strongbox: towards reliability and consistency in cloud-based

31
*-Box (star-box) Towards Reliability and Consistency in Dropbox-like File Synchronization Services Yupu Zhang, Chris Dragga, Andrea Arpaci-Dusseau, Remzi Arpaci-Dusseau University of Wisconsin - Madison 6/27/2013 1

Upload: others

Post on 04-Feb-2022

2 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: StrongBox: Towards Reliability and Consistency in Cloud-Based

*-Box (star-box)

Towards Reliability and Consistency in Dropbox-like File Synchronization Services

Yupu Zhang, Chris Dragga,

Andrea Arpaci-Dusseau, Remzi Arpaci-Dusseau

University of Wisconsin - Madison

6/27/2013 1

Page 2: StrongBox: Towards Reliability and Consistency in Cloud-Based

Cloud-Based File Synchronization Services

• Exploding in popularity – Numerous providers: Dropbox, Google Drive, SkyDrive …

– Large user base: Dropbox has more than 100 million users

• Key benefit – Automatic synchronization across clients/devices

– Reliable data storage on the server through replication

“your stuff is safe in Dropbox and will never be lost” - Dropbox tour, page 1

6/27/2013 2

Page 3: StrongBox: Towards Reliability and Consistency in Cloud-Based

Is Your Data Really Safe?

• Data corruption

– Uploaded from local machine to cloud

– Propagated to other devices/clients

6/27/2013 3

Page 4: StrongBox: Towards Reliability and Consistency in Cloud-Based

1

• Crash consistency

– Inconsistent data ends up everywhere

– “Out-of-sync” synchronization

1 1 2

Is Your Data Really Safe?

6/27/2013 4

1

1

after reboot

sync client thinks everything is in sync

Page 5: StrongBox: Towards Reliability and Consistency in Cloud-Based

Your Data is NOT Really Safe

• False sense of safety

– Many copies do NOT always make your data safe

• Why?

– Semantic gap between local file system and cloud

– Separately designed and loosely linked

6/27/2013 5

Page 6: StrongBox: Towards Reliability and Consistency in Cloud-Based

Project *-Box (star-box)

• Goal

– Close the gap between local file system and cloud

– Provide * without too much infrastructure changes

• * represents desired properties

– e.g., reliable, consistent, fast, private …

• Currently focus on two properties

– Reliable: Data corruption

– Consistent: Crash consistency

6/27/2013 6

Page 7: StrongBox: Towards Reliability and Consistency in Cloud-Based

Outline

• Introduction

• Data Corruption

• Crash Consistency

• Current Status

• Conclusion

6/27/2013 7

Page 8: StrongBox: Towards Reliability and Consistency in Cloud-Based

Corruption Problem

• Data corruption is not uncommon

– Comes from disk media, firmware, controllers [Bairavasundaram07, Anderson03]

– Remains local w/o synchronization

• With synchronization

– Corruption may propagate and pollute other copies

• Synchronization is a double-edged sword

– Make sure synchronized data is good

6/27/2013 8

Page 9: StrongBox: Towards Reliability and Consistency in Cloud-Based

I1 I

Dedup

NOT a duplicate duplicate

Disk

Background

9 6/27/2013

Dropbox Client

Local Database

File System

Dropbox Server

Name Attributes

foo I

Inotify

Rsync

C1 C2 C3

D1 D2

foo [v0] (4MB data chunks)

foo: inode

NO offline changes

D2’

overwrite D2

foo was modified I’

read foo (chunk by chunk)

D1 … …

C1’ C2 C3

4KB data blocks

D2’

C1’

foo [v1] (4MB data chunks)

Name Attributes

foo I’

Changed Multiple Times

Page 10: StrongBox: Towards Reliability and Consistency in Cloud-Based

D1 D1 D1 D1 Disk M1 I

Inject Corruption

10 6/27/2013

File System

Dropbox Server

Inotify

C1 C2 C3

D2

foo [v0] (4MB data chunks)

… …

Corrupt D1

foo: inode

4KB data blocks

Page 11: StrongBox: Towards Reliability and Consistency in Cloud-Based

I I

Local Database

Name Attributes

foo I

Dedup

Disk

Start Client

11 6/27/2013

Dropbox Client

File System

Dropbox Server

Inotify

Rsync

C1 C2 C3

D1 D2

foo [v0] (4MB data chunks)

