Database Recovery. Haengrae Cho Yeungnam University. Database recovery. Introduction to Database Systems

Transcription

1 Database Recovery Haengrae Cho Yeungnam University Database recovery. Introduction to Database Systems Report Yeungnam University, Database Lab. Chapter 1-1

2 1. Introduction to Database Recovery 2. Recovery in Client-Server DBMS 3. Recovery in Shared Disks 4. Disaster Recovery 5. Persistent Applications Yeungnam University, Database Lab. Chapter 1-2 Chapter 1 Principles of Transaction- Oriented Database Recovery : T. Haerder and A. Reuter, Principles of Transaction-Oriented Database Recovery, ACM Computing Surveys 15(4)

3 Table of Contents 1. Introduction 2. The Mapping Hierarchy of a DBMS 3. Crash Recovery 4. Archive Recovery Yeungnam University, Database Lab. Chapter Introduction What is a Transaction? Which Failures have to be Anticipated? Summary of Recovery Actions Yeungnam University, Database Lab. Chapter 1-5

4 1.1 What is a Transaction? Transaction : ñ : ú ñ Begin Transaction End Transaction partially committed committed active failed aborted Yeungnam University, Database Lab. Chapter 1-6 Example of a Transaction Begin_transaction Fund_Transfer begin input(account_debit, account_credit, amount); EXEC SQL SELECT balance INTO :d_balance FROM Accounts WHERE number = :account_debit; if d_balance < amount then { output( ); ABORT } else { EXEC SQL UPDATE Accounts SET balance = balance amount WHERE number = :account_debit; EXEC SQL UPDATE Accounts SET balance = balance amount WHERE number = :account_credit; COMMIT; Output( ) } end. Yeungnam University, Database Lab. Chapter 1-7

5 Transaction Properties Atomicity Consistency Isolation Durability Yeungnam University, Database Lab. Chapter Which Failures have to be Anticipated Transaction Failure Bad input, Overflow, Data not found, Deadlock, times per minute System Failure DBMS OS ( : ) Volatile memory Several times a week Media Failure HDD crash, H/W errors in the channel or disk controller, Nonvolatile memory Once or twice a year Yeungnam University, Database Lab. Chapter 1-9

6 1.3 Summary of Recovery Actions Transaction Consistent State : A database is consistent if and only if it contains the results of successful transactions. : Failure Commitþ Commitÿ þ. Action: UNDO, REDO T1 T2 T3 T4 T5 System Crash Yeungnam University, Database Lab. Chapter 1-10 Four Recovery Actions Transaction UNDO Global UNDO, þ Partial REDO, ÿ Global REDO (Archive Recovery), ÿ archive DB Yeungnam University, Database Lab. Chapter 1-11

7 2. Mapping Hierarchy of DBMS The Mapping Process: Objects and Operations The Storage Hierarchy: Implementation Environment Different Views of a Database Mapping Concepts for Updates Yeungnam University, Database Lab. Chapter Mapping Process Level of abstraction Nonprocedural access Navigational access Record & access path management Propagation control File management Pages Objects Relations File scan, Index scan Records, Access paths Files, Blocks Description SQL level interface Access path ñ procedural interface / Page Block mapping ñ ó ÿ block interface (OS) Yeungnam University, Database Lab. Chapter 1-13

8 2.2 The Storage Hierarchy Host Computer AP 1 AP 2 AP n Temporary Log Supports Transaction UNDO, Global UNDO, Partial REDO DBMS code Log Buffer DB Buffer Physical Copy of the Database Archive Log Supports Global REDO Archive Copy of the Database Yeungnam University, Database Lab. Chapter Different Views of a Database Current Database DBMS (DB buffer, Disk version of DB) Materialized Database DB There is no buffer. Recovery = Materialized DB Current DB Physical Database Page image disk block block ð. Example: indirect page allocation Yeungnam University, Database Lab. Chapter 1-15

