SQL Server Architecture - SQL CLR

SQL CLR
Microsoft SQL Server 2005 includes a component named SQL CLR ("Common Language Runtime") via which it integrates with .NET Framework. Unlike most other applications that use .NET Framework, SQL Server itself hosts the .NET Framework runtime, i.e., memory, threading and resource management requirements of .NET Framework are satisfied by SQLOS itself, rather than the underlying Windows operating system. SQLOS provides deadlock detection and resolution services for .NET code as well. With SQL CLR, stored procedures and triggers can be written in any managed .NET language, including C# and VB.NET. Managed code can also be used to define UDT's (user defined types), which can persist in the database. Managed code is compiled to .NET assemblies and after being verified for type safety, registered at the database. After that, they can be invoked like any other procedure. However, only a subset of the Base Class Library is available, when running code under SQL CLR. Most APIs relating to user interface functionality are not available.

When writing code for SQL CLR, data stored in SQL Server databases can be accessed using the ADO.NET APIs like any other managed application that accesses SQL Server data. However, doing that creates a new database session, different from the one in which the code is executing. To avoid this, SQL Server provides some enhancements to the ADO.NET provider that allows the connection to be redirected to the same session which already hosts the running code. Such connections are called context connections and are set by setting context connection parameter to true in the connection string. SQL Server also provides several other enhancements to the ADO.NET API, including classes to work with tabular data or a single row of data as well as classes to work with internal metadata about the data stored in the database. It also provides access to the XML features in SQL Server, including XQuery support. These enhancements are also available in T-SQL Procedures in consequence of the introduction of the new XML Datatype (query,value,nodes functions).

Read More

SQL Server Architecture - Data retrieval

Data retrieval
The main mode of retrieving data from an SQL Server database is querying for it. The query is expressed using a variant of SQL called T-SQL, a dialect Microsoft SQL Server shares with Sybase SQL Server due to its legacy. The query declaratively specifies what is to be retrieved. It is processed by the query processor, which figures out the sequence of steps that will be necessary to retrieve the requested data. The sequence of actions necessary to execute a query is called a query plan. There might be multiple ways to process the same query. For example, for a query that contains a join statement and a select statement, executing join on both the tables and then executing select on the results would give the same result as selecting from each table and then executing the join, but result in different execution plans. In such case, SQL Server chooses the plan that is expected to yield the results in the shortest possible time. This is called query optimization and is performed by the query processor itself.

SQL Server includes a cost-based query optimizer which tries to optimize on the cost, in terms of the resources it will take to execute the query. Given a query, then the query optimizer looks at the database schema, the database statistics and the system load at that time. It then decides which sequence to access the tables referred in the query, which sequence to execute the operations and what access method to be used to access the tables. For example, if the table has an associated index, whether the index should be used or not - if the index is on a column which is not unique for most of the columns (low "selectivity"), it might not be worthwhile to use the index to access the data. Finally, it decides whether to execute the query concurrently or not. While a concurrent execution is more costly in terms of total processor time, because the execution is actually split to different processors might mean it will execute faster. Once a query plan is generated for a query, it is temporarily cached. For further invocations of the same query, the cached plan is used. Unused plans are discarded after some time.

SQL Server also allows stored procedures to be defined. Stored procedures are parameterized T-SQL queries, that are stored in the server itself (and not issued by the client application as is the case with general queries). Stored procedures can accept values sent by the client as input parameters, and send back results as output parameters. They can call defined functions, and other stored procedures, including the same stored procedure (up to a set number of times). They can be selectively provided access to. Unlike other queries, stored procedures have an associated name, which is used at runtime to resolve into the actual queries. Also because the code need not be sent from the client every time (as it can be accessed by name), it reduces network traffic and somewhat improves performance. Execution plans for stored procedures are also cached as necessary.

