1 p2p data management, 2006, s. abiteboul1/89 information management in p2p serge abiteboul...

1

P2P Data Management, 2006, S. Abiteboul 1/89

Information Management in P2PSerge AbiteboulINRIA-Futurs and Univ. Paris 11

2

Introduction

3


Success stories at the time of the Internet bubble

Google: management of Web pages

Mapquest: management of maps

Amazone: book catalogue

eBay: product catalogue

Napster (emule, bearshare, etc.): music database

Flickr: picture database

Wikipedia: dictionary

del.icio.us: annotations

In France: Meetic: dating databaseKelkoo: comparative shopping

They are all aboutpublishing some database

4


The trend is towards peer-to-peerand interactivity

P2P: A large and varying number of computers cooperate to solve some particular task without any centralized authority

Goal: build an efficient, robust, scalable system based (typically) on inexpensive, unreliable computers distributed in a wide area network

seti@home; kazaa; cabal

Switch from centralized servers to communities and syndication

Interaction and Web 2.0

Motivations: Social, organizational

5


Information management in a P2P network

Private terminology: data ring

Information is heterogeneous, distributed, replicated, dynamic

Which info: Data + meta-data + knowledge + services

Peers are heterogeneous, autonomous and possibly mobile• From sensors to PDA to mainframe

Typically very large number of peers

Variety of requirements: QoS, performance, security, etc.

6


Acknowledgement

Xyleme: Scalable XML warehousing

Sophie Cluet, Guy Ferran (Xyleme) & many others

ActiveXML: Language for P2P data management

Omar Benjelloun (Google), Ioana Manolescu, Tova Milo (Tel Aviv) & many others

KadoP: P2P scalable XML indexing

Ioana Manolescu, Nicoleta Preda & others

Data Ring: Infrastructure for P2P data management

Alkis Polyzotis (UC Santa Cruz)

7


Outline

1. Introduction – the data ring

2. Calculus for P2P data management (ActiveXML)

3. Algebra for P2P data management (ActiveXML algebra)

4. Indexing in P2P (KadoP)

5. Conclusion

Goal of the tutorial: present issues and technology on p2p information management

Warning: it is very biased – it is not a survey

8


Outline





5. Conclusion

9


10


The information in a data ring

Data: tuples, collections, documents, relations…

Services: data sources, possibly some processing

Meta-data about resources: attribute/values pairs, annotations

Ontologies to explain data and metadata

View definitions

Data integration information, e.g., mappings between ontologies

Physical data: Indices and materialized views

11


Functionalities of the data ring

Storage, persistence, replication

Indexing, caching, querying, updating, optimization

Schema management, access control

Fault tolerance, self tuning, monitoring

Resource discovery, history, provenance, annotations, multi-linguism,

Semantic enrichment, uncertain data

Each functionality may be achieved by a peer or by the network

12


And now, what is a peer?

A mainframe database

A file system

Web server

A PC

A PDA

A telephone

A sensor

A home appliance

A car

A manufacturing tool

A telecom equipment

A toy

Another data ring

Any connected device or software with some information to share

A net address and some names of resources (e.g. document or service)

13


Advantages and disadvantages of P2P

Scaling

Performance• Optimization of parallelism• Avoid bottleneck• Replication

Availability • Replication

Cost• Avoid the cost of server• Share operational cost

Dynamicity• add/remove new data sources

Complexity

Performance• Cost for complex queries

• Communication cost

Availability• Peers can leave

Consistency maintenance • Difficult to support transaction

Quality• Difficult to guarantee quality

14


Crash course on Web standards

Data exchange format = XML • Labeled ordered trees

• Its main asset = XML schema

• There is much more

Distributed computing protocol = Web services• SOAP Simple Object Access Protocols

• WSDL Web Service Definition Language

• UDDI Universal Description, Discovery and Integration

• BPEL Business Process Execution Language

Query languages = XPath and XQuery• Declarative query language for XML + full-text + update language

Knowledge representation = Owl or RDF/S

SOAPWSDL

XML

XqueryXpath

OwlRDFS

15

Information used to live in islands but with the Web, this is changing:

uniform access to information… …the dream for distributed data management

16


Do you like the standards?

It is the wrong question!

