Database ,16 P2P

Distributed DBMS ©M. T. Özsu & P. Valduriez Ch.16/1
Outline
• Introduction
• Background
• Distributed Database Design
• Database Integration
• Semantic Data Control
• Distributed Query Processing
• Distributed Transaction Management
• Data Replication
• Parallel Database Systems
• Distributed Object DBMS
• Peer-to-Peer Data Management
➡ P2P infrastructure
➡ Schema mapping
➡ Querying over P2P systems
➡ Replica management in P2P systems
• Web Data Management
• Current Issues

Motivations
• P2P systems
➡Each peer can have same functionality
➡Decentralized control, large scale
➡Low-level, simple services
✦ File sharing, computation sharing, com. sharing
• Traditional distributed DBMSs
➡High-level data management services
✦ queries, transactions, consistency, security, etc.
➡Centralized control, limited scale
• P2P + distributed database
➡Why? How?

Why High-level P2P Data
Sharing?
• Professional community example
➡ Medical doctors in a hospital may want to share (some of) their patient data
for an epidemiological study
➡ They have their own, independent patient descriptions
➡ They want to ask queries such as “age and weight of male patients diagnosed
with disease X …” over their own descriptions
➡ They don’t want to create a database and set a centralized server

Problem Definition
• P2P system
➡ No centralized control, very large scale
➡ Very dynamic: peers can join and leave the network at any time
➡ Peers can be autonomous and unreliable
• Techniques designed for distributed data management need be extended
➡ Too static, need to be decentralized, dynamic and self-adaptive

Peer Reference Architecture

Potential Benefits of P2P
Systems
• Scale up to very large numbers of peers
• Dynamic self-organization
• Load balancing
• Parallel processing
• High availability through massive replication

P2P vs Traditional Distributed
DBMS
P2P Distributed
DBMS
Joining the
network
Upon peer’s
initiative
Controled by DBA
Queries No schema,
key-word based
Global schema,
static optimization
Query answers Partial Complete
Content location Using neighbors
or DHT
Using directory

Requirements for P2P Data
Management
• Autonomy of peers
➡ Peers should be able to join/leave at any time, control their data with respect
to other (trusted) peers
• Query expressiveness
➡ Key-lookup, key-word search, SQL-like
• Efficiency
➡ Efficient use of bandwidth, computing power, storage

Requirements for P2P Data
Management (cont’d)
• Quality of service (QoS)
➡ User-perceived efficiency: completeness of results, response time, data
consistency, …
• Fault-tolerance
➡ Efficiency and QoS despite failures
• Security
➡ Data access control in the context of very open systems

P2P Network Topologies
• Pure P2P systems
➡ Unstructured systems
✦ e.g. Napster, Gnutella, Freenet, Kazaa, BitTorrent
➡ Structured systems (DHT)
✦ e.g. LH* (the earliest form of DHT), CAN, CHORD, Tapestry, Freepastry, Pgrid,
Baton
• Super-peer (hybrid) systems
➡ e.g. Edutela, JXTA
• Two issues
➡ Indexing data
➡ Searching data

P2P Unstructured Network
• High autonomy (peer only needs to know neighbor to
login)
• Searching by
➡ flooding the network: general, may be inefficient
➡ Gossiping between selected peers: robust, efficient
• High-fault tolerance with replication
p2p
data
peer 1
p2p
data
peer 2
p2p
data
peer 3
p2p
data
peer 4

Search over Centralized Index
1. A peer asks the
central index
manager for resource
2. The response
identifies the peer
with the resource
3. The peer is asked for
the resource
4. It is transferred

Search over Distributed Index
1. A peer sends the
request for
resource to all its
neighbors
2. Each neighbor
propagates to its
neighbors if it
doesn’t have the
resource
3. The peer who has
the resource
responds by
sending the
resource