D1 … …

NO offline changes

foo: inode

4KB data blocks

Page 12: StrongBox: Towards Reliability and Consistency in Cloud-Based

I I’

D1 D1

Dedup

NOT a duplicate

Disk I

Data Write

12 6/27/2013

Dropbox Client

Local Database

File System

Dropbox Server

Inotify

Rsync

C1 C2 C3

D2

foo [v0] (4MB data chunks)

D2’

overwrite D2

foo was modified

read foo (chunk by chunk)

… …

C1’

D2’

C1’

foo [v1] (4MB data chunks)

D1

foo: inode

4KB data blocks

Name Attributes

foo I

Name Attributes

foo I’

Page 13: StrongBox: Towards Reliability and Consistency in Cloud-Based

I I’

D2 D1 D1

Dedup

NOT a duplicate

Disk I

Metadata Change

13 6/27/2013

Dropbox Client

Local Database

File System

Dropbox Server

Inotify

Rsync

C1 C2 C3

D2

foo [v0] (4MB data chunks)

touch -m foo’s metadata was changed

read foo (chunk by chunk)

… …

C1’

D2’

C1’

foo [v1] (4MB data chunks)

D1

foo: inode

4KB data blocks

Name Attributes

foo I

Name Attributes

foo I’

Page 14: StrongBox: Towards Reliability and Consistency in Cloud-Based

More Results

FS Service Data Writes Metadata Changes

mtime ctime atime

ext4 (Linux)

Dropbox L G L G L G L

ownCloud L G L G L L

FileRock L G L G L L

HFS+ (Mac OS X)

Dropbox L G L G L L

ownCloud L G L G L L

GoogleDrive L G L G L L

SugarSync L G L L L

Syncplicity L G L G L L

• Corruption is propagated when there is a change to file data • Even if there is no data change, corruption may still be uploaded

6/27/2013 14

L: Local corruption G: Global corruption

Page 15: StrongBox: Towards Reliability and Consistency in Cloud-Based

Summary

• Bad bits are promoted to resilient bad bits

– ALL copies polluted

– Cloud copies protected by checksum

• Fundamental problem, not implementation bugs

– FS monitoring services only provide file-level notification

– Sync clients cannot tell legitimate changes from corruption

• Redundant data on the cloud is not fully utilized

– If corruption can be detected, local FS can recover from corruption using cloud copies

6/27/2013 15

Page 16: StrongBox: Towards Reliability and Consistency in Cloud-Based

Outline

• Introduction

• Data Corruption

• Crash Consistency

• Current Status

• Conclusion

6/27/2013 16

Page 17: StrongBox: Towards Reliability and Consistency in Cloud-Based

Crash Recovery Techniques

• Copy-on-write (e.g., ZFS, btrfs)

– Always roll back to a consistent version

• Journaling (e.g., ext4)

– Data journaling mode • Both data and metadata are logged

• Provide data consistency

– Ordered journaling mode • Only journal metadata

• Data blocks are written before metadata is logged

• Cannot guarantee data consistency

6/27/2013 17

Page 18: StrongBox: Towards Reliability and Consistency in Cloud-Based

Disk I

I I’ I’ I’ D1’

1. Write dirty data blocks to home locations

2. Write metadata blocks to journal

3. Write journal commit block to the journal

4. Checkpoint journaled metadata blocks to home locations

Ordered Mode

18 6/27/2013

File System

D1 D2

D2’

D1

overwrite D1 &D2

D2’

1 2

C

3 4

foo: inode

4KB data blocks

D1’

1

Journal / Log

Page 19: StrongBox: Towards Reliability and Consistency in Cloud-Based

Disk

I I’ I’

I

D1’

• Crash during step 1

– Write dirty data blocks to home locations

Crash in Ordered Mode

19 6/27/2013

File System

D1 D2

D2’

D1

overwrite D1 &D2

D2’

1

foo: inode

4KB data blocks

D1’

1 Consistent Data Inconsistent Data

Page 20: StrongBox: Towards Reliability and Consistency in Cloud-Based

Ext4 Ordered Mode + Dropbox

• Case 1

– Inconsistent data is propagated

• Case 2

– Consistent data is NOT synchronized

6/27/2013 20

Page 21: StrongBox: Towards Reliability and Consistency in Cloud-Based

I I’