Correct questions: What can you do with it? What is missing?

Is Xquery the ultimate query language for the Web? No

• It is a language for querying centralized XML

• We will see what it is missing

We will not talk much about semantics

17


Automatic and distributed management of the data ring

No centralized server

No information administrator (no info manager)

Most users are non-experts• E.g., scientists

Requirements• Ease of deployment (zero-effort)

• Ease of administration (zero-effort)

• Ease of publication (epsilon-effort)

• Ease of exploitation (epsilon-effort)

• Participation in community building notably via annotations

Happy database admin

18


What should be made automatic

Self-statistics from the monitoring of the data ring• In particular, define the statistics that are needed*

Self-tuning based on the self-statistics• Choose the most appropriate organization

• Decide to install access structures: indexes, views, etc.

• Control replication of data and services

Self-healing• Recovery from errors

• E.g., replacement of a failing Web service

And automatic file management

19


Any hope?

Technology exists (database self-tuning, machine learning, etc.)

But self-tuning for databases has advanced very slowly

Why can this work?

1. There is no alternative (for db, this was just a cool gadget)

2. KISS (keep it simple stupid!)

3. The power of parallelism

This is assuming lots of machine have free cycles (true) and bandwidth is generous (not always true)

20


Distributed access control

Goal: Control access to ring resources

Access to resources is based on access rights (ACL)

Who is controlling ACLs?

A node manages ACLs for a collection of distributed resources• Easy but against the spirit and possible bottleneck

The network manages access control• Anybody can get the data

• The data is published with encryption and signatures; only nodes with proper access rights can perform reads/writes

• Some techniques exist

21


Monitoring

What is monitored?

Web service calls and database updates

The Web– Web pages– RSS feed

What is produced?

A stream of events– As a continuous service– As a RSS feed– As a Web site/page

Info-surveillance

Self-statistics and tracing

Basis for error diagnosis

22


Streams are everywhere

In query processing

In indexing (KadoP)

In recursive queries (AXML-QSQ)

In messaging, monitoring and pub/sub

That is why we will use an algebra over streams of trees and not simply an algebra over trees

23


Example: Edos distribution system

A system for the management of Linux distribution

Joint work with Mandriva Software and U. Tel Aviv

Community of open-source developers: thousands

System releases: about 10 000 software packages + metadata

Functionalities• Query the metadata

• Query subscription

• Retrieve packages

• Publish a new release or update an existing one

24


Exemple: WebContent

WebContent: an ANR platform for the management of web content

Web surveillance• Business, technical, web watching…

Participation of Gemo• WP3: knowledge

• WP5: P2P content management

Partners: CEA, EADS, Thales, Bongrain, Xyleme, Exalead, many research groups (UVSQ, Grenoble, Paris 6, etc.)

25


Taxonomy of such applications

Parameters• Number of peers and quantity of data

• How volatile the peers are

• The query/update workload

• The functionalities that are desired

Edos: peers and documents in thousands, mostly append for updates, peers not too volatile

An extreme: Google search engine in P2P for billions of documents using millions of hyper volatile peers

Mostly interested in the first case

26


Thesis

XQuery is fine for local XML processing and publishing

Not sufficient for distributed data management

The success of the relational model, i.e., of tables on a server :1. A logic for defining tables

2. An algebra for describing query plans over tables

By analogy, we need for trees in a P2P system1. A logic for defining distributed tree data and data services

2. An algebra for describing query plans over these

Proposal: ActiveXML logic and algebra

27


Outline





5. Conclusion

28

2. Active XML:a logic for distributed data management

29


The basis

AXML is a declarative language for distributed information management and an infrastructure to support the language in a P2P framework

Simple idea: XML documents with embedded service calls

Intensional data• Some of the data is given explicitly whereas for some, its definition (i.e.

the means to acquire it when needed) is given

Dynamic data• If the data sources change, the same document will provide different

information

30


Example(omitting syntactic details)

<resorts state=‘Colorado’> <resort> <name> Aspen </name> <sc> Unisys.com/snow(“Aspen”) </sc> <depth unit=“meter”>1</depth> <hotels ID=AspHotels > …. Yahoo.com/GetHotels(<city name=“Aspen”/>) </hotels> </resort> …</resorts>

May contain callsto any SOAP web service :

• e-bay.net, google.com…to any AXML web services

• to be defined

31


Marketing Philosophy

Manon: What’s the capital of Brazil?Dad: Let’s ask Wikipedia.com!