Read More

SQL Server Architecture - Concurrency and locking

Concurrency and locking
SQL Server allows multiple clients to use the same database concurrently. As such, it needs to control concurrent access to shared data, to ensure data integrity - when multiple clients update the same data, or clients attempt to read data that is in the process of being changed by another client. SQL Server provides two modes of concurrency control: pessimistic concurrency and optimistic concurrency. When pessimistic concurrency control is being used, SQL Server controls concurrent access by using locks. Locks can be either shared or exclusive. Exclusive lock grants the user exclusive access to the data - no other user can access the data as long as the lock is held. Shared locks are used when some data is being read - multiple users can read from data locked with a shared lock, but not acquire an exclusive lock. The latter would have to wait for all shared locks to be released. Locks can be applied on different levels of granularity - on entire tables, pages, or even on a per-row basis on tables. For indexes, it can either be on the entire index or on index leaves. The level of granularity to be used is defined on a per-database basis by the database administrator. While a fine grained locking system allows more users to use the table or index simultaneously, it requires more resources. So it does not automatically turn into higher performing solution. SQL Server also includes two more lightweight mutual exclusion solutions - latches and spinlocks - which are less robust than locks but are less resource intensive. SQL Server uses them for DMVs and other resources that are usually not busy. SQL Server also monitors all worker threads that acquire locks to ensure that they do not end up in deadlocks - in case they do, SQL Server takes remedial measures, which in many cases is to kill one of the threads entangled in a deadlock and rollback the transaction it started. To implement locking, SQL Server contains the Lock Manager. The Lock Manager maintains an in-memory table that manages the database objects and locks, if any, on them along with other metadata about the lock. Access to any shared object is mediated by the lock manager, which either grants access to the resource or blocks it.

SQL Server also provides the optimistic concurrency control mechanism, which is similar to the multiversion concurrency control used in other databases. The mechanism allows a new version of a row to be created whenever the row is updated, as opposed to overwriting the row, i.e., a row is additionally identified by the ID of the transaction that created the version of the row. Both the old as well as the new versions of the row are stored and maintained, though the old versions are moved out of the database into a system database identified as Tempdb. When a row is in the process of being updated, any other requests are not blocked (unlike locking) but are executed on the older version of the row. If the other request is an update statement, it will result in two different versions of the rows - both of them will be stored by the database, identified by their respective transaction IDs.

Read More

SQL Server Architecture - Logging and Transaction

Logging and Transaction
SQL Server ensures that any change to the data is ACID-compliant, i.e. it uses transactions to ensure that the database will always revert to a known consistent state on failure. Each transaction may consist of multiple SQL statements all of which will only make a permanent change to the database if the last statement in the transaction (a COMMIT statement) completes successfully. If the COMMIT successfully completes the transaction is safely on disk.

SQL Server implements transactions using a write-ahead log.

Any changes made to any page will update the in-memory cache of the page, simultaneously all the operations performed will be written to a log, along with the transaction ID which the operation was a part of. Each log entry is identified by an increasing Log Sequence Number (LSN) which is used to ensure that all changes are written to the data files. Also during a log restore it is used to check that no logs are duplicated or skipped. SQL Server requires that the log is written onto the disc before the data page is written back. It must also ensure that all operations in a transaction are written to the log before any COMMIT operation is reported as completed.

At a later point the server will checkpoint the database and ensure that all pages in the data files have the state of their contents synchronised to a point at or after the LSN that the checkpoint started. When completed the checkpoint marks that portion of the log file as complete and may free it (see Simple transaction logging vs Full transaction logging). This enables SQL Server to ensure integrity of the data, even if the system fails.

On failure the database log has to be replayed to ensure the data files are in a consistent state. All pages stored in the roll forward part of the log (not marked as completed) are rewritten to the database, when the end of the log is reached all open transactions are rolled back using the roll back portion of the log file.

