Database Systems (資料庫系統) Crash Recovery Chapter 18

(1)

Database Systems

( 資料庫系統 )

January 2, 2006

Last Day of Class 

(2)

Announcements

• Final exam covers chapters

[10,11,12,13,14,16,17.1~17.4, 18]

• The exam will be on 1/9/2006 9:10~12:00

– Place: CSIE 103/105 – Closed book

– No calculator

• Assignment #6 is due on 1/4 next Wed.

• Practicum #4 is due on 1/11.

(3)

Crash Recovery

(4)

Overview

• Many types of failures:

– Transaction failure: bad input, data not found, etc. – System crash: bugs in OS, DBMS, loss of power, etc. – Disk failure: disk head crash

• Recovery manager

is called after a system crash to

restore DBMS to a consistent state before the crash.

• Recovery manager ensures two transaction properties:

– Atomicity: undo all actions of uncommitted transactions.

– Durability: actions of committed transactions survives failures. (redo their update actions if they have not been written to disks).

(5)

ARIES Overview

• Assume HW support:

– Log actions on an independent “crash-safe” storage

• What are SW problems?

– Results of uncommitted transactions may be written to disks → undo them

– Results of committed transactions may not be written to the disk → redo them

– Questions:

• What are the states of transactions at the time of crash? • What are the states of page (dirty?) at the time of the crash? • Where to start undo & redo?

(6)

ARIES General Approach

• Before crash:

– Log changes to DB

(WAL)

– Checkpoints

• After crash:

– Analysis phase

• Figure out states of [committed vs.

uncommitted]

transactions & pages [dirty or clean]

• After crash [continue]:

– Redo phase

• Repeat actions from uncommitted &

committed transactions [till the time of the crash]

– Undo phase

• Undo actions of uncommitted transactions

• Do we really need

“redo” & “undo”?

(7)

Three Phases of ARIES

Oldest log record of transactions active at crash

Smallest recLSN in dirty page table at e nd of Analysis

Most recent checkpoint

CRASH (end of log) A B C Log Analysis Redo Undo

(8)

Steal Policy

• ARIES is designed to work with a

steal, no-force

approach.

• Steal property:

– Can the changes made to an object O in the buffer pool by T1 be written to disk before T1 commits?

– If yes, we have a steal (T2 steals a frame from T1).

– Say T2 wants to bring a new page (Q), and buffer pool replace the frame

containing O.

T1 . . W(O) O Buffer Pool T2 . . R(Q) Disk write Read (Q)

(9)

Force Policy

• When T1 commits, do we ens

ure that all changes T1 has m

ade are immediately forced to

disk?

• If yes, we have a

force

appro

ach.

T1 . . W(O) . Commit O Buffer Pool Disk write

(10)

Steal, No-Force Write Policies

• ARIES can recover crashes from DB with steal &

no-force write policy:

– Modified pages may be written to disk before a transaction commits.

– Modified pages may not be written to disk after a transaction commits.

• No-steal & Force write policy leads to low DB

performance.

(11)

ARIES

• ARIES is a recovery

algorithm that can work

with a steal, no-force write

policy.

• ARIES is invoked after a

crash. This process is

called

restart

.

• ARIES maintains a history

of actions executed by

DBMS in a

log

.

– The log is stored on stable storage and must survive crashes. (Use RAID-1 Mirrored)

(12)

Three Phases in ARIES (1)

• Goal:

– Restore the DB to the state (buffer pool & disk) before the crash

– AND without effects from actions of active (uncommitted)

transactions.

• Analysis Phase:

– Identify active transaction at the time of crash

• T1 and T3.

