Modern techniques for transaction- oriented database recovery - PowerPoint PPT Presentation

Modern techniques for transaction- oriented database recovery Caetano Sauer

My pleas 1.Demand on-demand recovery 2.Make the log great again

Demand on-demand recovery

ARIES Recovery info B A B A Transactions Pages requiring requiring Buffer pool Database undo redo 1. Log analysis LOG 2. REDO scan 3. UNDO scan system mostly unavailable during three phases

ARIES restart Recovery info B A B A Transactions Pages requiring requiring Buffer pool Database undo redo 1. Log analysis LOG 2. REDO scan 3. UNDO scan system mostly unavailable during three phases 6

Instant recovery (post-failure) lock(f) Recovery info Running txn. fix(B) Transactions Pages B requiring requiring lock undo redo conflict! + locks Buffer pool Aborting txn. LOG ... (pre-failure) per-page log chains redo and undo on demand, without waiting for log scans same recovery actions, different schedule

Failure classes Failure class Loss Typical cause Response Transaction Single-transaction progress Deadlock, constraint violation Rollback System Server process (in-memory state) Software fault, power loss Restart Media Persistent database contents Hardware fault Restore Single page Local integrity Partial writes, wear-out Repair

Make the log great again

Throw away the database! Main memory Main memory write- read- optimized optimized Log DB ? Dual storage Single storage “... a DBMS is really two DBMSs, one managing the database as we know it and a second one managing the log.” Michael Stonebraker

The log as a partitioned B-tree A B A B A B P0 P1 P2 LSN domain x y z w like an LSM, but with page identifiers and their log records

The log as a partitioned B-tree (implementation) Partitioned log index sorted unsorted partitions partition A B A B A B page → LSN mapping log appends A B (txn. commit) P0 P1 P2 LSN domain x y z u w (sort low-water mark)

Merging Before: A B A B A B � x, y � � y, z � � z, w � merge After: A B A B � x, z � � z, w �

FineLine Write-ahead logging: FineLine: fetch Data Data structure structure read reorg. Log Database write index append WAL append Log records Log records volatile persistent volatile persistent

FineLine recovery Data structure Indexed log Log records volatile persistent

FineLine recovery fetch Indexed log volatile persistent

FineLine recovery fetch Indexed log volatile persistent

FineLine recovery fetch Indexed log volatile persistent

In-memory FineLine database fetch Indexed log append volatile persistent ➢ nodes recovered automatically during fetch ➢ volatile structures = in-memory database ➢ no undo, no dirty pages, no checkpoints, no offline log scans

In-memory DBMS databases Buffer mgmt. Log-structured Physiological storage logging

My pleas 1.Demand on-demand recovery 2.Make the log great again

Standing on the shoulders of Giants

Thank you! csauer@tableau.com