Manon: How do I get a cheap ticket to Galapagos?Dad: Let’s place a subscription on LastMinute.com!

Manon: What are the countries in the EC?Dad: France, Germany, Holland, Belgium, and hum… Let’s ask YouLists.com for more!

Active answer = intensional and dynamic and flexible

Embedding calls in data is an old idea in database

32


Active XML peer

Peer-to-peer architecture

Each Active XML peer • Repository: manages Active XML data • Web client: calls the services inside a document

• Web server: provides (parameterized) queries/updates over the repository as web services

Exchange of AXML instead of XML

AXMLpeerso

ap

33


What is an AXML peer?

PC• Now open source ObjectWeb – queries in OQL

Peer on a mass storage system• eXist (open source XML database) queries in XQuery• Xyleme queries in XyQL

PDA or cell phone• Persistence in file system and XPATH

On going: the entire network• Data is stored in a P2P network - KadoP

More: java card, a relational database…

34


A key issue: call activation

When to activate the call?• Explicit pull mode: active databases

• Implicit pull mode: deductive databases

• Push mode: query subscription

What to do with its result?

How long is the returned data valid? • Mediation and caching

Where to find the arguments?• Under the service call: XML,XPATH or a service call

35


Hi John, what is the phone number of the Prime Minister of France?• Find his name at whoswho.com then look in the phone dir• Look in the yellow pages for deVillepin’s in phone dir of www.gov.fr• (33) 01 56 00 01

Another key issue: what to send?

Send some AXML tree t• As result of a query or as parameter of a call

The tree t contains calls, do we have to evaluate them? • If I do, I may introduce service calls, do we have to evaluate all

these calls before transmitting the data?

36


Active XMLcool idea – complex problems

Blasphemous claim:

Active XML is the proper paradigm for data exchange!

Not XML + not XQuery

Brings to a unique setting

distributed db, deductive db, active db, stream data

warehousing, mediation

This is unreasonable? Yes!

Plenty of works ahead… to make it work

But first, the algebra

37


Outline



3. Algebra for P2P data management (ActiveXML algebra)• Query processing

• Query optimization


5. Conclusion

38

3. Active XML algebra

39


Motivation

Relational model: centralized tables

optimization: algebraic expression and rewriting

Active XML model: distributed trees

optimization: algebraic expression and rewriting

Distributed query optimization based on algebraic rewriting of Active XML trees

Based on experiences with AXML optimization

40


Active XML peers

We focus on positive AXML • Set-oriented data

• Positive/monotone services

Services = tree-pattern-query-with-join queries

Services produce streams • Optimized by a local query optimizer

• Evaluated by a local query processor

• Out of our scope

π

join

π

outputstream

inputstream

inputstream

Localqueryprocessing

41


The problem

An AXML system • A set of peers

• For each peer a set of documents and services

• Extensional data is distributed

• Intensional data (knowledge) is distributed

– Defined using query services (TPQJ queries)

– These services are generic: any peer can evaluate a query

A query q to some peer

Evaluate the answer to q with optimal response time

42


(AXML) algebraic expressions:

AXML algebra

AXML logic

l

E1 E2 En…

s@p

E1 E2 En…d@p

send@p

#n@p’ E1

eval@p

E1

receive@p

E1

Each such expression lives at some peerIncludes the AXML trees

43


Algebraic expressions annotations

Executing service call:

Terminated service call:

Subtletyq@p(5): definition of intensional data

eval(q@p(5)): request to evaluate it; during query optimization

q@p(5): query is being evaluated; during query processing

q@p(5): query evaluation is complete

44


Evaluation rules: local rules

for l ≠ sc, s ≠ send, receive

l

t1 t2 tn…

eval@p l

eval@p

t1

eval@p

t2

eval@p

tn…

→

tn

s@p

t1 t2 …

eval@p ●s@p

eval@p

t1

eval@p

t2

eval@p

tn…

→ E1

eval@p

eval@p E1

eval@p→

45


Evaluation rules: transfer rules

Site p asks p’ to do the work and send the result to p

s@p’

t1 t2 …

eval@p

#x@p

s@p’

t1 t2 …

receive@p

#x@p

s@p’

t1 t2 …

eval@p’

newRoot()@p’

send@p’