– Identify dirty pages in the buffer pool: • P1, P3 and P5. LSN LOG 10 Update: T1 writes P5 20 Update: T2 writes P3 30 T2 commits 40 T2 ends 50 Update: T3 writes P1 60 Update: T3 writes P3 CRASH, RESTART

(13)

Three Phases in ARIES (1)

• Redo Phase:

– Repeat actions (active & committed), starting from a

redo point in the log, and restores the database state to what it was at the time of crash. LSN LOG 10 Update: T1 writes P5 20 Update: T2 writes P3 30 T2 commits 40 T2 ends 50 Update: T3 writes P1 60 Update: T3 writes P3 CRASH, RESTART

(14)

Three Phases in ARIES (2)

• Undo Phase:

– Undo actions of active

transactions in reverse-time order, so that DB reflects only the actions of committed

transactions.

• LSNs 70 -> 60 -> 50 -> 10

• Why in reserve-time order?

– Consider forward-time order. P5 will be restored to a value written by T1 (in LSN #10), rather than before it.

LSN LOG 10 Update: T1 writes P5 20 Update: T2 writes P3 30 T2 commits 40 T2 ends 50 Update: T3 writes P1 60 Update: T3 writes P3 70 Update: T3 writes P5 CRASH, RESTART

(15)

Three Principles of ARIES

• Write-Ahead Logging (WAL)

– Update to a DB object is first recorded in the log.

– The log record must be forced to a stable storage before the writing the DB object to disk.

• How is WAL different from Force-Write? – Forcing the log vs. data to disk.

• Repeating History During Redo

– On restart, redo the actions (recorded in the log) to bring the system back to the exact state at the time of crash. Then undo the actions of active (not committed) transactions.

• Logging Changes During Undo

(16)

How is ARIES explained?

• Describe the needed data structure for recovery

– WAL

– Data page

– Transaction table

– Dirty page table

– Checkpoints

• Describe the algorithm

– no crash during recovery

– crash during recovery

(17)

Log Structure

• Log contains history of actions

executed by the DBMS.

• A DB action is recorded in a

log

record.

• Log Tail

: most recent portion of the

log in main memory.

– It is periodically forced to stable storage.

– Aren’t all records in a log in stable

storage? No, only when writes to disk or commits.

• Log Sequence Number (LSN):

unique ID for each log record.

LSN LOG 10 Update: T1 writes P5 20 Update: T2 writes P3 30 T2 commits 40 T2 ends 50 Update: T3 writes P1 60 Update: T3 writes P3

(18)

Data Page

• PageLSN:

the LSN of the most

recent log record that made a c

hange to this page.

– Every page in the DB must have a pageLSN.

– What is P3’s pageLSN? • 60 or 20

– It is used in the Redo phase of the algorithm. LSN LOG 10 Update: T1 writes P5 20 Update: T2 writes P3 30 T2 commits 40 T2 ends 50 Update: T3 writes P1 60 Update: T3 writes P3

(19)

What Actions to Record Log?

• A log is written for each of the following actions:

– Updating a page: when a transaction writes a DB object, it write an update type record. It also updates pageLSN to this record’s LSN.

– Commit: when a transaction commits, it force-writes a commit type log record to stable storage.

– Abort: when a transaction is aborted, it writes an abort type log record.

– End: when a transaction is either aborted or committed, it writes an end-type log record.

– Undoing an update: when a transaction is rolled back (being

aborted, or crash recovery), it undoes the updates and it writes a compensation log record (CLR).

(20)

Log Record

prevLS

N transID Type PageID / Length Offset Before-image After-image Fields common to all log

records Additional fields for update log records

T1000 update P500 / 3 21 ABC DEF

T2000 update P600 / 3 41 HIJ KLM

T1000 update P500 / 3 20 GDE QRS

T1000 update P505 / 3 21 TUV WXY

PrevLSN: LSN of the previous log record in the same transaction. It forms a single linked list of log records going back in time. It will be used for recovery.

(21)

Compensation Log Record (CLR)

CLR is written when undoing an update (T1000 30) after an abort (or during crash recovery).

