Concurrency Control and Recovery

Zachary G. IvesUniversity of Pennsylvania

CIS 650 – Implementing Data Management Systems

February 7, 2005

Some content on recovery courtesy Hellerstein, Ramakrishnan, Gehrke

2

Administrivia

No normal office hour this week (out of town)

Upcoming talks of interest: George Candea, Stanford, recovery-oriented

computing, 2/24 Mike Swift, U Wash., restartable OS device drivers, 3/1 Andrew Whitaker, U Wash., paravirtualization, 3/15 Sihem Amer-Yahia, AT&T Research, text queries over

XML, 3/17 Krishna Gummadi, U Wash., analysis of P2P systems,

3/29 Muthian Sivathanu, U Wisc., speeding disk access, 3/31

3

Today’s Trivia Question

4

Recall the Fundamental Concepts of Updatable DBMSs

Transactions as atomic units of operation always commit or abort can be terminated and restarted by the DBMS – an essential

property typically logged, restartable, recoverable

ACID properties: Atomicity: transactions may abort (“rollback”) due to error

or deadlock (Mohan+) Consistency: guarantee of consistency between transactions Isolation: guarantees serializability of schedules (Gray+,

Kung & Rob.) Durability: guarantees recovery if DBMS stops running

(Mohan+)

5

Serializability and Concurrent Deposits

Deposit 1 Deposit 2read(X.bal) read(X.bal)X.bal := X.bal + $50 X.bal:= X.bal + $10write(X.bal) write(X.bal)

6

Violations of Serializability

Dirty data: data written by an uncommitted transaction; a dirty read is a read of dirty data (WR conflict)

Unrepeatable read: a transaction reads the same data item twice and gets different values (RW conflict)

Phantom problem: a transaction retrieves a collection of tuples twice and sees different results

7

Two Approaches to Serializability (or other Consistency Models)

Locking – a “pessimistic” strategy First paper (Gray et al.): hierarchical locking,

plus ways of compromising serializability for performance

Optimistic concurrency control Second paper (Kung & Robinson): allow writes

by each session in parallel, then try to substitute them in (or reapply to merge)

8

Locking

A “lock manager” grants and releases locks on objects

Two basic types of locks: Shared locks (read locks): allow other shared locks to be

granted on the same item Exclusive locks (write locks): do not coexist with any other

locks

Generally granted in two phase locking (2PL) model: Growing phase: locks are granted Shrinking phase: locks are released (no new locks granted) Well-formed, two-phase locking guarantees serializability Strict 2PL: shrinking phase is at the end of the transaction

(Note that deadlocks are possible!)

9

Gray et al.: Granularity of Locks

For performance and concurrency, want different levels of lock granularity, i.e., a hierarchy of locks: database extent table page row attribute

But a problem arises: What if T1 S-locks a row and T2 wants to X-lock a table? How do we easily check whether we should give a lock to

T2?

10

Intention Locks

Two basic types: Intention to Share (IS): a descendant item will be

locked with a share lock Intention Exclusive lock: a descendant item will be

locked with an exclusive lock Locks are granted top-down, released bottom-up

T1 grabs IS lock on table, page; S lock on row T2 can’t get X-lock on table until T1 is done

But T3 can get an IS or S lock on the table

11

Lock Compatibility Matrix

IS IX S SIX X

IS Y Y Y Y N

IX Y Y N N N

S Y N Y N N

SIX Y N N N N

X N N N N N

12

Lock Implementation

Maintain as a hash table based on items to lock Lock/unlock are atomic operations in critical

sections First-come, first-served queue for each locked

object All adjacent, compatible items are a compatible group The group’s mode is the most restrictive of its members

What if a transaction wants to convert (upgrade) its lock? Should we send it to the back of the queue? No – will almost assuredly deadlock! Handle conversions immediately after the current group

13

Degrees of Consistency

Full locking, guaranteeing serializability, is generally very expensive

So they propose several degrees of consistency as a compromise (these are roughly the SQL isolation levels): Degree 0: T doesn’t overwrite dirty data of other

transactions Degree 1: above, plus T does not commit writes before

EOT Degree 2: above, plus T doesn’t read dirty data Degree 3: above, plus other transactions don’t dirty any

data T read

14

Degrees and Locking

Degree 0: short write locks on updated items

Degree 1: long write locks on updated items

Degree 2: long write locks on updated items, short read locks on read items

Degree 3: long write and read locks

Does Degree 3 prevent phantoms? If not, how do we fix this?

