Basics of SQL Transactions

Transcription

1 Basics of SQL Transactions Big Picture for understanding COMMIT and ROLLBACK of SQL transactions Files, Buffers,, Service Threads, and Transactions

2 (Flat) SQL Transaction [BEGIN TRANSACTION] [SET TRANSACTION ISOLATION.. ] SELECT INSERT UPDATE DELETE Database Transaction log COMMIT or ROLLBACK

3 Overview of a Database Server Instance connections (sessions) transactions - SQL commands & results network main memory listener & service threads bufferpool(s) DBMS instance processes (threads) x startup disc DBMS software instance control files database control files database files tablespace files transaction logs log cache(s) trace & alert files

4 Database Instance Architectures DB2 Instance Oracle Instance control files directories bufferpools bufferpools control files directories bufferpools database database database For every database the active transaction logs build a circular chain Oracle calls the chained active transaction logs as redo-logs SQL Server Instance A server Instance may include just one or multiple databases depending on the DBMS product. In the following we will focus only on the simple case of a single database System databases master bufferpool database database The active database transaction logs are allocated as Virtual Log Files in a single file building a circular chain Martti Laiho

5 Files and Caches listening for connection requests of clients Applications (clients) Transaction control, etc WAL write-ahead-logging protocol Cache of before and after images of rows of active transactions data pages for rows Cache of data pages for fast data processing minimizing disk I/O Checkpoint LRU protocol for freeing pages on disks building the tablespace for tables and indices (for history) Circular chain of

6 Structures of Data Pages Control data including pointers, dirty bit, LSN, etc PCTFREE percentage of originally free space on pages Page header Record of the row row Page header Record rest of the row Slot directory Slot directory Address of a row is RID (ROWID) File : Page : Slot Every data file has an internal file in the database and the file space is managed as a sequence of pages. Page is the ordinal number of the page in the file. A typical page size nowadays is 4, 8, 16 or 32 KB. A row may continue on other pages or original record may contain only a link to the new address, preserving the original RID value of the row s Address. Martti Laiho

7 SQL Session: A client starts its SQL connection.. connect?

8 .. gets a service thread connect?

9 .. Client enters a SQL command to update a row (still on disk) This first SQL command starts a new transaction update T set.. where id=.. Transaction tr begin The transaction gets transaction ID () and a begin_transaction record is written to the

10 DBMS reads the page of the row to bufferpool update T set.. where id=.. Transaction tr... begin....

11 The row is updated and the updated page is marked with the Dirty Bit update T set.. where id=.. Transaction tr begin dirty bit

12 A log record of the before and after image of the row is written to log cache update T set.. where id=.. Transaction tr dirty bit begin tr & before image & after image

13 On commit the log records of the transaction are written to the transaction log COMMIT Transaction tr dirty bit begin begin commit

14 .. but in case of ROLLBACK the before images are returned to original addresses ROLLBACK Transaction tr dirty bit begin begin rollback Martti Laiho

15 .. Later the client closes its SQL connection disconnect dirty bit begin commit

16 .. Some other pages may also get updated by SQL sessions of clients tr(s)... X xx X xx begin commit

17 From time to time a checkpoint stops the services.. Checkpoint tr(s)... X xx write-ahead-logging (WAL) 1. begin commit Martti Laiho

18 and all dirty pages are written to data files clearing the dirty bits Checkpoint tr(s) begin commit X xx Martti Laiho

19 and writes checkpoint record tr(s) Checkpoint... Checkpoint record begin commit - Chk Begin - tr list - Chk End CBg CEn Martti Laiho X xx CBg CEn

20 After the checkpoint the serving of clients can continue tr(s)... begin commit X xx CBg CEn

21 Contents of full transaction log files are copied to archive.. tr(s)... begin commit X xx CBg CEn begin commit

22 .. and the space of the copied transaction log file can be reused tr(s)... X xx CBg CEn begin commit

23 A logging test using SQL Server 2012 based on the original exercise by Kari Silpiö Page header Page ID 1:77 Slots

24 Some Notes.. The Least Recently Used (LRU) procedure is needed when DBMS needs to recover free space in the bufferpool(s). The pages which have been unused for the longest time and have dirty bit reset, will be cleared for new pages to be read from the data files. The active, circular transaction log files are the most important files of a database, since they contain data of the latest committed transactions! Dual logging (replicated) is recommended for these, and on dedicated disks. Based on the last checkpoint record and the circular transaction log a DBMS can recover the database after a soft crash automatically by roll-back recovery of failed and roll-forward recovery of the committed transactions to the level of the latest committed transasction before the soft crash. In case of hardware crash the database content need to be restored from the database backup and the applying roll-forward recovery of the transaction history from the archive and the latest circular logs, and finally rollback recovery of the latest failed transactions. These topics are covered in separate tutorials.