Pastiche: Making Backup Cheap

1
Pastiche Making Backup Cheap Easy

• Authors L. Cox, C. Murray, B. Noble
• Presenter David Spain

2
Premise

• Backup is expensive hard-to-manage and overly
  centralized
• File systems are only 53 full!

3
Central Technologies

• Pastry Peer-to-Peer self-administering network
  substrate
• Content-based Indexing finds similar chunks of
  data within files
• Convergent encryption hosts can use the
  encrypted representation w/o sharing keys

4
Some Terms

• Backup Buddy 5 of them
• Abstract used to approximate FS overlap
• Chunkstore FS for chunks!
• Skeleton Metadata on archive state

5
Pastry (2.1)

• NodeId uniformly distributed! (how?)
• Proximity metric network cost
• Per node
• leaf set closest nodes, 50/50 above below
• neighbor set proximal nodes, 50/50
• routing table log N hops expected
• Discovery protocol to join

6
Content Based Indexing (2.2)

• Anchors Rabin fingerprints
• Chunks
• delimited by anchors
• Intra file granularity
• Named via SHA-1 hash

7
Convergent Encryption (2.3)

• Need to guarantee privacy!
• Avoid duplicate of encrypted data!

8
(Continued 2.3/3)

• File owner does
• Chunk is SHA-1d ? Handle
• Keygen(Handle) ? Key
• Handle is SHA-1d ? ChunkId
• Encrypt the chunk
• Send to Backup Buddy (many)
• Buddy does
• Store encrypted, update owners, refcount

9
Abstract (3.1)

• Ideal Buddy holds a superset of my info
• Abstract random selection of the signatures
  generated by Rabin fingerprinting.
• Justification heavy overlap makes this enough
  (linux example)

10
Finding Buddies (3.2)

• Buddy search
• Probe a row on the Pastry routing table
• Coverage compare abstract
• Lighthouse Sweep probe ordering technique
• Pick based on network proximity
• Coverage-rate overlay file system overlap

11
Backup Protocol (3.3)

• Skeleton kept locally archived
• Snapshot
• Add set, Delete set, metadata list
• Signed requests
• Buddy does fetch of not stored chunks

12
Failure Malice (3.5)

• Request a random sample
• Am I being fooled?
• Buddy can drop state if it gets no requests Be
  careful.
• Collusion large node space coverage

13
Greed (3.6)

• Equivalence Classes
• Crypto Puzzles
• Electronic Currency

14
Implementation (4)

• Chunkstore FS
• daemon notified on file close
• LRU cache of recently used files
• Backup Daemon
• Uses RPC2
• Communication
• Coordinates file locking with the FS daemon

15
ChunkStore Evaluation (5.12)

• 7.4 total overhead
• Fingerprints
• Chunking
• Claim implementation
• Rewrite lists
• Per file, update 3 files
• Backup / Restore
• Comparable to copy over NFS

16
Eval Finding Buddies(5.3)

• Abstracts can be small!
• Lighthouse query finds buddies seems obvious.
  Did I miss something?

17
Probably Recoverable?

• How many chunks do I query before believing I can
  recover?
• Compute
• P(corruption)
• P(dropped chunks)
• paper is clear ?

18
Related Work / Conclusion

• Abstracts are justified by similar findings
• Most work is on centralized backup
• Some others require more than one node during
  recovery
• Pastiche Good. What author wouldnt say it?