15

What If We Don’t Want to Lock?

Conflicts may be very uncommon – so why incur the overhead of locking? Typically hundreds of instructions for every

lock/unlock Examples: read-mostly DBs; large DB with few

collisions; append-mostly; hierarchical data

Kung & Robinson – break lock into three phases: Read – and write to private copy of each page (i.e.,

copy-on-write) Validation – make sure no conflicts between

transactions Write – swap the private copies in for the public ones

16

Validation

Goal: guarantee that only serializable schedules result in merging Ti and Tj writes

Approach: find an equivalent serializable schedule: Assign each transaction a number Ensure equivalent serializable schedule as follows:

If TN(Ti) < TN(Tj) then we must satisfy one of:

1. Ti finishes writing before Tj starts reading (serial)

2. WS(Ti) disjoint from RS(Tj) and Ti finishes writing before Tj writes

3. WS(Ti) disjoint from RS(Tj) and WS(Ti) disjoint from WS(Tj), and Ti finishes read phase before Tj completes its read phase

17

Why Does This Work?

Condition 1 – obvious since it’s serial Condition 2:

No W-R conflicts since disjoint In all R-W conflicts, Ti precedes Tj since Ti reads before it

writes (and that’s before Tj)

In all W-W conflicts, Ti precedes Tj

Condition 3: No W-R conflicts since disjoint No W-W conflicts since disjoint In all R-W conflicts, Ti precedes Tj since Ti reads before it

writes (and that’s before Tj)

18

The Achilles Heel

How do we assign TNs? Not optimistically – they get assigned at the

end of read phase Note that we need to maintain all of the read

and write sets for transactions that are going on concurrently – long-lived read phases cause difficulty here Solution: bound buffer, abort and restart

transactions when out of space Drawback: starvation – need to solve by locking the

whole DB!

19

Serial Validation

Simple: writes won’t be interleaved, so test1. Ti finishes writing before Tj starts reading (serial)2. WS(Ti) disjoint from RS(Tj) and Ti finishes writing before Tj

writes

Put in critical section: Get TN Test 1 and 2 for everyone up to TN Write

Long critical section limits parallelism of validation, so can optimize: Outside critical section, get a TN and validate up to there Before write, in critical section, get new TN, validate up to

that, writeReads: no need for TN – just validate up to highest TN

at end of read phase (no critical section)

20

Parallel Validation

For allowing interleaved writes Save active transactions (finished reading, not

writing) Abort if intersect current read/write set Validate:

CRIT: Get TN; copy active set; add self to active set Check (1), (2) against everything from start to finish Check (3) against all active set If OK, write CRIT: Increment TN counter, remove self from active

Drawback: might conflict in condition (3) with someone who gets aborted

21

Who’s the Top Dog?Optimistic vs. Non-Optimistic

Drawbacks of the optimistic approach: Generally requires some sort of global state, e.g., TN

counter If there’s a conflict, requires abort and full restart

Study by Agrawal et al. comparing optimistic vs. locking: Need load control with low resources Locking is better with moderate resources Optimistic is better with infinite or high resources

Both of these provided isolation; transactions and policies ensure consistency – what about atomicity, durability?

22

Rollback and Recovery

The Recovery Manager provides:Atomicity:

Transactions may abort (“rollback”) to start or to a “savepoint”.

Durability: What if DBMS stops running? (Causes?)

Desired behavior after system restarts:– T1, T2 & T3 should

be durable– T4 & T5 should be

aborted (effects not seen)

crash!T1T2T3T4T5

Assumptions in Recovery SchemesWe’re using concurrency control via locks

Strict 2PL at least at the page, possibly record level

Updates are happening “in place” No shadow pages: data is overwritten on (deleted from)

the disk

ARIES: Algorithm for Recovery and Isolation Exploiting Semantics Attempts to provide a simple, systematic simple scheme

to guarantee atomicity & durability with good performance

Let’s begin with some of the issues faced by any DBMS recovery scheme…

Managing Pages in the Buffer PoolBuffer pool is finite, so…

Q: How do we guarantee durability of committed data?