P2P Structured Network
• Simple API with put(key, data) and get(key)
➡ The key (an object id) is hashed to generate a peer id, which stores the
corresponding data
• Efficient exact-match search
➡ O(log n) for put(key, data), get(key)
• Limited autonomy since a peer is responsible for a range of keys
p2p
d(k1)
peer 1
p2p
d(k2)
peer 2
p2p
d(k3)
peer 3
p2p
d(k4)
peer 4
h(k1)= p1 h(k2)= p2 h(k3)= p3 h(k4)= p4
Distributed Hash Table (DHT)

Super-peer Network
• Super-peers can perform complex functions (meta-data
management, indexing, access control, etc.)
➡ Efficiency and QoS
➡ Restricted autonomy
➡ SP = single point of failure use several super-peers
p2sp
data
peer 1
p2sp
data
peer 2
p2sp
data
peer 3
p2sp
data
peer 4
sp2sp
sp2p
sp2sp
sp2p

Search over a Super-peer
System
1. A peer sends the
request for resource to
all its super-peer
2. The super-peer sends
the request to other
super-peers if necessary
3. The super-peer one of
whose peers has the
resource responds by
indicating that peer
4. The super-peer notifies
the original peer

P2P Systems Comparison
Requirements Unstructured DHT Super-peer
Autonomy high low avg
Query exp. high low high
Efficiency low high high
QoS low high high
Fault-tolerance high high low
Security low low high

P2P Schema Mapping
• Problem: support decentralized schema mapping so that a query expressed
on one peer’s schema can be reformulated to a query on another peer’s
schema
• Main approaches
➡ Pairwise schema mapping
➡ Mapping based on machine learning
➡ Common agreement mapping

Pairwise Schema Mapping
• Each user defines the
mapping between the local
schema and the schema of
any other peer that contains
data that are of interest
• Relying on the transitivity
of the defined mappings,
the system tries to extract
mappings between schemas
that have no defined
mapping

• This approach is generally used when the shared data are defined based
on ontologies and taxonomies as proposed for the semantic web
• It uses machine learning techniques to automatically extract the mappings
between the shared schemas
• The extracted mappings are stored over the network, in order to be used
for processing future queries
Mapping Based on Machine
Learning

Common Agreement Mapping
• Some (cooperating)
peers must agree on a
Common Schema
Description (CSD)
• Given a CSD, a peer
schema can be specified
using views
➡ Similar to the LAV
approach
• When a peer decides to
share data, it needs to
map its local schema to
the CSD

Querying over P2P Systems
• P2P networks provide basic query routing
➡ Sufficient for simple, exact-match queries, e.g. with a DHT
• Supporting more complex queries, particularly in DHTs, is difficult
• Main types of complex queries
➡ Top-k queries
➡ Join queries
➡ Range queries

Top-k Query
• Returns only k of the most relevant answers, ordered by relevance
➡ Like for a search engine
• Scoring function (sf) determines the relevance (score) of answers to the
query
• Example
SELECT *
FROM Patient P
WHERE (P.disease = “diabetes”) AND
(P.height < 170) AND (P.weight > 70)
ORDER BY sf(height, weight)
STOP AFTER 10
➡ Example of sf: weight − (height -100)

General Model for Top-k Queries
• Suppose we have:
➡ n data items
✦ items can be document, tuples, etc.
➡ m lists of n data items such that
✦ Each data item has
✓ a local score in each list
✓ an overall score computed based on its local scores in all lists using a given scoring function
✦ Each list
✓ contains all n data items (or item ids)
✓ is sorted in decreasing order of the local scores
• The objective is:
➡ Given a scoring function, find the k data items whose overall scores are the highest

Execution Cost of Top-k Queries
• Two modes of access to a sorted list
➡ Sorted (sequential) access
✦ Starts with the first data item, then accesses each next item
➡ Random access
✦ Looks up a given data item in the list by its identifier (e.g. TID)
• Given a top-k algorithm A and a database D (i.e. set of sorted lists), the cost
of executing A over D is:
➡ Cost(A, D) = (#sorted-access * sorted-access-cost) + (#random-access *
random-access-cost)