The database engine usually checkpoints quite frequently. However, in a heavily loaded database this can have a significant performance impact. It is possible to reduce the frequency of checkpoints or disable them completely but the rollforward during a recovery will take much longer

Read More

SQL Server Architecture - Buffer Management

Buffer management
SQL Server buffers pages in RAM to minimize disc I/O. Any 8 KB page can be buffered in-memory, and the set of all pages currently buffered is called the buffer cache. The amount of memory available to SQL Server decides how many pages will be cached in memory. The buffer cache is managed by the Buffer Manager. Either reading from or writing to any page copies it to the buffer cache. Subsequent reads or writes are redirected to the in-memory copy, rather than the on-disc version. The page is updated on the disc by the Buffer Manager only if the in-memory cache has not been referenced for some time. While writing pages back to disc, asynchronous I/O is used whereby the I/O operation is done in a background thread so that other operations do not have to wait for the I/O operation to complete. Each page is written along with its checksum when it is written. When reading the page back, its checksum is computed again and matched with the stored version to ensure the page has not been damaged or tampered with in the meantime.

Read More

SQL Server Architecture - Protocol Layer

Protocol layer
Protocol layer implements the external interface to SQL Server. All operations that can be invoked on SQL Server are communicated to it via a Microsoft-defined format, called Tabular Data Stream (TDS). TDS is an application layer protocol, used to transfer data between a database server and a client. Initially designed and developed by Sybase Inc. for their Sybase SQL Server relational database engine in 1984, and later by Microsoft in Microsoft SQL Server, TDS packets can be encased in other physical transport dependent protocols, including TCP/IP, Named pipes, and Shared memory. Consequently, access to SQL Server is available over these protocols. In addition, the SQL Server API is also exposed over web services.

Read More

SQL Server Architecture - Data Storage

Data storage
The main unit of data storage is a database, which is a collection of tables with typed columns. SQL Server supports different data types, including primary types such as Integer, Float, Decimal, Char (including character strings), Varchar (variable length character strings), binary (for unstructured blobs of data), Text (for textual data) among others. The rounding of floats to integers uses either Symmetric Arithmetic Rounding or Symmetric Round Down (Fix) depending on arguments: SELECT Round(2.5, 0) gives 3.

Microsoft SQL Server also allows user-defined composite types (UDTs) to be defined and used. It also makes server statistics available as virtual tables and views (called Dynamic Management Views or DMVs). In addition to tables, a database can also contain other objects including views, stored procedures, indexes and constraints, along with a transaction log. A SQL Server database can contain a maximum of 231 objects, and can span multiple OS-level files with a maximum file size of 220 TB. The data in the database are stored in primary data files with an extension .mdf. Secondary data files, identified with a .ndf extension, are used to store optional metadata. Log files are identified with the .ldf extension.

Storage space allocated to a database is divided into sequentially numbered pages, each 8 KB in size. A page is the basic unit of I/O for SQL Server operations. A page is marked with a 96-byte header which stores metadata about the page including the page number, page type, free space on the page and the ID of the object that owns it. Page type defines the data contained in the page - data stored in the database, index, allocation map which holds information about how pages are allocated to tables and indexes, change map which holds information about the changes made to other pages since last backup or logging, or contain large data types such as image or text. While page is the basic unit of an I/O operation, space is actually managed in terms of an extent which consists of 8 pages. A database object can either span all 8 pages in an extent ("uniform extent") or share an extent with up to 7 more objects ("mixed extent"). A row in a database table cannot span more than one page, so is limited to 8 KB in size. However, if the data exceeds 8 KB and the row contains Varchar or Varbinary data, the data in those columns are moved to a new page (or possibly a sequence of pages, called an Allocation unit) and replaced with a pointer to the data.

For physical storage of a table, its rows are divided into a series of partitions (numbered 1 to n). The partition size is user defined; by default all rows are in a single partition. A table is split into multiple partitions in order to spread a database over a cluster. Rows in each partition are stored in either B-tree or heap structure. If the table has an associated index to allow fast retrieval of rows, the rows are stored in-order according to their index values, with a B-tree providing the index. The data is in the leaf node of the leaves, and other nodes storing the index values for the leaf data reachable from the respective nodes. If the index is non-clustered, the rows are not sorted according to the index keys. An indexed view has the same storage structure as an indexed table. A table without an index is stored in an unordered heap structure. Both heaps and B-trees can span multiple allocation units.

Read More