attempt to minimize the scans by fitting as many
  columns in the MGM memory

21
Examples

• Customer Table
• Cust_id integer
• Fname char(50)
• Lname char(50)
• Address1 char(200)
• Address2 char(200)
• State char(2)
• zipcode integer
• Number of Rows 500,000

22
ExamplesMemory for Incore Sort
23
ExamplesNumber of Table Scans
24
Confidence

• A factor in the number of samples used by update
  statistics medium

25
Resolution

• Percentage of data that is represented in a
  distribution bin
• Example
• 100,000 rows in the table
• Resolution of 2
• Each bin will represent 2,000 rows

26
Improvements in update statistics in 7.31.UD2

• UPDATE STATISTICS CAN NOT ALLOCATE SORT MEMORY
  BETWEEN 4 AND 100 MB
• The default has been raised from 4MB to 15MB
• User can now configure the amount of memory
• UPDATE STATISTICS USES SLOW SCANNING TECHNOLOGY
  WHEN SCANNING A TABLE -- ENABLE LIGHT SCANS
• Implemented light scans
• Set oriented reads

27
Improvements in update statistics in 7.31.UD2

• THE PLAN UPDATE STATISTICS MAKE WHEN SCANNING IS
  NOT VIEWABLE BY THE DBA
• Set explain will now print the scan path and
  resource usage
• UPDATE STATISTICS LOW ON FRAGMENT INDEXES RUNS
  SERIALLY AND VERY SLOW
• With PDQ turned on each index fragment will be
  scanned in parallel
• PDQ at 1 means 10 of the index fragments scanned
  in parallel, while PDQ at 10 means all the index
  fragments will be scanned in parallel

28
Improvements in update statistics in 7.31.UD2

• ERROR 126 WHEN EXECUTING UPDATE STATISTICS AND
  STORED PROCECURE (ALSO ERRORS 312/100)
• Errors when trying to insert the distributions
  because set lock mode to wait was not handled
  properly inside update statistics
• SCANNING AN INDEX WHICH IS FRAGMENT IS SLOW DUE
  TO THE INEFFICIENT MERGE IN THE FRAGMENT MANAGER
• Binary search used instead of the previous nest
  loop merge when ordering index fragments
• Most noticeable when the number of fragments in
  an index is large

29
Tuning with the New Statistics

• Turn on PDQ when running update statistics, but
  only for tables
• Avoid PDQ when updating statistics for procedures
• When running high or medium increase the memory
  update statistics has to work with
• Enable parallel sorting (i.e. PSORT_NPROCS)

30
Considerations

• Change the RESOLUTION to 1.5
• Increasing the number of bins for the
  distributions
• Increasing the sample size for update statistics
  medium

31
Example

• Following Example
• Table size 215,000 rows
• Row size 445 bytes
• Uniprocessor

32
Example of the current update statistics

• Table jmiller.t9
• Mode HIGH
• Number of Bins 267 Bin size 1082
• Sort data 101.4 MB
• Sort memory granted 4.0 MB
• Estimated number of table scans 10
• PASS 1 c9
• PASS 2 c5
• PASS 3 c7
• PASS 4 c6
• ..
• PASS 10 c4

• Completed pass 1 in 0 minutes 24 seconds
• Completed pass 2 in 0 minutes 20 seconds
• Completed pass 3 in 0 minutes 17 seconds
• Completed pass 4 in 0 minutes 17 seconds
• Completed pass 5 in 0 minutes 17 seconds
• Completed pass 6 in 0 minutes 15 seconds
• Completed pass 7 in 0 minutes 14 seconds
• Completed pass 8 in 0 minutes 15 seconds
• Completed pass 9 in 0 minutes 16 seconds
• Completed pass 10 in 0 minutes 14 seconds

Total Time 146 seconds
33
The new Defaults in 7.31.UD2
Table jmiller.t9 Mode
HIGH Number of Bins 267 Bin size 1082 Sort
data 101.4 MB Sort memory granted
15.0 MB Estimated number of table scans
7 PASS 1 c9,c8,c10,c5,c7 PASS 2 c6,c1 PASS 3
c3 PASS 4 c2 PASS 5 c4

• Completed pass 1 in 0 minutes 34 seconds
• Completed pass 2 in 0 minutes 19 seconds
• Completed pass 3 in 0 minutes 16 seconds
• Completed pass 4 in 0 minutes 14 seconds
• Completed pass 5 in 0 minutes 15 seconds

Total Time 98 seconds
New Memory Default
34
Enabling PDQ with update statistics

• Table jmiller.t9
• Mode HIGH
• Number of Bins 267 Bin size 1082
• Sort data 101.4 MB
• PDQ memory granted 106.5 MB
• Estimated number of table scans 1
• PASS 1 c1,c2,c3,c4,c5,c6,c7,c8,c9,c10
• Index scans disabled
• Light scans enabled
• Completed pass 1 in 0 minutes 29 seconds

PDQ Memory
Features Enabled
Total Time 29 seconds