1. ARIES: Algorithm for Recovery and Isolation Exploiting Semantics

2. Outline
Durability
Atomicity
ARIES
Recovery in Postgres

3. Durability Schedule: Begin T1 Write(P1, x+1) Commit Buffer Volatile
Expensive
P1: x=42
Non-volatile

4. Make commits cheap again
Schedule:
Begin T1
Write(P1, x+1)
Commit
LSN TransID Type Pageid RedoInfo
T update P x + 1
T commit
Log
tail:
LSN: log sequence number

5. Long live the commits
Stable storage: abstract storage for log that survives any crash
In practice, approximated by keeping multiple log copies

6. Durability revisited Schedule: Begin T1 Write(P1, x+1) Commit Buffer
Log tail
Volatile
P1: x=43
LSN 1,2
P1: ?
LSN 1,2
Non-volatile
Stable storage
Disk

7. Write-ahead logging aka WAL
Persist an action instead of a page it affects
Accumulate log entries in memory
Log entry: necessary info to redo an action
Write ahead the log tale to stable storage
in batches
Write ahead: write a log entry before the corresponding page
Log tail: collection of log entries in memory
Stable storage: the one that survives any crash
Commited transaction: the one with the „commit“ log record on stable storage

8. REDOing things pageLSN: 1 P1: x=43 LSN TransID Type Pageid RedoInfo
T update P x + 1
pageLSN: 1 P1: x=43
pageLSN: LSN of the last log record that modifed this page

9. REDO
LSNs
Crash
REDO
REDO log record if pageLSN < record LSN

10. REDO logs Provide durable and efficient commits Normal operation:
Employ WAL: flush all log records up to and including commit record of the tranasction being commited
update pageLSN
Recovery
Replay the log entry if its LSN > pageLSN

11. Outline
Durability
Atomicity
ARIES
Recovery in Postgres

12. Atomicity Schedule: Begin T2 Write(P1, x=43) Abort Buffer Volatile
Non-volatile

13. REDO recap
LSNs
Crash
REDO
We are here

14. Transaction Table (TT)
What to undo
Any tranasction without the commit record
Transaction Table (TT)
TransID lastLSN Status
T Aborted
lastLSN: LSN of the last log record on stable storage

15. How to undo LSN TransID Type Pageid RedoInfo 1 T2 update P1 43
UndoInfo prevLSN
null
prevLSN: LSN of the record to undo next

16. UNDO logs Analyze the log to build the Transaction Table
Undo the largest lastLSN by applying the UndoInfo
For this tranasction, insert prevLSN into TT
Log UNDOs in a special compensational log record (CLR)
CLRs make UNDOs idempotent
Same procedure applies for aborting a single tranasction

17. UNDO after REDO
LSNs
Crash
REDO TT
UNDO

18. Outline
Durability
Atomicity
ARIES
Recovery in Postgres

19. Algorithm for Recovery and Isolation Exploiting Semantics
ARIES
Algorithm for Recovery and Isolation Exploiting Semantics
LSNs
Crash
Analysis
Transaction
and Dirty Page
tables
REDO
UNDO

20. Distinctive features Uses write-ahead logging
Repeats entire history, including uncommited transactions
Logs UNDOs

21. OS and hardware support
Assumption
Support
Configure file system
Atomic page write
DRAM with ECC
CRC for log entries
Correct memory content
Log copies in distant
places
Stable storage for logs

22. OS and hardware support
Assumption
Support
1. Disable write-back caches
2. Synchronous commits
3. Use UPS units
Logs actually reach
stable storage

23. Outline
Durability
Atomicity
ARIES
Recovery in Postgres

24. Recovery in Postgres REDO log (Xlog) only
Since version 7.1 (2001)
No UNDO log, presumably due to MVCC
Transactional DDL
WAL for file operations
Different format of log for different purposes
Hot-stand-by, point-in-time recovery etc.
Requires complex setup
Hard to ensure durability
Non-trivial performance-durability trade-offs

25. References

26. References

27. Algorithms for Recovery and Isolation Exploiting Semantics.
ARIES
Algorithms for Recovery and Isolation Exploiting Semantics.
LSNs
Crash
Analysis
Transaction
and Dirty Page
tables
REDO
UNDO