I have read that one of the tradeoffs for adding table indexes in SQL Server is the increased cost of insert/update/delete queries to benefit the performance of select queries.
I can conceptually understand what happens in the case of an insert because SQL Server has to write entries into each index matching the new rows, but update and delete are a little more murky to me because I can't quite wrap my head around what the database engine has to do.
Let's take DELETE as an example and assume I have the following schema (pardon the pseudo-SQL)
TABLE Foo
col1 int
,col2 int
,col3 int
,col4 int
PRIMARY KEY (col1,col2)
INDEX IX_1
col3
INCLUDE
col4
Now, if I issue the statement
DELETE FROM Foo WHERE col1=12 AND col2 > 34
I understand what the engine must do to update the table (or clustered index if you prefer). The index is set up to make it easy to find the range of rows to be removed and do so.
However, at this point it also needs to update IX_1 and the query that I gave it gives no obvious efficient way for the database engine to find the rows to update. Is it forced to do a full index scan at this point? Does the engine read the rows from the clustered index first and generate a smarter internal delete against the index?
It might help me to wrap my head around this if I understood better what is going on under the hood, but I guess my real question is this. I have a database that is spending a significant amount of time in delete and I'm trying to figure out what I can do about it.
When I display the execution plan for the deletion, it just shows an entry for "Clustered Index Delete" on table Foo which lists in the details section the other indices that need to be updated but I don't get any indication of the relative cost of these other indices.
Are they all equal in this case? Is there some way that I can estimate the impact of removing one or more of these indices without having to actually try it?
Nonclustered indexes also store the clustered keys.
It does not have to do a full scan, since:
your query will use the clustered index to locate rows
rows contain the other index value (c3)
using the other index value (c3) and the clustered index values (c1,c2), it can locate matching entries in the other index.
(Note: I had trouble interpreting the docs, but I would imagine that IX_1 in your case could be defined as if it was also sorted on c1,c2. Since these are already stored in the index, it would make perfect sense to use them to more efficiently locate records for e.g. updates and deletes.)
All this, however has a cost. For each matching row:
it has to read the row, to find out the value for c3
it has to find the entry for (c3,c1,c2) in the nonclustered index
it has to delete the entry from there as well.
Furthermore, while the range query can be efficient on the clustered index in your case (linear access, after finding a match), maintenance of the other indexes will most likely result in random access to them for every matching row. Random access has a much higher cost than just enumerating B+ tree leaf nodes starting from a given match.
Given the above query, more time is spent on the non-clustered index maintenance - the amount depends heavily on the number of records selected by the col1 = 12 AND col2 > 34
predicate.
My guess is that the cost is conceptually the same as if you did not have a secondary index but had e.g. a separate table, holding (c3,c1,c2) as the only columns in a clustered key and you did a DELETE for each matching row using (c3,c1,c2). Obviously, index maintenance is internal to SQL Server and is faster, but conceptually, I guess the above is close.
The above would mean that maintenance costs of indexes would stay pretty close to each other, since the number of entries in each secondary index is the same (the number of records) and deletion can proceed only one-by-one on each index.
If you need the indexes, performance-wise, depending on the number of deleted records, you might be better off scheduling the deletes, dropping the indexes - that are not used during the delete - before the delete and adding them back after. Depending on the number of records affected, rebuilding the indexes might be faster.
So I was reading up on indexes and their implementation, and I stumbled upon this website that has a brief explanation of b-tree indexes:
http://20bits.com/articles/interview-questions-database-indexes/
The b-tree index makes perfect sense for indexes that are only on a single column, but let's say I create an index with multiple columns, how then does the b-tree work? What is the value of each node in the b-tree?
For example, if I have this table:
table customer:
id number
name varchar
phone_number varchar
city varchar
and I create an index on: (id, name, city)
and then run the following query:
SELECT id, name
FROM customer
WHERE city = 'My City';
how does this query utilize the multiple column index, or does it not utilize it unless the index is created as (city, id, name) or (city, name, id) instead?
With most implementations, the key is simply a longer key that includes all of the key values, with a separator. No magic there ;-)
In your example the key values could look something like
"123499|John Doe|Conway, NH"
"32144|Bill Gates| Seattle, WA"
One of the characteristics of these indexes with composite keys is that the intermediate tree nodes can be used in some cases to "cover" the query.