A: Need policy on what happens when a transaction completes, what transactions can do to get more pages

Force write of buffer pages to disk at commit? Provides durability But poor response time

Steal buffer-pool frames from uncommited Xacts? If not, poor throughput If so, how can we ensure

atomicity?

Force

No Force

No Steal Steal

Trivial

Desired

More on Steal and Force

STEAL (why enforcing Atomicity is hard) To steal frame F: Current page in F (say P) is

written to disk; some Xact holds lock on P What if the Xact with the lock on P aborts? Must remember the old value of P at steal time (to

support UNDOing the write to page P)

NO FORCE (why enforcing Durability is hard) What if system crashes before a modified page is

written to disk? Write as little as possible, in a convenient place,

at commit time, to support REDOing modifications

Basic Idea: Logging

Record REDO and UNDO information, for every update, in a log Sequential writes to log (put it on a separate

disk) Minimal info (diff) written to log, so multiple

updates fit in a single log page

Log: An ordered list of REDO/UNDO actions Log record contains:

<XID, pageID, offset, length, old data, new data>

and additional control info (which we’ll see soon)

Write-Ahead Logging (WAL)

The Write-Ahead Logging Protocol:1. Force the log record for an update before the

corresponding data page gets to disk Guarantees Atomicity

2. Write all log records for a Xact before commit• Guarantees Durability (can always rebuild from the log)

Is there a systematic way of doing logging (and recovery!)? The ARIES family of algorithms

WAL & the Log

Each log record has a unique Log Sequence Number (LSN) LSNs always increase

Each data page contains a pageLSN The LSN of the most recent log record

for an update to that page

System keeps track of flushedLSN The max LSN flushed so far

WAL: Before a page is written, pageLSN flushedLSN

LSNs

DB

pageLSNs

RAM

flushedLSN

pageLSN

Log recordsflushed to disk

“Log tail” in RAM

Log Records

Possible log record types: Update Commit Abort End (signifies end of

commit or abort) Compensation Log

Records (CLRs) To log UNDO actions

prevLSNXIDtype

lengthpageID

offsetbefore-imageafter-image

LogRecord fields:

updaterecordsonly

Other Log-Related State

Transaction Table: One entry per active Xact Contains XID, status

(running/commited/aborted), and lastLSN

Dirty Page Table: One entry per dirty page in buffer pool Contains recLSN – the LSN of the log record

which first caused the page to be dirty

Normal Execution of an Xact

Series of reads & writes, followed by commit or abortWe will assume that write is atomic on disk

In practice, additional details to deal with non-atomic writes

Strict 2PLSTEAL, NO-FORCE buffer management, with

Write-Ahead Logging

Checkpointing

Periodically, the DBMS creates a checkpoint Minimizes recovery time in the event of a system crash Write to log:

begin_checkpoint record: when checkpoint began end_checkpoint record: current Xact table and dirty page table A ‘fuzzy checkpoint’:

Other Xacts continue to run; so these tables accurate only as of the time of the begin_checkpoint record

No attempt to force dirty pages to disk; effectiveness of checkpoint limited by oldest unwritten change to a dirty page. (So it’s a good idea to periodically flush dirty pages to disk!)

Store LSN of checkpoint record in a safe place (master record)

The Big Picture: What’s Stored Where

DB

Data pageseachwith apageLSN

Xact TablelastLSNstatus

Dirty Page TablerecLSN

flushedLSN

RAM

prevLSNXIDtype

lengthpageID

offsetbefore-imageafter-image

LogRecords

LOG

master record

Simple Transaction Abort

For now, consider an explicit abort of a Xact No crash involved

We want to “play back” the log in reverse order, UNDOing updates Get lastLSN of Xact from Xact table Can follow chain of log records backward via the

prevLSN field Before starting UNDO, write an Abort log record

For recovering from crash during UNDO!

Abort, cont.

To perform UNDO, must have a lock on data!No problem – no one else can be locking it

Before restoring old value of a page, write a CLR: You continue logging while you UNDO!! CLR has one extra field: undoNextLSN

Points to the next LSN to undo (i.e. the prevLSN of the record we’re currently undoing).