9 Types of Update Operations Modification of Page Contents at Normal Processing DB buffer Current DB WritingaModifiedPagetoaBlockonDisk physical database ó ( ) memory ÿ, volatile Propagation ÿ materialized DB update-in-place: write operation = propagation shadow paging: write operation propagation Yeungnam University, Database Lab. Chapter Mapping Concept for Updates Page Allocation Principles Direct (update-in-place) Propagation is interruptable by a system crash. Indirect (shadow paging) Propagation is not interruptable by a system crash. System Buffer Page i read write after modification Disk Block i System Buffer Page i read write after modification Disk Block i Block j Yeungnam University, Database Lab. Chapter 1-17

10 Two Types of Propagation ATOMIC Supporting all or nothing properties for propagation Maintaining two versions of a page (by chained IO) 1: time complexity & space overhead 2: longer recovery time ATOMIC Pages are written to blocks by an update-in-place policy. Propagation is vulnerable to system crashes. Efficient Atomic Propagation? Using a small amount of nonvolatile RAM. Yeungnam University, Database Lab. Chapter Crash Recovery State of Database after Crash Types of Log Information to Support Recovery Action Checkpoint Example of Logging and Recovery Components Evaluation of Logging and Recovery Concepts Yeungnam University, Database Lab. Chapter 1-19

11 3.1 State of Database After Crash ATOMIC Propagation Nothing is known about the state of the materialized database. It must be characterized as chaotic. Crash. ATOMIC Propagation ÿ propagation. Access path level (level 3) ñ ATOMIC Propagation Yeungnam University, Database Lab. Chapter Types of Log Information REDO Log & UNDO Log REDO Log Crash ÿ DB þ Transaction Commit Action UNDO Log þ DB ÿ, Crash ÿ Transaction Buffer Management Dependencies between Buffer Manager & Recovery : STEAL, STEAL : FORCE, FORCE Yeungnam University, Database Lab. Chapter 1-21

12 Dependencies between Buffer Manager and Recovery Components STEAL STEAL FORCE UNDO FORCE UNDO REDO REDO Yeungnam University, Database Lab. Chapter 1-22 Classification of Log Data (1) Criteria for Classifying Log Data Type of objects to be logged Physical logging: Bit pattern Logical logging: Operator argument Type of information to be logged State: DB Transition: ( ) Logical Physical State Before images After images Transition Actions(DML statements) EXOR differences Yeungnam University, Database Lab. Chapter 1-23

13 Classification of Log Data (2) Physical State Logging on Page Level ÿ (before & after image) Size of log record Physical Transition Logging on Page Level þ image ó XOR( ) old log = new (REDO), new log = old(undo) image REDO UNDO ÿ, XOR ó 0 compression Yeungnam University, Database Lab. Chapter 1-24 Classification of Log Data (3) Physical State Logging on Access Path Level ÿ logging : record level logging : Reduce the size log record Transition Logging on Access Path Level Physical transition logging Physical state logging on access path level ú : (operator image) vs. (address image) Logical transition logging Access path level ó Record-Oriented level Yeungnam University, Database Lab. Chapter 1-25

14 Classification of Log Data (4) Logical Logging on the Record-Oriented Level (Insert, Update, Delete) Record-ID (OLD/NEW) Value REDO: DML, UNDO: Inverse DML ú ATOMIC propagation Logging technique Level no. Expenses during normal processing Expenses for recovery actions Physical state 2 High Low Physical transition 2 Medium Low Physical state 3 Low Low Logical transition 4 Very Low Medium Yeungnam University, Database Lab. Chapter Checkpoints Problems at System Restart? ó DB Checkpoint Steps <Begin-Checkpoint> ÿ <End-Checkpoint> Transaction-Oriented Checkpoints Transaction-Consistent Checkpoints Action-Consistent Checkpoints Fuzzy Checkpoints Yeungnam University, Database Lab. Chapter 1-27

