MMDB:Log&Recovery
2009-09-17 14:27:50| 分类:
dbms
| 标签:
|举报
|字号大中小 订阅
Repeat history
Low Amount of disk I/O
- Redo logs are written to disk
- No undo logs to disk normally
Ping-Pong Checkpointing Scheme
- Write to alternate disk copies on successive checkpoints
resilient to crashes during checkpointing
Only relevant undo logs written to disk when checkpointing
- undo logs of transactions that are active at the end of the checkpoint
Fuzzy checkpoints
- Updates execute concurrently with checkpoints
e.g. No page latching during checkpoints
- Only dirty pages written to disk during checkpoints
Single pass recovery algorithm
Do not write logs for index.
- If crash occurs, the index will be rebuild
Extended Logging Support:
Logging support for high concurrency
- Multi-level logging
Physical undo during updates, logical undo afterward
- Post-commit actions
Actions that cannot be rolled back (e.g. freeing storage space)
Support for user-defined operation logging
Time stamp recovery
- Recover the database to a transaction-consistent point in the past
Turn redo logging off
- Support atomic transactions without requiring them to be durable
评论这张
转发至微博
转发至微博
评论