CLR records undoNextLSN, which is the LSN of the next log record that is to be undone for T1000, which is the prevLSN of log record

prevL

SN LSN transID Type PageID / Length Offset Before-image After-image

00 T1000 update P500 / 3 21 ABC DEF

10 T2000 update P600 / 3 41 HIJ KLM

20 T2000 update P500 / 3 20 GDE QRS

30 T1000 update P505 / 3 21 TUV WXY

40 T1000 abort 50 T1000 CLR / undo 30 P505 / 3 21 TUV pr ev LS N un do N ex tL S N

(22)

More on CLR

• If a system crash occurs after CLR log record is written

to stable storage, do we need to undo CLR actions?

– Why not?

– The reason is that a transaction must be rolled back, so CLR

undo action is required as a part of the recovery. So no need to undo it.

– Note that in restart, ARIES does redo before undo. So redo the recovery CLR actions.

• This bounds the number of CLRs in the log during restart

(even if you have repeated crashes).

(23)

Crash after Abort Example: No Need to

Undo An Undo Action

prevL

SN LSN transID Type PageID / Length Offset Before-image After-image

- 00 T1000 update P500 / 3 21 ABC DEF

00 10 T1000 update P505 / 3 21 TUV WXY

System Crash

20 T1000 CLR/undo 10 P505 / 3 21 TUV System Crash

30 T1000 CLR/undo 00 P500 / 3 21 ABC

(24)

Other Recovery-Related Structures

• Transaction Table

: one

entry for each active

(uncommitted)

transaction. Each

entry has

– Transaction ID

– lastLSN

: the last LSN

log record for this

transaction.

– How is it used? (In

Undo)

LSN Trans ID Type PageID / Length 00 T1000 update P500 / 3 10 T2000 update P600 / 3 20 T2000 update P500 / 3 30 T1000 update P505 / 3 pageID recLSN P500 00 P600 10 P505 30 transID lastLSN T1000 30 T2000 20

(25)

Other Recovery-Related Structures

• Dirty Page Table:

one

entry for each dirty

page (not written to

disk) in the buffer pool.

– recLSN: LSN of the first log record that caused this page to become dirty.

– How is it used? (In Redo)

LSN Trans ID Type PageID / Length 00 T1000 update P500 / 3 10 T2000 update P600 / 3 20 T2000 update P500 / 3 30 T1000 update P505 / 3 pageID recLSN P500 00 P600 10 P505 30 transID lastLSN T1000 30 T2000 20

(26)

Write-Ahead Log (WAL)

• Before writing a page (P) to disk, every update log

record that describes a change to P must be forced to

stable storage.

• A committed transaction forces its log records (including

the commit record) to stable storage.

• (Non-forced approach + WAL) vs. (Forced approach) at

Transaction Commit Time:

– Non-forced approach + WAL mean log records are written to stable storage, but not data records.

– Forced approach means data pages are written to disk. – Log records are smaller than data pages!

(27)

Checkpointing

• A checkpoint is a snapshot of DBMS state stored in

stable storage.

• Checkpointing in ARIES has three steps:

(1) write begin_checkpoint record to log

(2) write the state of transaction table and dirty page table + end_checkpoint record to log

(3) write a special master record containing LSN of begin_checkpoint log record.

• Why checkpointing?

– The restart process will look for the most recent checkpoint & start analysis from there.

(28)

Recovering from a System Crash

• Recovering will use information from

– Write-ahead log

• The most recent checkpoint

• Compensation Log Records

– undoNextLSN: the LSN of the next log record that is to be undone

– Transaction table

• active (not committed) transactions

• lastLSNs: the LSN of the most recent log record for this transaction. (analysis)

– Dirty page table

• dirty (not written to disk) pages

• recLSNs: LSN of the first log record that caused this page to become dirty

(29)

Analysis Phase

• It finds out three things:

– A point in the log to start REDO.

• Earliest update log that may not have been written to disk.

– Dirty pages in the buffer pool at the time of crash -> restore the dirty page table to the time of crash.

– Active transactions at time of crash for UNDO -> restore the transaction table to the time of crash.

(30)

Analysis Phase: Algorithm

1. Find the most recent begin_checkpoint log record.

2. Initialize transaction & dirty page tables from the ones

saved in the most recent checkpoint.