For example, if the query is to find the Name and City given the ID, since the ID is first in the index, the index can search by this efficiently. Once in the intermediate node, it can "parse" the Name and City, from the key, and doesn't need to go to the leaf node to read the same.
If however the query wanted also to display the phone number, then the logic would follow down the leaf when the full record is found.
Imagine that the key is represented by a Python tuple (col1, col2, col3) ... the indexing operation involves comparing tuple_a with tuple_b ... if you have don't know which value of col1 and col2 that you are interested in, but only col3, then it would have to read the whole index ("full index scan"), which is not as efficient.
If you have an index on (col1, col2, col3), then you can expect that any RDBMS will use the index (in a direct manner) when the WHERE clause contains reference to (1) all 3 columns (2) both col1 and col2 (3) only col1.
Otherwise (e.g. only col3 in the WHERE clause), either the RDBMS will not use that index at all (e.g. SQLite), or will do a full index scan (e.g. Oracle) [if no other index is better].
In your specific example, presuming that id is a unique identifier of a customer, it is pointless to have it appear in an index (other than the index that your DBMS should set up for a primary key or column noted as UNIQUE).
Some implementations simply concatenate the values in the order of the columns, with delimiters.
Another solution is to simply have a b-tree within a b-tree. When you hit a leaf on the first column, you get both a list of matching records and a mini b-tree of the next column, and so on. Thus, the order of the columns specified in the index makes a huge difference on whether that index will be useful for particular queries.
Here's a related question I wrote last week:
Does SQL Server jump leaves when using a composite clustered index?
"The index will be ordered by the first key element, then by second key element and so on"
https://www.qwertee.io/blog/postgresql-b-tree-index-explained-part-1/
In Oracle a composite key index can be used even though the leading columns are not filtered. This is done through three mechanisms:
A fast full index scan, in which multiblock reads are used to traverse the entire index segment.
An index full scan, in which the index is read in the logical order of the blocks (I believe I read that in recent versions Oracle can use multiblock reads for this, but really you should count on single block reads)
An inddex skip scan, where a very low cardinality for the non-predicated leading columns allows Oracle to perform multiple index range scans, one for each unique value of the leading column(s). These are pretty rare in my experience.
Look for articles by Richard Foote or Jonathan Lewis for more information on Oracle index internals.
Other than the "composite key" mechanism already described, one possibility is a kdtree which works like a binary tree, but as you traverse each level you cycle through k dimensions. That is, the first level of the tree separates the first dimension into two parts, the second level splits the second dimension, the k+1th level splits the first dimension again, etc.. This allows for efficient partitioning of data in any number of dimensions. This approach is common in "spatial" databases (e.g., Oracle Spatial, PostGIS, etc.), but probably not as useful in "regular" multi-indexed tables.
http://en.wikipedia.org/wiki/Kd-tree
It can use the (id,name,city) index to satisfy a "City = ? " predicate, but very very inefficently.
In order to use the index to satisfy this query it would need to walk most of tree structure looking for entries with the desired city. This is still probably an order of magnatude faster than scanning the table!
An index of (city,name,id) would be the best index for your query. It would find all the desired city entries easily and would not need to access the underlying table to get the id and name values.
I have a limited exposure to DB and have only used DB as an application programmer. I want to know about Clustered and Non clustered indexes.
I googled and what I found was :
A clustered index is a special type of index that reorders the way
records in the table are physically
stored. Therefore table can have only
one clustered index. The leaf nodes
of a clustered index contain the data
pages. A nonclustered index is a
special type of index in which the
logical order of the index does not
match the physical stored order of
the rows on disk. The leaf node of a
nonclustered index does not consist of
the data pages. Instead, the leaf
nodes contain index rows.
What I found in SO was What are the differences between a clustered and a non-clustered index?.
Can someone explain this in plain English?
With a clustered index the rows are stored physically on the disk in the same order as the index. Therefore, there can be only one clustered index.
With a non clustered index there is a second list that has pointers to the physical rows. You can have many non clustered indices, although each new index will increase the time it takes to write new records.
It is generally faster to read from a clustered index if you want to get back all the columns. You do not have to go first to the index and then to the table.
Writing to a table with a clustered index can be slower, if there is a need to rearrange the data.
