Recovery

Disk & Buffer Management

A naive policy implements the following:

• No Steal Policy

  Don’t allow buffer-pool frames with uncommited updates to be replace (pin those pages)

  • No need for UNDO logging

  • Cause bad performance taking up buffer

• Force Policy

  Force all updates onto the disk before commit.

  • No need for REDO logging

  • Bad performance (random I/O writes for commits)

What we actually want is the opposite.

• No Force Policy

  Flush as little as possible to a REDO log prior to commit. This enforces durability (if commited -> REDO).

• Steal Policy

  Remember old values of flushed pages to an UNDO log. This enforaces atomicity (if abort -> UNDO).

Log Records

An ordered list of records that allow REDO/UNDO

• Transaction ID

• Page ID

• Offset

• Length

• Old data

• New data

ARIES Protocol

This log record follows the ARIES protocol.

• prevLSN: LSN of the previous log record written by the specific transaction.

• Transaction Table A table of active transactions

• Dirty Page Table A table of dirty pages stores information for UNDO.

  • recLSN: LSN of the first action that dirtied the page.

• Log record types:

  • Update, Commit, Abort

  • Checkpoint (for log maintenance)

  • Componensation Log Records (CLRs, for UNDO actions)

  • End (end of commit or abort)

• Abort and Checkpoint When abort and checkpointing we need to implement UNDO. Before we start the UNDO we need

  • get lastLSN

  • write abort log

  • follow the chain of log records using prevLSN

  • Componensation Log Record (CLR): A log to insure crashes while undoing.

    • Contains REDO info

    • UNDO persists.

  • Write end log record.

ARIES Recovery Phases

During a crash the following is ran

1. Analysis (Checkpoint)

   If checkpoint exists, figure out which transaction was commited since checkpoint and which failed.

   • End record

     • Remove transaction from table

   • Update record

     • If the page is not in the dirty page table, add to it. Set recLSN=LSN

   • Else

     • Add transaction to table. Set lastLSN=LSN

     • change transaction status on commit.

2. REDO

   Redo all actions since the smallest recLSN in the dirty page.

   • Start at the smallest recLSN (oldest update log that dirtied the page).

   • For each CLR, REDO the aciton unless either

     • Affected page isn’t in the dirty page table

     • Affected page is in the dirty page table but has recLSN > LSN

     • pageLSN >= LSN

   • Apply REDO:

     • Reapply logged action

     • Set pageLSN=LSN

3. UNDO Undo the effects of the failed transaction.

   • Start with the transaction with the largest lastLSN. For each transaction lastLSN (stored in a list of undo),

     • If the lastLSN log is a CLR

       • Add the nextLSN to the undo list if it isn’t null.

       • Otherwise write an end record.

     • Otherwise the record is an update

       • Write a CLR for the undo in the log.

       • Perform the undo.

       • Add the prevLSN to the list of undo

SQL Commands

• BEGIN

  Start a transaction that won’t commit until the COMMIT is ran.

• COMMIT

  Commit the transaction if all transactions above (until the BEGIN) suceeded.

• ROLLBACK

  Rollback the transaction.

• SAVEPOINT <name>

  Save the current.

• RELEASE SAVEPOINT <name>

  Delete the savepoint.

• ROLLBACK TO SAVEPOINT <name>

  Rollback to the savepoint.

Why Database Crash?

From the most common to least:

1. Operation Error

   The user caused the error.

2. Configuration Error

   Error on requirements (e.g., disk space, file permission, etc).

3. Software Failure

   Bugs (DBMS bugs, security flaws, OS bugs, etc).

4. Hardware Failure