Data Warehousing and Mining - CS437 GIK Institute

1. Lecture 16
Dr. Fawad Hussain
Primary and Secondary IndicesPrimary and Secondary IndicesPrimary and Secondary IndicesPrimary and Secondary Indices

2. Use multiple indices for certain types of queries.
Example:
select account-number
from account
where branch-name = “Perryridge” and balance = 1000
Possible strategies for processing query using indices on single
attributes:
1. Use index on branch-name to find accounts with balances of $1000; test branch-name =
“Perryridge”.
2. Use index on balance to find accounts with balances of $1000; test branch-name = “Perryridge”.
3. Use branch-name index to find pointers to all records pertaining to the Perryridge branch.
Similarly use index on balance. Take intersection of both sets of pointers obtained.
Multiple-Key Access

3. With the where clause
where branch-name = “Perryridge” and balance = 1000
the index on the combined search-key will fetch only records that satisfy
both conditions.
Using separate indices in less efficient — we may fetch many records (or
pointers) that satisfy only one of the conditions.
Can also efficiently handle
where branch-name - “Perryridge” and balance < 1000
Indices on Multiple Attributes

4. Sample account File

5. Hash Function of branch-name

6. Bitmap Indices on Relation customer-info

7. Primary Index
Secondary Index
Sparse Index vs Dense Index
IndexingTechniques
Primary versus secondary indexing.
Single index access versus scanning.
Combining multiple indexes.
What we studied

8. PI for a table (inTeradata) is a specification of its partitioning
column(s).
PI may be defined as unique (UPI) or non-unique (NUPI).
Automatic enforcement of uniqueness when UPI is specified.
PI provides an implicit access path to any row just by knowing its
value.
Only one PI per table.
PI can be on multiple columns i.e. composite.
Primary Index

9. Primary index selection criteria:
Common join and retrieval key.
Distributes rows evenly across database partitions.
Less than 10,000 rows per PI value when non-unique.
WHY?
Primary Index

10. Trick question: What should be the primary index of the transaction
table for a large financial services firm?
create table tx
(tx_id decimal (15,0) NOT NULL
,account_it decimal (10,0) NOT NULL
,tx_amt decimal (15,2) NOT NULL
,tx_dt date NOT NULL
,tx_cd char (2) NOT NULL
....
) primary index (???);
Ans: It depends
Primary Index

11. Almost all joins and retrievals will come in through the account
_id foreign key.
Want account_id as NUPI.
If data is “lumpy” when distributed on account_id or if accounts
have very large numbers of transactions (e.g., an institutional
account could easily have 10,000+ transactions).
Want tx_id as UPI for good data distribution.
Primary Index

12. Joins and access via primary index are very efficient due toTeradata’s
sophisticated row hashing algorithms that allow going directly to the data
block containing the desired row.
Single I/O operation for accessing a data row via UPI.
Single I/O operation for accessing a data row via NUPI whenever all rows
with the same PI value fit into a single block.
SingleVAMP operation for indexed retrieval.
No spool space required.
Primary Index

13. Primary index is free!
No storage cost.
No index build required.
This is a direct result of the underlying hash-based file system
implementation.
OLTP databases use a page-based file system and therefore do not deliver
this performance advantage.
Primary Index

14. Any index that is NOT a PI
SI structures are implemented using the same underlying structure
as base tables (often referred to as sub-tables).
SI may be defined as unique (USI) or non-unique (NUSI).
Automatic enforcement of uniqueness when USI is specified.
Up to thirty-two SI’s per table inTeradata.
Unlike a primary index, SI are not “free” in terms of storage.
SI is NOT required BUT desired.
Secondary Index(SI)

15. Lecture 17
Dr. Fawad Hussain
Primary and Secondary IndicesPrimary and Secondary IndicesPrimary and Secondary IndicesPrimary and Secondary Indices----IIIIIIII

16. Primary index selection criteria:
Common join and retrieval key.
Distributes rows evenly across database partitions.
Less than 10,000 rows per PI value when non-unique.
WHY?
Recall - 1

17. Secondary Index (SI)
Any index that is NOT a PI
SI structures are implemented using the same underlying structure as
base tables (often referred to as sub-tables).
SI may be defined as unique (USI) or non-unique (NUSI).
Automatic enforcement of uniqueness when USI is specified.
Up to thirty-two SI’s per table inTeradata.
Unlike a primary index, SI are not “free” in terms of storage.
SI is NOT required BUT desired.
Recall - 2

18. A non-unique secondary index (NUSI) is partitioned so that each index
entry is co-located on the sameVamp (VirtualAccess Module Processor)
with its corresponding row in the base table.
Each row access via a NUSI is a singleVamp operation because the NUSI
entry and data row are co-located.
NUSI access is always performed in parallel across allVamp whenever it
is appropriate to do so.
Secondary Index (NUSI)

19. Compressed ROWID index structure:
Hash on index value to get block location (ROWID for sub-table).
Store index value just once followed by all ROWIDs in base table corresponding to the
index value.
Sorted by ROWID to facilitate maximum efficiency when accessing base table,
performing updates and deletes, etc.
Additional blocks allocated when NUSI is non-selective and compressed ROWID
structure for the index value exceeds 64K.
Secondary Index (NUSI)

20. Secondary Index (NUSI)
Non Unique Secondary
Index Value
Non Unique Secondary
Index Value
Hashing AlgorithmHashing Algorithm
NUSI Sub-table
Base Table

21. Building a NUSI helps when the selectivity of the indexed column is very
high.
Cost-based optimizer will determine when to access via NUSI:
Number of rows selected by NUSI must be less than number of blocks in the table
to justify access via NUSI (assumes even distribution of rows with NUSI value
within table).
Must also consider cost for reading the NUSI sub-table and building ROWID spool
file.
Note that the extreme efficiency of table scanning inTeradata reduces the need for
secondary indexing as compared to other databases.
When to build NUSI?