A clustered index means you are telling the database to store close values actually close to one another on the disk. This has the benefit of rapid scan / retrieval of records falling into some range of clustered index values.
For example, you have two tables, Customer and Order:
Customer
----------
ID
Name
Address
Order
----------
ID
CustomerID
Price
If you wish to quickly retrieve all orders of one particular customer, you may wish to create a clustered index on the "CustomerID" column of the Order table. This way the records with the same CustomerID will be physically stored close to each other on disk (clustered) which speeds up their retrieval.
P.S. The index on CustomerID will obviously be not unique, so you either need to add a second field to "uniquify" the index or let the database handle that for you but that's another story.
Regarding multiple indexes. You can have only one clustered index per table because this defines how the data is physically arranged. If you wish an analogy, imagine a big room with many tables in it. You can either put these tables to form several rows or pull them all together to form a big conference table, but not both ways at the same time. A table can have other indexes, they will then point to the entries in the clustered index which in its turn will finally say where to find the actual data.
In SQL Server, row-oriented storage both clustered and nonclustered indexes are organized as B trees.
(Image Source)
The key difference between clustered indexes and non clustered indexes is that the leaf level of the clustered index is the table. This has two implications.
The rows on the clustered index leaf pages always contain something for each of the (non-sparse) columns in the table (either the value or a pointer to the actual value).
The clustered index is the primary copy of a table.
Non clustered indexes can also do point 1 by using the INCLUDE clause (Since SQL Server 2005) to explicitly include all non-key columns but they are secondary representations and there is always another copy of the data around (the table itself).
CREATE TABLE T
(
A INT,
B INT,
C INT,
D INT
)
CREATE UNIQUE CLUSTERED INDEX ci ON T(A, B)
CREATE UNIQUE NONCLUSTERED INDEX nci ON T(A, B) INCLUDE (C, D)
The two indexes above will be nearly identical. With the upper-level index pages containing values for the key columns A, B and the leaf level pages containing A, B, C, D
There can be only one clustered index per table, because the data rows
themselves can be sorted in only one order.
The above quote from SQL Server books online causes much confusion
In my opinion, it would be much better phrased as.
There can be only one clustered index per table because the leaf level rows of the clustered index are the table rows.
The book's online quote is not incorrect but you should be clear that the "sorting" of both non clustered and clustered indices is logical, not physical. If you read the pages at leaf level by following the linked list and read the rows on the page in slot array order then you will read the index rows in sorted order but physically the pages may not be sorted. The commonly held belief that with a clustered index the rows are always stored physically on the disk in the same order as the index key is false.
This would be an absurd implementation. For example, if a row is inserted into the middle of a 4GB table SQL Server does not have to copy 2GB of data up in the file to make room for the newly inserted row.
Instead, a page split occurs. Each page at the leaf level of both clustered and non clustered indexes has the address (File: Page) of the next and previous page in logical key order. These pages need not be either contiguous or in key order.
e.g. the linked page chain might be 1:2000 <-> 1:157 <-> 1:7053
When a page split happens a new page is allocated from anywhere in the filegroup (from either a mixed extent, for small tables or a non-empty uniform extent belonging to that object or a newly allocated uniform extent). This might not even be in the same file if the filegroup contains more than one.
The degree to which the logical order and contiguity differ from the idealized physical version is the degree of logical fragmentation.
In a newly created database with a single file, I ran the following.
CREATE TABLE T
(
X TINYINT NOT NULL,
Y CHAR(3000) NULL
);
CREATE CLUSTERED INDEX ix
ON T(X);
GO
--Insert 100 rows with values 1 - 100 in random order
DECLARE #C1 AS CURSOR,
#X AS INT
SET #C1 = CURSOR FAST_FORWARD
FOR SELECT number
FROM master..spt_values
WHERE type = 'P'
AND number BETWEEN 1 AND 100
ORDER BY CRYPT_GEN_RANDOM(4)
OPEN #C1;
FETCH NEXT FROM #C1 INTO #X;
WHILE ##FETCH_STATUS = 0
BEGIN
INSERT INTO T (X)
VALUES (#X);
FETCH NEXT FROM #C1 INTO #X;
END
Then checked the page layout with
SELECT page_id,
X,
geometry::Point(page_id, X, 0).STBuffer(1)
FROM T
CROSS APPLY sys.fn_PhysLocCracker( %% physloc %% )
ORDER BY page_id
The results were all over the place. The first row in key order (with value 1 - highlighted with an arrow below) was on nearly the last physical page.