Basic Top-k Algorithms
• Fagin’s Algorithm (FA)
➡ General model of top-k queries using sorted lists
➡ Simple algorithm
✦ Do sorted access in parallel to the lists until at least k data items have been seen in all
lists
• Threshold Algorithm (TA)
➡ Proposed independently by several groups
➡ Efficient algorithm over sorted lists
➡ The basis for many TA-style distributed algorithms
✦ Mainly for DHTs
✦ Algorithms for unstructured or super-peer simpler

TA
• Similar to FA in doing sorted access to the lists
➡ But different stopping condition
• Unlike FA, no need to wait until the lists give k items
➡ Once an item has been seen from a sorted access, get all its scores through
random access
• But how do we know that the scores of seen items are higher than those of unseen
items?
➡ Use a threshold (T) to predict maximum possible score of unseen items
✦ based on the last scores seen in the lists under sorted access
➡ Then stop when there are at least k data items whose overall score ≥ T

TA Example
• Assume sf( ) = s1 + s2 + s3, k = 3,
Y: {top seen items with overall scores}
• At position 1
➡ Look up the local scores of items d1, d2 and d3 in other
lists using random access and compute their overall
scores (which are 65, 63 and 70, respectively)
➡ Y = {(d1,70) (d2,65) (d3,63)}, T = 30 + 28 + 30 = 88
• Then
➡ At position 2, Y = {(d3,70) (d4,70) (d5,65)}, T = 84
➡ At position 3, Y = {(d3,71) (d5,70) (d8,70)}, T = 80
➡ At position 4, Y = same, T = 75
✦ which is less than the overall score of the three data items in
Y. Thus, TA stops
• Note that the contents of Y at position 6 is
exactly the same as at position 3

Improvement over TA: BPA
• Best Position Algorithm
• Main idea: keep track of the positions (and scores) of the items seen under sorted
or random access
➡ Enables BPA to stop as soon as possible
✦ In the previous example, BPA stops at position 3
• Best position = the greatest seen position in a list such that any position
before it is also seen
➡ Thus, we are sure that all positions between 1 and best position have been seen
• Stopping condition
➡ Based on best positions overall score, i.e. the overall score computed based on
the best positions in all lists

Top-k Query Processing in DHTs
• Given
➡ A DHT network
➡ T : a set of tuples which are stored in the DHT
➡ Q : a top-k query
➡ sf : a scoring function
• Goal
➡ Retrieve efficiently the k tuples stored in the DHT whose scores are the
highest according to Q and sf
➡ While avoiding centralized data storage
• DHTop algorithm
➡ TA style

Data Storage Mechanism
• Two complementary methods for storing tuples in the DHT
➡ Tuple storage: each tuple of a relation is stored in the DHT using its tuple’s
identifier (e.g. primary key)
✦ Allows to retrieve tuples using their identifier
➡ Attribute storage: the values of some attributes of a tuple are stored
individually in the DHT
✦ Acts as secondary indices
✦ Good support for exact match queries

Tuple Storage Method
• Let
➡ R (a1, a2, …, am) be a relation
➡ t 〈v1, v2, …, vm〉 be a tuple of R
➡ v1 be the primary key of tuple t
➡ h : a hash function that hashes its inputs into a DHT key
• To store tuple t in the DHT, a peer does as follows:
➡ ts_key = h(R, v1);
➡ put(ts_key, v1, v2, …, vm );

Attribute-value storage
• Domain partitioning
➡ Let Da be the domain of an attribute a
➡ We partition Da into n nonempty sub-domains d1, d2, ..., dn
➡ Example:
✦ Attribute “weight”, with domain [0..200] in kilograms, can be partitioned into
n=40 sub-domains [0..5), [5..10), …, [190..195), [195..200]
• Exploitation by peers
➡ All peers know the sub-domains of each attribute
➡ Thus, each peer can locally execute the following function:
✦ sd(a, v) : returns the sub-domain of attribute a to which value v belongs