22. A unique secondary index (USI) is partitioned by the unique column
upon which the index is built.
Row access via a USI is a twoVamp operation.
First I/O is initiated on theVamp with the USI entry.
Second I/O is initiated on theVamp with the data row entry.
Secondary Index (USI)

23. When to Build a USI?
To allow data access without allVAMP operations.
Increased efficiency for (very) high selectivity retrievals.
Obtain co-location of index with frequently joined tables.
When to build USI?

24. Example:
create table order_header
(order_id decimal(12, 0) NOT NULL
,customer_id decimal(9, 0) NOT NULL
,order_dt date NOT NULL
...
)
primary index( customer_id );
create unique index oh_order_idx (order_id) on order_header;
create table order_detail
(order_id decimal(12, 0) NOT NULL
,product_id integer NOT NULL
,extended_price_amt decimal(15,2) NOT NULL
,item_cnt integer NOT NULL
...
)
primary index( order_id );
When to build USI?

25. Example: How many customers ordered green socks in the last month? (Assume that
green socks is quite selective)
select count(distinct order_header.customer_id)
from order_header
,order_detail ,product
where order_header.order_id = order_detail.order_id
and order.order_dt > add_months(date, -1)
and order_detail.product_id = product.product_id
and product.product_subcategory_cd = 'SOCKS'
and product.color_cd = 'GREEN'
;
The order_id USI on order_header table obviates the need for allVamp duplication of spool result from
order detail to product join when joining to the order header table.
When to build USI?

26. Lecture 18
Dr. Fawad Hussain
Primary and Secondary IndicesPrimary and Secondary IndicesPrimary and Secondary IndicesPrimary and Secondary Indices----IIIIIIIIIIII

27. Example: What is the average age (in years) of customers who live in California or
Massachusetts, completed a graduate degree, are consultants, and have a hobby of
volleyball or chess?
select avg( (days(date) - days(customer.birth_dt)) / 365.25
)
from customer
where customer.state_cd in (‘CA’ , MA’)
and customer.education_cd = ‘G’
and customer.occupation_cd = ‘CONSULTANT’
and customer.hobby_cd in (‘VOLLEYBALL’,‘CHESS’)
;
A Simple Query

28. Assume:
20M customers.
128 byte rows.
64K data block size.
Results in approximately 512 rows per block and a total of 39,063 blocks
in the customer table.
Note: We are ignoring block overhead for purposes of simplicity in calculations.
Sample Query Structure

29. Assume:
8% of customers live in California.
4% of customers live in Massachusetts.
4% of customers have completed a graduate degree.
6% of customers are consultants.
2% of customers have a primary hobby of chess.
3% of customers have a primary hobby of volleyball.
Data Demographics

30. Must read every block in the table.
Apply where clause predicates to determine which customers to include
in average.
Adjust numerator and denominator of average as appropriate.
Total I/O count = 39,063
Note: Data demographics have no (minimal) impact on query performance when using
a full table scan operation.
Full Table Scan

31. B-tree or hash organization of column values:
Index entries store row IDs (RIDs), lists of RIDs, or pointers to lists of
RIDs.
Originally designed for columns with many unique values (OLTP
legacy).
Assuming an eight byte RID, we will get 8096 RIDs per 64K block.
Single Index Structure

32. Optimizer chooses index with best selectivity based on values specified
in query.
1. Access next (first) index entry corresponding to specified column
value(s).
2. Use RID from index entry to locate row with specified column value.
3. Validate remaining predicates to qualify row.
4. Adjust average as appropriate.
5. Go to 2 until no more matching index values.
Single Index Access

33. What are my indexing choices?
state_cd (8% + 4% = 12% selectivity)
education_cd (4% selectivity)
occupation_cd (6% selectivity)
hobby_cd (2% + 3% = 5% selectivity)
Choose education_cd because it has best selectivity.
Single Index Access

34. Access via index on education_cd:
800,000 RIDs (4% of 20M)
99 blocks of RIDs to read
But...4% selectivity with 512 rows per block in the base table means that
800,000 selected RIDs will cause access to every block in the base table!
Total I/O count = 39,063 + 99 = 39,162
Worse than full table scan!
Single Index Access Performance

35. How do we calculate that number?
The selectivity on education is 4% (or 8 lac rows). If these rows were
consecutively distributed, then we would have to access
8lac rows/512 (rows per block)= 1563 blocks.
However, assuming equal distribution, and 4% selectivity, we will use
probability to find the distribution.That is, 4% selectivity gives us
0.04*512=20.48 rows (that are desired) are found in each block.
Hence, total I/O required is
8lacs /20.48 = 39062.5 bocks
Plus the additional 99 for accessing the index (8 lac Rows and 8096 RIDs
/block=99 blocks of RIDs to read)
The calculations