#x@p

→ &

46


Evaluation rules: more transfer rules

When a query is evaluated, results appear

They are sent to the place that requested them

Also some rules for eof

→ &

#x@p

Z

…

send@p’

eval@p’

●s@p’

…

#x@p

Z

…

send@p’

eval@p’

●s@p’

…

#x@p

Z

…

47


Evaluation

Reminder: setting• An AXML system

• A request to evaluate query q at peer p – eval@p( q )

Rewrite the trees in peer workspaces until termination of the process

Results

For positive XML, this process converges… to a possibly infinite state

This process computes the answer to q

May be fairly inefficient: need for optimization!

48

Optimization

More rewrite rules to evaluate a query more efficiently

49


Query optimization

Well-known optimization techniques for distributed data management

Pushing selections

Semijoin reducers

Horizontal, vertical, hybrid decomposition

Recursive query processing and query-subquery

Some “specific” AXML optimizations

Pushing queries over service calls

Lazy service call evaluation …

Optimizing subscription management

All are captured by the algebraic framework

50


Example: pushing selections

Same rule applies if d@p2 is replaced by a continuous query

q

eval@p

d@p2

q1@p

eval@p

@p2(q2@p2(d@p2))

→ q1

eval@p

(q2(d@p2))

→ q1@p

eval@p

eval@p2

(q2(d))

→

Suppose q ≡ q1((q2)):

51


Example: interleaving of processing and optimization

At peer i: di = ri di+1

Query at p1: (d1)

1. (d1) → (r1) (d2)eval@p1((r1) (d2)) → eval@p1((r1)) eval@p1((d2))eval@p1((r1)) → ●(r1) (starts streaming data)

2. (d2) → (r2) (d3); (r2) starts streaming data

3. (d3) → (r3) (d4) …

(r1) (r2) (r3) (r4)

…

52


Transfer and load balancing rules

Peer p1 delegates the evaluation of E to p2

E

eval@p1

#x@p1

→

eval@p2

send@p2

#x@p1

newRoot@p2()

eval@p1

#x@p1

send@p1

E

53




eval(E)

#x@p1

→

send@p2

#x@p1

newRoot@p2()

eval@p1

#x@p1

send@p1

eval@p2

E

54




eval(E)

#x@p1

→

newRoot@p2()#x@p1

send@p2

#x@p1 eval@p2

E

&

55




eval(E)

#x@p1

→

newRoot@p2()#x@p1

send@p2

#x@p1eval(E)

&

56




#x@p1

eval(E) →

newRoot@p2()#x@p1

send@p2

#x@p1eval(E)

&eval(E)

57




E

eval@p1

#x@p1

→

eval@p2

send@p2

#x@p1

newRoot@p2()

eval@p1

#x@p1

send@p1

E

←

58


Back to interleaved execution and optimization

Data transfers reducedMore work for p1: merging all the streams

(r1) (r2) (r3) (r4)

…

(r1) (r2) (r3) (r4) …

Hierarchical stream merging

… …

Repeated transfers

59


Example: Horizontal and vertical decomposition

A relation d over ABC that is split both horizontally and vertically d = (d1 d2) d3

d1 = B<5 (d') and d2 = B>=5 (d')

d', d1, d2 over AB and d3 over BC; each di is at a peer pi

Consider the query B=0@p(d) → B=0@p( (B<5 (d') B>=5 (d'))) d3@p3 )→ B=0 @p( d1@p1 d3@p3 )→ @p (#x@preceive(d1@p1), #y@preceive(d3@p3)) & send@p1(#x@p; B=0@p1(d1@p1) ) & send@p3(#y@p; d3@p3)

60


Common sub-expression elimination

eval@p(E), #x@preceive@p(E) →

eval@p(#x@p), #x@preceive@p(E)

eval@p

E

#x@p

E

receive@p

eval@p #x@p

E

receive@p#x@p→

& &

61


In spite of the two calls to s@p, the function is evaluated only once

Common sub-expression elimination

eval@p’

q@p’

s@p s@p

q@p’