3. Scan forward the records from begin_checkpoint log

record to the end of the log. For each log record LSN,

update trans_tab and dirty_page_tab as follows:

– If we see an end log record for T, remove T from trans_tab. – If we see a log record for T’ not in trans_tab, add T’ in

trans_tab. If T’ is in the trans_tab, then set T’s lastLSN field to LSN.

– If we see an update/CLR log record for page P and P is not in the dirty page table, add P in dirty page table and set its

(31)

Analysis Phase: Example (1)

• After system crash,

both table are lost.

• No previous

checkpointing, initialize

tables to empty.

LSN TransID Type PageID 00 T1000 update P500 10 T2000 update P600 20 T2000 update P500 30 T1000 update P505 40 T2000 commit System Crash pageID recLSN transID lastLSN

(32)

Analysis Phase: Example (2)

• Scanning log 00:

– Add T1000 to transaction table.

– Add P500 to dirty page table.

LSN TransID Type PageID

00 T1000 update P500 10 T2000 update P600 20 T2000 update P500 30 T1000 update P505 40 T2000 commit System Crash pageID recLSN P500 00 transID lastLSN T1000 00

(33)

Analysis Phase: Example (3)

• Scanning log 10:

– Add T2000 to transaction table.

LSN TransID Type PageID 00 T1000 update P500 10 T2000 update P600 20 T2000 update P500 30 T1000 update P505 40 T2000 commit System Crash pageID recLSN P500 00 P600 10 transID lastLSN T1000 00 T2000 10

(34)

Analysis Phase: Example (4)

• Scanning log 20:

– Set lastLSN to 20

LSN TransID Type PageID 00 T1000 update P500 10 T2000 update P600 20 T2000 update P500 30 T1000 update P505 40 T2000 commit System Crash pageID recLSN P500 00 P600 10 transID lastLSN T1000 00 T2000 20

(35)

Analysis Phase: Example (5)

• Scanning log 30:

LSN TransID Type PageID 00 T1000 update P500 10 T2000 update P600 20 T2000 update P500 30 T1000 update P505 40 T2000 commit System Crash pageID recLSN P500 00 P600 10 P505 30 transID lastLSN T1000 30 T2000 20

(36)

Analysis Phase: Example (6)

• Scanning log 40:

– Remove T2000 from transaction table.

– We are done!

• The redo point starts at

00. • Why?

– P500 is the earliest log that may not have been written to disk before crash

.

• We have restored

transaction table & dirty

page table.

LSN TransID Type PageID 00 T1000 update P500 10 T2000 update P600 20 T2000 update P500 30 T1000 update P505 40 T2000 Commit System Crash pageID recLSN P500 00 P600 10 P505 30 transID lastLSN T1000 30 T2000 10

(37)

Redo Phase: Algorithm

• Scan forward from the redo point (LSN 00). For each

update/CLR-undo log record LSN, we need to perform

redo unless one of the conditions holds:

– The affected page is not in the dirty page table

• Changes to this page have been written to disk, because it is not dirty. So no need to redo.

– The affected page is in the dirty page table, but recLSN > LSN. • Update has also been written to disk, because the page’s recLSN

(oldest log record causing this page to be dirty) is after LSN. – pageLSN >= LSN

• Update has been written to disk, because a later update has been written (pageLSN = the most recent LSN to update the page on disk).

(38)

Redo Phase: Example (1)

• Scan forward from the redo

point (LSN 00).

• Assume that P600 has been written to disk.

– But it can still be in the dirty page table.

• Scanning 00:

– P500 is in the dirty page table.