Fragmentation can be reduced or removed by rebuilding or reorganizing an index to increase the correlation between logical order and physical order.
After running
ALTER INDEX ix ON T REBUILD;
I got the following
If the table has no clustered index it is called a heap.
Non clustered indexes can be built on either a heap or a clustered index. They always contain a row locator back to the base table. In the case of a heap, this is a physical row identifier (rid) and consists of three components (File:Page: Slot). In the case of a Clustered index, the row locator is logical (the clustered index key).
For the latter case if the non clustered index already naturally includes the CI key column(s) either as NCI key columns or INCLUDE-d columns then nothing is added. Otherwise, the missing CI key column(s) silently gets added to the NCI.
SQL Server always ensures that the key columns are unique for both types of indexes. The mechanism in which this is enforced for indexes not declared as unique differs between the two index types, however.
Clustered indexes get a uniquifier added for any rows with key values that duplicate an existing row. This is just an ascending integer.
For non clustered indexes not declared as unique SQL Server silently adds the row locator into the non clustered index key. This applies to all rows, not just those that are actually duplicates.
The clustered vs non clustered nomenclature is also used for column store indexes. The paper Enhancements to SQL Server Column Stores states
Although column store data is not really "clustered" on any key, we
decided to retain the traditional SQL Server convention of referring
to the primary index as a clustered index.
I realize this is a very old question, but I thought I would offer an analogy to help illustrate the fine answers above.
CLUSTERED INDEX
If you walk into a public library, you will find that the books are all arranged in a particular order (most likely the Dewey Decimal System, or DDS). This corresponds to the "clustered index" of the books. If the DDS# for the book you want was 005.7565 F736s, you would start by locating the row of bookshelves that is labeled 001-099 or something like that. (This endcap sign at the end of the stack corresponds to an "intermediate node" in the index.) Eventually you would drill down to the specific shelf labelled 005.7450 - 005.7600, then you would scan until you found the book with the specified DDS#, and at that point you have found your book.
NON-CLUSTERED INDEX
But if you didn't come into the library with the DDS# of your book memorized, then you would need a second index to assist you. In the olden days you would find at the front of the library a wonderful bureau of drawers known as the "Card Catalog". In it were thousands of 3x5 cards -- one for each book, sorted in alphabetical order (by title, perhaps). This corresponds to the "non-clustered index". These card catalogs were organized in a hierarchical structure, so that each drawer would be labeled with the range of cards it contained (Ka - Kl, for example; i.e., the "intermediate node"). Once again, you would drill in until you found your book, but in this case, once you have found it (i.e, the "leaf node"), you don't have the book itself, but just a card with an index number (the DDS#) with which you could find the actual book in the clustered index.
Of course, nothing would stop the librarian from photocopying all the cards and sorting them in a different order in a separate card catalog. (Typically there were at least two such catalogs: one sorted by author name, and one by title.) In principle, you could have as many of these "non-clustered" indexes as you want.
Find below some characteristics of clustered and non-clustered indexes:
Clustered Indexes
Clustered indexes are indexes that uniquely identify the rows in an SQL table.
Every table can have exactly one clustered index.
You can create a clustered index that covers more than one column. For example: create Index index_name(col1, col2, col.....).
By default, a column with a primary key already has a clustered index.
Non-clustered Indexes
Non-clustered indexes are like simple indexes. They are just used for fast retrieval of data. Not sure to have unique data.
Clustered Index
A clustered index determines the physical order of DATA in a table. For this reason, a table has only one clustered index(Primary key/composite key).
"Dictionary" No need of any other Index, its already Index according to words
Nonclustered Index
A non-clustered index is analogous to an index in a Book. The data is stored in one place. The index is stored in another place and the index has pointers to the storage location. this help in the fast search of data. For this reason, a table has more than 1 Nonclustered index.