#x@p’

s@preceive@p’

s@p

eval@p’

newRoot()@p

send@p

#x@p’ s@p

→ &

newRoot()@p’

send@p’

#y@p’ #x@p’

q@p’

#x@p’

receive@p’

s@p

#y@p’

receive@p’

s@p

newRoot()@p

send@p

#x@p’ s@p

→& &

62


Example: recursive query processing

Using a pseudo Datalog syntax

s1@p($x, $y) d2@p'($x, $z), s2@p'($z, $y)

s2@p'($x, $y) d1@p($x, $z), s1@p($y, $z)

After rewriting:

on p : #x@p receive@p(q1@p'(d2@p', s2@p') )

root@p send@p(#y@p', q2@p( d1@p, #x@p) )

on p' : root@p' send@p'(#x@p,

q1@p'(d2@p', #y@p' receive@p'(s2@p') ) )

63


q@any: where q is a query• Any peer that has some query processor for q can do it

f@any: where f is a processing service call• Example: decryption or gene comparison

q over a P2P collection

Generic and global services

eval@p

q @ coll

eval@p

q @→ index

q

eval@p

q@p1→eval@p

q@p2

64


The AXML algebra – conclusion

Captures distributed XML query processing/optimization

Based on a communication model a la CCS

Algebraic – stream-oriented

Orthogonal to the local XML query optimizer

Orthogonal to the network support (DHT, small world etc.)

What is not yet available? A cost model and heuristics

65


Outline





5. Conclusion

66

4. P2P XML indexing and query processing

67


Efficient evaluation of tree-pattern-queries

Many optimization techniques

We are interested here in distributed query evaluation/optimization

1) We consider XML indexing

2) Holistic twig join that is based on indexing

3) P2P indexing

4) P2P query processing

5) Optimizing P2P indexing

68


XML indexing: structural identifiers

1

2

3 5

6

7

84

6

6

6

8

8

8

X ancestor of Y <=>pre(X) < pre(Y) andpost(X) ≥ post(Y)

X parent of Y <=>X ancestor of Y andlevel(X) = level(Y) - 1Structural IDs = Prefix-Postfix

A

B

D E

C

F

“John” G

0

1 1

22 2

3

443

-Level

69


Holistic Twig Join

Input a document and a tree pattern query

Find the bindings of the query in the document

Holistic = holistique

(le tout et pas juste les parties)

Twig = brindille

Join = you know…

Sounds like Harry Potter?

70


Query evaluation over a document

Ids for A(1,8,0)…

Ids for D

Ids for “John”

“John”

Ids for C

Ids are sorted in lexicographical orderGoals is to find “matching Ids”

A

DC

71


The Holistic Twig Join Algorithma

b

c

a (3,5

c (4,5)

a (5,5)

c (6,8)

c (7,8)

a (8,8)

b (10,11)

c (11,11)

b (12,14)

b (13,14)

c (14,14)

a (16,17)

c (17,17)

b (19,22)

b (20,21)

c (22,22)

c (23,25)

b (24,25)

c (25,25)

a (2,8) c (9,14) a (15,17) a (18,25)

c (21,21)

r (1,25)0

1

2

3

4

level

72


Legend:

c11c9

a

c8c7

a6

b3

c6c4

The Holistic Twig Join Algorithm

a1

a2

a3

a4

a5 a7

b1 b2 b4

b5

c1 c2

c3

c5c9

b

c

(a7, b4, c8), (a7, b5, c8),

Sc

Sb

Sa

a7

b4

c8

b6

b5

c10

(a7, b4 ,c9)

b6

c11

(a7 ,b6 ,c11)

Head of the streamFind the match for the query sub-tree determined by this node !!!The ID is present also in the stack

Stacks

This is the end

73

P2P XML processing

74


XML indexing in Xyleme

History• 1999: INRIA research project • 2000: Creation of a spin-off• 2006: About 25 people

Technology• A scalable XML repository• A content warehouse • On a cluster of Linux PC

XML query processing • Twig join• Index is distributed• Keyword-based vs. document

based

LAN

Put(C;[d,p,6,6,1])Put(“John”;[d,p,3,1,2])

hash(C)

hash(“John)

