Redo Log

Format

Example

By the way, for a sparse page, there are logs(MLOG_COMP_LIST_START ,MLOG_COMP_LIST_END and etc) that store a list of operation.

Mini-Transaction

Atomic operations

• Group Redo Log: make multi redo log atomic.
• Single Redo Log

Group Redo Log

Single Redo Log

Mini-Transaction

Mini-Transaction is a atomic operation, including a group redo log or a Single atomic redo log.

Relationship

Redo Log Buffer

Redo Log Block

Write Redo Log to Redo Log Buffer

Concurrency of Redo Log Buffer

When write to disk

• Not enough log buffer space.
• transaction is committing.
• write redo log before writing dirty.
• background thread write periodically.
• power off
• checkpoint update.

Redo Log File

File Group

File Format

Log File header

Log File checkpoint

Flags

• Long Sequence Number
• Flushed_to_disk_lsn
• oldest_modification
• checkpoint_lsn

Variables	Meaning	How to update
Long Sequence Number	Next redo start number.	do operation
Flushed_to_disk_lsn	lsn that has been written to disk	write redo log page in redo log bugger to disk
oldest_modification	oldest lsn of dirty page.	periodically flush
checkpoint_lsn	checkpoint of transaction	execute checkpoint: update checkpoint_lst and write checkpoint to log file

rank

checkpoint_lsn <= oldest_modification <= flushed_to_disk_lsn <= Long Sequence Number

How to update checkpoint_lsn : checkpoint

• read current oldest_modification and set it to checkpoint_lsn
• write current checkpoint’s information to redo log file(actually the first log filie’s checkpoint area).

Crash Recovery

Recovery Range

Look at rank picture.

Recovery range is from checkpoint_lsn to flushed_to_disk_lsn, namely part 4.

but part 1 has been write to disk before crash, so it will be passed.

Start Point

get checkpoint_lsn from first log’s checkpoint area.

End Point

scan from checkpoint until log file’s LOG_BLOCK_HDR_DATA_LEN is not 512.

How to pass part 1 while recovering

There are a FIL_PAGE_LSN in redo log file header.

if this value is bigger than last checkpoint_lsn, meaning that it’s more updated, it will passed.

Optimization

While reading redo log file, arrange redo log by hash table. Key is redo log’s space ID and page number.

arrange them by writing order to keep original order.

If this page has been passed(namely has been written be crash), it will not wirte again.

• minimize disk I/O by hash table
• avoid duplicate write

Safety vs Concurency

innodb_flush_log_at_trx_commit when flush redo while committing.

• 0: not right now, depends on background thread.
• default 1: now. write to disk directly.
• 2: wriiten by buffer IO. live in System’s page buffer.