Attribute-Value Storage (cont’d)
• Let
➡ R(a1, a2, …, am) be a relation
➡ t v1, v2, …, vm be a tuple of R
➡ ts_keyt : the key used for storing tuple t in the DHT, i.e. h(R, v1);
• For storing the value v2 of attribute a2 in the DHT :
➡ let di be the sub-domain to which v2 belongs, i.e. di = sd(a2, v2);
avs_key h(R, a2, di);
put(avs_key, v2 , ts_keyt );
• Two important features:
1. After retrieving an attribute value, we are able to retrieve its entire tuple
2. The values that are “relatively close” are stored at the same peer

DHTop
• Two phases, executed by the peer where Q is originated
• Phase 1: Prepare ordered lists of attributes’ sub-domains
➡ For each scoring attribute a :
✦ Step 1: create a list of a’s sub-domains, denoted by La
✦ Step 2: remove useless sub-domains by considering Q’s conditions
✓ E.g. condition a = c , remove all sub-domains except sub-domain of c
✦ Step 3: sort La according to the positive impact of its sub-domains on the
scoring function

DHTop (cont’d)
• Phase 2: Continuously retrieve attribute values and their tuples until finding
the k top tuples
➡ For each scoring attribute a do in parallel
i=1
Repeat
1. Ask the peer p that maintains attribute values for La[i] (i.e. i-th sub-
domain in the list) to return its attribute values in order of their positive
impact on the scoring function
2. After receiving each attribute value v
o Retrieve the entire tuple
o Check End-Condition: are there at least k tuples (among the
retrieved tuples) whose scores are greater than the threshold?
o If end-condition holds, then exit
3. If all values maintained by p are received then
o i = i + 1
Until (end-condition holds) or (i ≥ number of a’s sub-domains)

DHTop Threshold
• Let a1, a2, ..., am be the scoring attributes, and v1, v2, ..., vm be the last values
retrieved for them
• Let f be the scoring function
• The threshold is = f(v1, v2, …, vm)

Top-k Query Processing in
Unstructured Systems
• Fully Decentralized (FD) algorithm
➡ Fully decentralized query execution
➡ No centralized information
➡ Considers dynamicity of peers
➡ Simple, efficient and fault-tolerant
✦ Many optimizations possible

FD Algorithm Overview
1. Query Forward
➡ If TTL > 0, send Q to your neighbors
2. Local query execution
➡ Extract the k local top scores
3. Wait
➡ Wait to receive from the neighbors their k top scores
➡ The wait time is based on the received TTL, network parameters and peer’s local
processing parameters
4. Merge-and-Backward
➡ Extract k top scores from the local and received scores
➡ Send the k top scores to your parent
5. Data retrieval (at the query originator)
➡ Retrieve the overall top-k data using the final top scores

State Transition Diagram

FD Optimizations
• What if a peer receives a score-list after its wait time has expired?
➡ Urgent Score-list: a mechanism to backward a score-list to another peer still in
the wait phase
• What if the parent of a peer p leave the system?
➡ If p has a neighbor that is not p’s child, the score-list can be sent to it as an
urgent score-list
➡ If p has not such a neighbor, it can send its merged score-list directly to the
query originator
• Return results ASAP
➡ Without waiting for the entire result set

Join Query Processing in DHTs
• A DHT relies on hashing to store and locate data
➡ Basis for parallel hash join algorithms
• Basic solution in the context of the PIER P2P system
➡ Let us call it PIERjoin
➡ Assume that the joined relations and the result relations have a home which
are the peers that store horizontal fragments of the relation
✦ Recall def. of home from Chapter 8
➡ Make use of the put method for distributing tuples onto a set of peers based
on their join attribute so that tuples with the same join attribute values are
stored at the same peers
➡ Then apply the probe/join phase

PIERjoin Algorithm
1. Multicast phase
➡ The query originator peer multicasts Q to all peers that store tuples of the
join relations R and S, i.e., their homes.
2. Hash phase
➡ Each peer that receives Q scans its local relation, searching for the tuples
that satisfy the select predicate (if any)
➡ Then, it sends the selected tuples to the home of the result relation, using
put operations
➡ The DHT key used in the put operation uses the home of the result
relation and the join attribute
3. Probe/join phase
➡ Each peer in the home of the result relation, upon receiving a new
tuple, inserts it in the corresponding hash table, probes the opposite hash
table to find matching tuples and constructs the result joined tuples