75


Query processing over a distributed collection

Ids for A(p12,d456, 1,7,0)…

Ids for D

Ids for John

A

DC

“John”

Ids for C

Ids include peerId and docIdIds are sorted in lexicographical orderGoals is to find “matching Ids” in the collection

76


XML indexing in KadoP

Use structural Ids

Publish them via a DHT• Distributed Hash Table

• Peers come and go

• Locate(k):: log(n) messages to fing the peer in charge of key k

• Put(k,v)

• Get(k): retrieves all the values for k

We use Pastry • We also tried P2PSim and JXTA

DHT

put(C;[d,p,6,6,1])

put(“John”;[d,p,3,1,2])

hash(C)

hash(“John)

posting for C

put(C;[d,p,6,6,1])

77


XML query processing in KadoP

Given a tree pattern query Q

Evaluate an index query indexQ to locate the peers that can provide some answers

• indexQ is a twig join

Ship Q to these peers and evaluate it there

If indexQ is imprecise, many false positive• Example: ship Q to all peers (maximal parallelism)

• Example: Instead of structural Ids, just use (label/word,peerId,docId)

78


KadoP architecture

KadoP peerpublish & query

KadoP Engine

DHT locate, put, get & delete Index

Indexing

Queryprocessing

ActiveXMLengine

Web interface

Semantic layerExternalLayer

LogicalLayer

PhysicalLayer

79


Some technical issues

Our goal: manage millions of documents with a large number of peers

First experiments were a disaster

Replace the index storage of the DHT in a FS by storage in a database (Berkeley DB)

Extend the API of the DHT to have Append and not only Read/Write

Extend the API of the DHT to have a streaming exchange of postings• Useful because the XML algebra works better with streams

Now it scales but there is the issue of long postings

80


The issue of long postings: Google in P2P

Using keyword distribution

Suppose • Peer for Ullman is in Europe• Peer for XML is in US

we have to ship one long posting between US and Europe

For a large number of users, we absorb all the bandwidth of Internet backbone

Need for replication

Even for thousands of peers, the exchange of long postings is an issue

DHTUllmanxml

Ullmann & xml?

81


Intensional indexing in KadoP

long postingh(Name)

Distributed B-tree

Long posting = bad response time

1. No long posting

2. get h(name) then parallel fetch

3. Possibility to optimize further

f(docId55..docId75)

may be it does not match

no need to call f

h(Name)

f g h i

82


More optimization

• Standard for P2P keyword search– Gap compression and adaptive set intersection

• Standard distributed query optimization techniques– Ship smallest list – Load balancing– Caching– Replication

• Semi-join techniques notably Bloom semi-join

83


Outline





5. Conclusion

84

6. Conclusion

85


Logic for distributed data management• Opinion: XQuery is a language for local XML management

• Proposal: ActiveXML

Algebraic foundation of distributed query optimization• Proposal: ActiveXML algebra

P2P (Active) XML indexing• KadoP is now being tested and we are working on optimization

Software• ActiveXML is open-source – see activexml.net

• KadoP soon will be – already available upon request

• EDOS distribution system as well

Conclusion

86


Lots of related work and related systems

This is going very fast in system devepments

Structured P2P nets: Pastry, Chord

Content delivery net: Coral, Akamai

XML repositories: Xyleme, DBMonet

Multicas systemst: Avalanche, Bullet

File sharing systems: BitTorrent, Kazaa

Pub/Sub systems: Scribe, Hyper

Distributed storage systems: OceanStore, GoogleFS

Etc.

Fundamental research is somewhat left behind

87


Issues

P2P query optimization

P2P access control

P2P archiving

P2P self tuning

P2P monitoring

P2P knowledge management [SomeWhere]

Also: analysis and verification of these systems• E.g., termination, error detection, diagnosis

88


Find your own topic

Pick your favorite problem for data or knowledge management and study it in a P2P setting

with gigabytes of data and thousands of peers

If you find it boring, consider it

with terabytes of data and millions of peers

89


Merci

Merci

1 p2p data management, 2006, s. abiteboul1/89 information management in p2p serge abiteboul...

Documents

data ring data

p2p data managementactivexml

data ring slide

data sources

data ring information

data ring storage

p2p kadop

interactivity p2p