"Biology Book" at starting there is a separate index to point Chapter location and At the "END" there is another Index pointing the common WORDS location
A very simple, non-technical rule-of-thumb would be that clustered indexes are usually used for your primary key (or, at least, a unique column) and that non-clustered are used for other situations (maybe a foreign key). Indeed, SQL Server will by default create a clustered index on your primary key column(s). As you will have learnt, the clustered index relates to the way data is physically sorted on disk, which means it's a good all-round choice for most situations.
Clustered Index
A Clustered Index is basically a tree-organized table. Instead of storing the records in an unsorted Heap table space, the clustered index is actually B+Tree index having the Leaf Nodes, which are ordered by the clusters key column value, store the actual table records, as illustrated by the following diagram.
The Clustered Index is the default table structure in SQL Server and MySQL. While MySQL adds a hidden clusters index even if a table doesn't have a Primary Key, SQL Server always builds a Clustered Index if a table has a Primary Key column. Otherwise, the SQL Server is stored as a Heap Table.
The Clustered Index can speed up queries that filter records by the clustered index key, like the usual CRUD statements. Since the records are located in the Leaf Nodes, there's no additional lookup for extra column values when locating records by their Primary Key values.
For example, when executing the following SQL query on SQL Server:
SELECT PostId, Title
FROM Post
WHERE PostId = ?
You can see that the Execution Plan uses a Clustered Index Seek operation to locate the Leaf Node containing the Post record, and there are only two logical reads required to scan the Clustered Index nodes:
|StmtText |
|-------------------------------------------------------------------------------------|
|SELECT PostId, Title FROM Post WHERE PostId = #P0 |
| |--Clustered Index Seek(OBJECT:([high_performance_sql].[dbo].[Post].[PK_Post_Id]), |
| SEEK:([high_performance_sql].[dbo].[Post].[PostID]=[#P0]) ORDERED FORWARD) |
Table 'Post'. Scan count 0, logical reads 2, physical reads 0
Non-Clustered Index
Since the Clustered Index is usually built using the Primary Key column values, if you want to speed up queries that use some other column, then you'll have to add a Secondary Non-Clustered Index.
The Secondary Index is going to store the Primary Key value in its Leaf Nodes, as illustrated by the following diagram:
So, if we create a Secondary Index on the Title column of the Post table:
CREATE INDEX IDX_Post_Title on Post (Title)
And we execute the following SQL query:
SELECT PostId, Title
FROM Post
WHERE Title = ?
We can see that an Index Seek operation is used to locate the Leaf Node in the IDX_Post_Title index that can provide the SQL query projection we are interested in:
|StmtText |
|------------------------------------------------------------------------------|
|SELECT PostId, Title FROM Post WHERE Title = #P0 |
| |--Index Seek(OBJECT:([high_performance_sql].[dbo].[Post].[IDX_Post_Title]),|
| SEEK:([high_performance_sql].[dbo].[Post].[Title]=[#P0]) ORDERED FORWARD)|
Table 'Post'. Scan count 1, logical reads 2, physical reads 0
Since the associated PostId Primary Key column value is stored in the IDX_Post_Title Leaf Node, this query doesn't need an extra lookup to locate the Post row in the Clustered Index.
Clustered Index
Clustered indexes sort and store the data rows in the table or view based on their key values. These are the columns included in the index definition. There can be only one clustered index per table, because the data rows themselves can be sorted in only one order.
The only time the data rows in a table are stored in sorted order is when the table contains a clustered index. When a table has a clustered index, the table is called a clustered table. If a table has no clustered index, its data rows are stored in an unordered structure called a heap.
Nonclustered
Nonclustered indexes have a structure separate from the data rows. A nonclustered index contains the nonclustered index key values and each key value entry has a pointer to the data row that contains the key value.
The pointer from an index row in a nonclustered index to a data row is called a row locator. The structure of the row locator depends on whether the data pages are stored in a heap or a clustered table. For a heap, a row locator is a pointer to the row. For a clustered table, the row locator is the clustered index key.
You can add nonkey columns to the leaf level of the nonclustered index to by-pass existing index key limits, and execute fully covered, indexed, queries. For more information, see Create Indexes with Included Columns. For details about index key limits see Maximum Capacity Specifications for SQL Server.