CLRs never Undone (but they might be Redone when repeating history: guarantees Atomicity!)

At end of UNDO, write an “end” log record.

Transaction Commit

Write commit record to log All log records up to Xact’s lastLSN are

flushed Guarantees that flushedLSN lastLSN Note that log flushes are sequential,

synchronous writes to disk Many log records per log page

Commit() returns Write end record to log

Crash Recovery: Big Picture

Start from a checkpoint (found via master record)

Three phases:1. Figure out which Xacts

committed since checkpoint, which failed (Analysis)

2. REDO all actions– (repeat history)

3. UNDO effects of failed Xacts

Oldest log rec. of Xact active at crash

Smallest recLSN in dirty page table after Analysis

Last chkpt

CRASH

A R U

Recovery: The Analysis PhaseReconstruct state at checkpoint

via end_checkpoint record

Scan log forward from checkpoint End record: Remove Xact from Xact table Other records: Add Xact to Xact table, set

lastLSN=LSN, change Xact status on commit Update record: If P not in Dirty Page Table,

Add P to D.P.T., set its recLSN=LSN

Recovery: The REDO PhaseWe repeat history to reconstruct state at crash:

Reapply all updates (even of aborted Xacts!), redo CLRs

Scan forward from log rec containing smallest recLSN in D.P.T. For each CLR or update log rec LSN, REDO the action unless:

Affected page is not in the Dirty Page Table, or Affected page is in D.P.T., but has recLSN > LSN, or pageLSN (in DB) LSN

To REDO an action: Reapply logged action Set pageLSN to LSN. No additional logging!

Recovery: The UNDO Phase

ToUndo = { l | l a lastLSN of a “loser” Xact}Repeat:

Choose largest LSN among ToUndo. If this LSN is a CLR and undoNextLSN==NULL

Write an End record for this Xact. If this LSN is a CLR, and undoNextLSN != NULL

Add undoNextLSN to ToUndo (Q: what happens to other CLRs?)

Else this LSN is an update. Undo the update, write a CLR, add prevLSN to ToUndo.

Until ToUndo is empty.

Example of Recovery

begin_checkpoint

end_checkpoint

update: T1 writes P5

update T2 writes P3

T1 abort

CLR: Undo T1 LSN 10

T1 End

update: T3 writes P1

update: T2 writes P5

CRASH, RESTART

LSN LOG

00

05

10

20

30

40

45

50

60

Xact TablelastLSNstatus

Dirty Page TablerecLSN

flushedLSN

ToUndo

prevLSNs

RAM

Example: Crash During Restart

begin_checkpoint, end_checkpoint

update: T1 writes P5

update T2 writes P3

T1 abort

CLR: Undo T1 LSN 10, T1 End

update: T3 writes P1

update: T2 writes P5

CRASH, RESTART

CLR: Undo T2 LSN 60

CLR: Undo T3 LSN 50, T3 end

CRASH, RESTART

CLR: Undo T2 LSN 20, T2 end

LSN LOG00,05

10

20

30

40,45

50

60

70

80,85

90

Xact TablelastLSNstatus

Dirty Page TablerecLSN

flushedLSN

ToUndo

undonextLSN

RAM

Additional Crash Issues

What happens if system crashes during Analysis? During REDO?

How do you limit the amount of work in REDO? Flush asynchronously in the background. Watch “hot spots”!

How do you limit the amount of work in UNDO? Avoid long-running Xacts.

Summary of Logging/Recovery

Recovery Manager guarantees Atomicity & Durability

Use WAL to allow STEAL/NO-FORCE w/o sacrificing correctness

LSNs identify log records; linked into backwards chains per transaction (via prevLSN)

pageLSN allows comparison of data page and log records

Summary, Continued

Checkpointing: A quick way to limit the amount of log to scan on recovery.

Recovery works in 3 phases: Analysis: Forward from checkpoint Redo: Forward from oldest recLSN Undo: Backward from end to first LSN of oldest

Xact alive at crash

Upon Undo, write CLRs Redo “repeats history”: Simplifies the

logic!