15 Transaction-Oriented Checkpoints ä FORCE ú EOT Begin-Checkpoint Begin-Checkpoint òò Hot spot ó T1 T2 T3 Transaction-Oriented Checkpoints C(T1) C(T2) System Crash Yeungnam University, Database Lab. Chapter 1-28 Transaction-Consistent Checkpoints ä Global checkpoint transaction-consistent DB Checkpoint signal, checkpoint T1 Checkpoint Signal T2 Checkpoint Cn Generated Cn-1 T3 Processing Delay for New Transactions Cn T4 System Crash Yeungnam University, Database Lab. Chapter 1-29

16 Action-Consistent Checkpoints ä TCC : Long processing delay ñ action level ó delay! Global UNDO ó TCC T1 Checkpoint Signal Checkpoint Cn Generated T4 T2 T5 T3 T7 T6 Cn-1 Processing Delay for Actions System Crash Yeungnam University, Database Lab. Chapter 1-30 Cn Fuzzy Checkpoints ä TCC ACC Checkpoint buffer dirty page Checkpoint ú Checkpoint data! Log record Transaction table Dirty page table ó dirty page ÿ, recovery Hot spot page!! Clipping count Fuzzy checkpoint ACC Yeungnam University, Database Lab. Chapter 1-31

17 3.4 Example of Logging and Recovery Components Combination of recovery concepts Propagation Strategy Page Replacement EOT Processing Checkpoint Scheme Example of Implementation Technique Atomic, Steal, Force, TOC Atomic, Steal, Force, TCC Atomic, Steal, Force, ACC Yeungnam University, Database Lab. Chapter 1-32 Atomic, Steal, Force, TOC Force REDO log temporary log ÿ. Archive log. Steal UNDO log. Atomic EOT processing interruptable Example of Implementation Technique T: A A, B B EOT processing of T Write UNDO log for A and B (in case Force fails) Propagate A and B Write REDO log for A and B to the archive log Discard the UNDO entries for A and B WriteaCOMMITlog Yeungnam University, Database Lab. Chapter 1-33

18 Atomic, Steal, Force, TCC Steal large main memory Steal Force, main memory DB Normal transaction processing disk write REDO log shutdown data (TCC) High performance transaction processing Yeungnam University, Database Lab. Chapter 1-34 Atomic, Steal, Force, ACC System R s Shadow-Page Mechanism Shadow page : Logical page LP1 P1 LP1, P1 P1: shadow version, P1 : current version Page Map: (LP1, P1, P1 ), P1 ó in-place update ACC Checkpoint shadow version discard Current version shadow version Steal & Force, UNDO log REDO log Yeungnam University, Database Lab. Chapter 1-35

19 3.5 Evaluation of Logging and Recovery Concepts propagation strategy ATOMIC ATOMIC buffer replacement STEAL STEAL STEAL STEAL EOT processing FORCE FORCE FORCE FORCE FORCE FORCE FORCE FORCE checkpoint type TOC TCC ACC Fuzzy TOC TCC Fuzzy TOC TCC ACC TOC TCC materialized DB state after system failure DC DC DC DC DC DC DC TC TC AC TC TC cost of transaction UNDO cost of partial REDO at restart cost of global UNDO at restart overhead during normal processing frequency of checkpoints checkpoint cost Yeungnam University, Database Lab. Chapter Archive Recovery Two Ways of DB Recovery at Failure Physical Database Temporary Log File Archive Copy Archive Log Scenario for Archive Recovery Archive Copy Two extreme cases On-line dump of consistent version of DB: expensive If archive copy is too old, recovery should redo too much. Two solutions Fuzzy dump: Copy the DB on the fly Incremental dump: ÿ off-line process dump Yeungnam University, Database Lab. Chapter 1-37

20 Scenario for Archive Recovery Normal DBMS Execution Updating the Archive Copy by an Independent Process Archive Copy Generated Off-line Generation n Archive Log DB Archive Copy Generated during Normal Execution ( Fuzzy Dump ) Incremental Dump Archive Copy (Generation n1) Supplement the Archive Version by Latest Increments (2) REDO Changes up to the Most Recent Transaction Consistent State Archive Recovery DB Copy Most Recent Archive Version (1) Yeungnam University, Database Lab. Chapter 1-38