Reference: https://learn.microsoft.com/en-us/sql/relational-databases/indexes/clustered-and-nonclustered-indexes-described
Let me offer a textbook definition on "clustering index", which is taken from 15.6.1 from Database Systems: The Complete Book:
We may also speak of clustering indexes, which are indexes on an attribute or attributes such that all of tuples with a fixed value for the search key of this index appear on roughly as few blocks as can hold them.
To understand the definition, let's take a look at Example 15.10 provided by the textbook:
A relation R(a,b) that is sorted on attribute a and stored in that
order, packed into blocks, is surely clusterd. An index on a is a
clustering index, since for a given a-value a1, all the tuples with
that value for a are consecutive. They thus appear packed into
blocks, execept possibly for the first and last blocks that contain
a-value a1, as suggested in Fig.15.14. However, an index on b is
unlikely to be clustering, since the tuples with a fixed b-value
will be spread all over the file unless the values of a and b are
very closely correlated.
Note that the definition does not enforce the data blocks have to be contiguous on the disk; it only says tuples with the search key are packed into as few data blocks as possible.
A related concept is clustered relation. A relation is "clustered" if its tuples are packed into roughly as few blocks as can possibly hold those tuples. In other words, from a disk block perspective, if it contains tuples from different relations, then those relations cannot be clustered (i.e., there is a more packed way to store such relation by swapping the tuples of that relation from other disk blocks with the tuples the doesn't belong to the relation in the current disk block). Clearly, R(a,b) in example above is clustered.
To connect two concepts together, a clustered relation can have a clustering index and nonclustering index. However, for non-clustered relation, clustering index is not possible unless the index is built on top of the primary key of the relation.
"Cluster" as a word is spammed across all abstraction levels of database storage side (three levels of abstraction: tuples, blocks, file). A concept called "clustered file", which describes whether a file (an abstraction for a group of blocks (one or more disk blocks)) contains tuples from one relation or different relations. It doesn't relate to the clustering index concept as it is on file level.
However, some teaching material likes to define clustering index based on the clustered file definition. Those two types of definitions are the same on clustered relation level, no matter whether they define clustered relation in terms of data disk block or file. From the link in this paragraph,
An index on attribute(s) A on a file is a clustering index when: All tuples with attribute value A = a are stored sequentially (= consecutively) in the data file
Storing tuples consecutively is the same as saying "tuples are packed into roughly as few blocks as can possibly hold those tuples" (with minor difference on one talking about file, the other talking about disk). It's because storing tuple consecutively is the way to achieve "packed into roughly as few blocks as can possibly hold those tuples".
Clustered Index:
Primary Key constraint creates clustered Index automatically if no clustered Index already exists on the table. Actual data of clustered index can be stored at leaf level of Index.
Non Clustered Index:
Actual data of non clustered index is not directly found at leaf node, instead it has to take an additional step to find because it has only values of row locators pointing towards actual data.
Non clustered Index can't be sorted as clustered index. There can be multiple non clustered indexes per table, actually it depends on the sql server version we are using. Basically Sql server 2005 allows 249 Non Clustered Indexes and for above versions like 2008, 2016 it allows 999 Non Clustered Indexes per table.
Clustered Index - A clustered index defines the order in which data is physically stored in a table. Table data can be sorted in only way, therefore, there can be only one clustered index per table. In SQL Server, the primary key constraint automatically creates a clustered index on that particular column.
Non-Clustered Index - A non-clustered index doesn’t sort the physical data inside the table. In fact, a non-clustered index is stored at one place and table data is stored in another place. This is similar to a textbook where the book content is located in one place and the index is located in another. This allows for more than one non-clustered index per table.It is important to mention here that inside the table the data will be sorted by a clustered index. However, inside the non-clustered index data is stored in the specified order. The index contains column values on which the index is created and the address of the record that the column value belongs to.When a query is issued against a column on which the index is created, the database will first go to the index and look for the address of the corresponding row in the table. It will then go to that row address and fetch other column values. It is due to this additional step that non-clustered indexes are slower than clustered indexes
Differences between clustered and Non-clustered index
There can be only one clustered index per table. However, you can
create multiple non-clustered indexes on a single table.
Clustered indexes only sort tables. Therefore, they do not consume
extra storage. Non-clustered indexes are stored in a separate place
from the actual table claiming more storage space.
Clustered indexes are faster than non-clustered indexes since they
don’t involve any extra lookup step.
For more information refer to this article.