I Disk D1

Dedup

Case 1 Crash

21 6/27/2013

Dropbox Client

File System

Dropbox Server

Inotify

Rsync

C1 C2 C3

D2

foo [v0] (4MB data chunks)

D2’

… …

foo: inode

4KB data blocks

overwrite D1 &D2

D1’ D1’

crash AFTER database is changed

inconsistent data on disk

Local Database

Name Attributes

foo I

Name Attributes

foo I’

NOT fully updated

foo was modified

Page 22: StrongBox: Towards Reliability and Consistency in Cloud-Based

I I

C1

Dedup

Disk

Case 1 Reboot

22 6/27/2013

Dropbox Client

Local Database

File System

Dropbox Server

Inotify

Rsync

C2 C3

D1 D2

foo [v0] (4MB data chunks)

D1’ … …

foo: inode

4KB data blocks

C1’

foo [v1] (4MB data chunks)

inconsistent data D1’ D2 on cloud

Name Attributes

foo I’

NOT fully updated

Sync!

inconsistent data on disk

Page 23: StrongBox: Towards Reliability and Consistency in Cloud-Based

I I’

I Disk D1

Dedup

Case 2 Crash

23 6/27/2013

Dropbox Client

Local Database

File System

Dropbox Server

Inotify

Rsync

C1 C2 C3

D2

foo [v0] (4MB data chunks)

D2’

… …

foo: inode

4KB data blocks

overwrite D1 &D2 (O_SYNC)

D1’ D1’ D2’

crash BEFORE database is changed

consistent data on disk

Name Attributes

foo I

Page 24: StrongBox: Towards Reliability and Consistency in Cloud-Based

I I

C1

Dedup

Disk

Case 2 Reboot

24 6/27/2013

Dropbox Client

Local Database

File System

Dropbox Server

Inotify

Rsync

C2 C3

D1 D2’

foo [v0] (4MB data chunks)

D1’ … …

NO offline changes

foo: inode

4KB data blocks

Server and other devices still have v0

This machine has v1

Name Attributes

foo I

NO sync!

consistent data on disk

Page 25: StrongBox: Towards Reliability and Consistency in Cloud-Based

Recover using Data on Cloud?

• Data on server does not always reflect a consistent state on disk

– Dropbox uploads data asynchronously

– Dropbox reorders file uploading

– Actively modified files may get delayed

• When crash occurs, files on server could be inconsistent with respect to disk

6/27/2013 25

Page 26: StrongBox: Towards Reliability and Consistency in Cloud-Based

Summary

• Inconsistent content gets propagated

• “Out-of-sync” files may exist

– Different client/devices see different versions of the same file

• Need in-depth communication between local FS and cloud

– Cloud has very weak sense of actual FS state

6/27/2013 26

Page 27: StrongBox: Towards Reliability and Consistency in Cloud-Based

Outline

• Introduction

• Data Corruption

• Crash Consistency

• Current Status

• Conclusion

6/27/2013 27

Page 28: StrongBox: Towards Reliability and Consistency in Cloud-Based

Current Status of *-Box

• Finished

– Data checksumming in ext4

– Dropbox-aided corruption recovery in ext4

– Fine-grained inotify in Linux

• Add a ranged file-update notification

• Adapt ownCloud (an open-source sync service) to use it

• In-progress

– Use in-memory snapshot to facilitate crash recovery in ext4 ordered mode

6/27/2013 28

Page 29: StrongBox: Towards Reliability and Consistency in Cloud-Based

Outline

• Introduction

• Data Corruption

• Crash Consistency

• Possible Solutions

• Conclusion

6/27/2013 29

Page 30: StrongBox: Towards Reliability and Consistency in Cloud-Based

Conclusion

• Many copies do NOT always make your data safe

– Propagation of corrupt data and inconsistent state

– Synchronized files are out-of-sync

• Propose *-Box project

– Solve problems by reducing the semantic gap between existing local FS and cloud storage

• Ultimately may need a cohesive system that provides capabilities unachievable in isolation

6/27/2013 30

Page 31: StrongBox: Towards Reliability and Consistency in Cloud-Based

Thank you!

Questions?

Advanced Systems Lab (ADSL)

University of Wisconsin-Madison

http://www.cs.wisc.edu/adsl

Wisconsin Institute on Software-defined Datacenters in Madison

http://wisdom.cs.wisc.edu/

6/27/2013 31