– 00(recLSN) = 00 (LSN) – -10 (pageLSN) < 00 (LSN) – Redo 00

• Scanning 10:

LSN TransID Type PageID

00 T1000 update P500 10 T2000 update P600 (disk) 20 T2000 update P500 30 T1000 update P505 40 T2000 commit System Crash pageID recLSN P500 00 P600 10 P505 30 transID lastLSN T1000 30

(39)

Redo Phase: Example (2)

• Scanning 10:

– 10 (pageLSN) == 10 (LSN) – Do not redo 10

LSN TransID Type PageID 00 T1000 update P500 10 T2000 update P600 (disk) 20 T2000 update P500 30 T1000 update P505 40 T2000 commit System Crash pageID recLSN P500 00 P600 10 P505 30 transID lastLSN T1000 30

(40)

Undo Phase: Algorithm

• It scans backward in time from the end of the log.

• It needs to undo all actions from active (not committed)

transactions. They are also called

loser transactions

.

– Same as aborting them.

• Analysis phase gives the set of loser transactions, called

ToUndo

set.

• Repeatedly choose the record with the largest LSN value

in this set and processes it, until ToUndo is empty.

– If it is a CLR and undoNextLSN value is not null, use

undoNextLSN value in ToUndo. If undoNextLSN is null, this transaction is completely undo.

– If it is an update record, a CLR is written and restore the data record value to before-image. Use prevLSN value in ToUndo.

(41)

Undo Phase: Example (1)

• The only loser transaction is

T1000.

• ToUndo set is {T1000:30}

LSN TransID Type PageID 00 T1000 update P500 10 T2000 update P600 (disk) 20 T2000 update P500 30 T1000 update P505 40 T2000 commit System Crash pageID recLSN P500 00 P600 10 P505 30 transID lastLSN T1000 30

(42)

Undo Phase: Example (2)

• The only loser transaction is

T1000.

• ToUndo set is {T1000:30}

• Undoing LSN:30

– Write CLR:undo record log. – ToUndo becomes {T1000:00}

• Undoing LSN:00

– Write CLR:undo record log. – ToUndo becomes null.

– We are done.

LSN TransID Type PageID 00 T1000 update P500 10 T2000 update P600 (disk) 20 T2000 update P500 30 T1000 update P505 40 T2000 commit System Crash 50 T1000 CLR:undo:30 P505 60 T1000 CLR:undo:00 P500 un do N ex tL S N

(43)

Crashes During Restart (1)

• After T1 aborts, undo actions

from T1.

• Undo LSN #10: write

CLR:undo record log for LSN #10. • Dirty pages: – P1 (recLSN=50) – P3(20) – P5(10). • Loser transaction: – T2(lastLSN=60) – T3(50)

• Redo phases starts at 10. • Undo LSN #60. LSN LOG 00, 05 begin_checkpoint, end_checkpoint 10 Update: T1 write P5 20 Update: T2 writes P3 30 T1 aborts 40, 45 CLR: Undo T1 LSN 10, T1 end 50 Update: T3 writes P1 60 Update: T2 writes P5 CRASH, RESTART 70 CLR: Undo T2 LSN 60 80, 85 CLR: Undo T3 LSN 50, T3 end CRASH, RESTART un do N ex tL S N

(44)

Crashes During Restart (2)

• Undo LSN #50: write CLR:

undo record log.

– T3 is completely undone.

– LSN #85,80,70 are written to stable storage.

• Crash occurs after restart. – Loser transaction is T2.

– Read LSN #70, set ToUndo to #20.

– Undo #20: write another CLR. – Done LSN LOG 00, 05 begin_checkpoint, end_checkpoint 10 Update: T1 write P5 20 Update: T2 writes P3 30 T1 aborts 40, 45 CLR: Undo T1 LSN 10, T1 end 50 Update: T3 writes P1 60 Update: T2 writes P5 CRASH, RESTART 70 CLR: Undo T2 LSN 60 80, 85 CLR: Undo T3 LSN 50, T3 end CRASH, RESTART un do N ex tL S N

(45)

Crashes During Analysis/Redo

• Restart from Analysis phases. Why?

– Crashes during analysis: all work is lost (nothing written to disk or log).

– Crashes during redo: some data records may be written to disk prior to a crash, but safe to start with analysis. Some of the update log that was redone the first time will not be redone a second time.