Range Query Processing
• Range query
➡ WHERE clause of the form “attribute A in range [a; b]”
• Difficult to support in structured P2P systems, in particular, DHTs
➡ Hashing tends to destroy the ordering of data that is useful in finding ranges
quickly
• Two main approaches for supporting range queries in structured P2P
systems
➡ Extend a DHT with proximity or order-preserving properties
✦ Problem: data skew that can result in peers with unbalanced ranges, which hurts
load balancing
➡ Maintain the key ordering with a tree-based structure
✦ Better at maintaining balanced ranges of keys

BATON
• BATON (BAlanced Tree Overlay Network)
• Organizes peers as a balanced binary tree
➡ Each node of the tree is maintained by a peer
➡ The position of a node is determined by a (level, number) tuple, with level
starting from 0 at the root, number starting from 1 at the root and sequentially
assigned using in-order traversal
➡ Each tree node stores links to its parent, children, adjacent nodes and selected
neighbor nodes that are nodes at the same level
➡ Two routing tables: a left routing table and a right routing table store links to
the selected neighbor nodes

BATON Structure-tree Index
• Each node (or peer) is assigned a range of values
➡ Maintained at the routing table of each link
➡ Required to be to the right of the range managed by its left subtree and less than
the range managed by its right subtree

Range Query Processing in
BATON
Input: Q, a range query in the form [a,b]
Output: T: result relation
1. Search for the peer storing the lower bound of the range
At query originator node do
find peer p that holds value a
send Q to p
2. A peer p that receives Q (from query originator or its left adjacent
peer) searches for local tuples and sends Q to its right adjacent node
At each peer p that receives Q
Tp = Range(p) [a,b]
send Tp to query originator
If (Range(RighAdjacent(p)) [a,b]) not empty
send Q to right adjacent peer of p
• With X nodes covering the range, Q is answered in O(log n + X) steps

Example of Range Query
Execution
• Consider Q with range [7; 45] issued
at node 7
• First, execute an exact match query
looking for a node containing the
lower bound of the range (see dashed
line)
• Since the lower bound is in node 4’s
range, check locally for tuples
belonging to the range and forward Q
to its adjacent right node (node 9)
• Node 9 checks for local tuples
belonging to the range and forwards
Q to node 2
• Nodes 10, 5, 1 and 6 receive Q, check
for local tuples and contact their
respective right adjacent node until
the node containing the upper bound
of the range is reached

Replica Consistency
• To increase data availability and access performance, P2P systems
replicate data, however, with very different levels of replica
consistency
➡ The earlier, simple P2P systems such as Gnutella and Kazaa deal only
with static data (e.g., music files) and replication is “passive” as it occurs
naturally as peers request and copy files from one another
(basically, caching data)
• In more advanced P2P systems where replicas can be updated, there is
a need for proper replica management techniques
• Replica consistency in DHTs
➡ Basic support - Tapestry
➡ Replica reconciliation - OceanStore

Tapestry
• Decentralized object location and routing on top of a structured overlay
• Routes messages to logical end-points (i.e., not associated with physical
location), such as nodes or object replicas.
➡ This enables message delivery to mobile or replicated endpoints in the
presence of network instability
• Location and routing
➡ Let O be an object identified by id(O), the insertion of O involves two nodes:
the server node (noted ns) that holds O and the root node (noted nr) that holds
a mapping in the form (id(O); ns) indicating that the object identified by id(O)
is stored at node ns
➡ The root node is dynamically determined by a globally consistent
deterministic algorithm

Object Publishing in Tapestry
• When O is inserted into ns, ns
publishes id(O) at its root node by
routing a message from ns to nr
containing the mapping (id(O); ns)
• This mapping is stored at all nodes
along the message path
• During a location query (e.g.,
“id(O)?”, the message that looks for
id(O) is initially routed towards nr,
but it may be stopped before
reaching it once a node containing
the mapping (id(O); ns) is found
• For routing a message to id(O)’s
root, each node forwards this
message to its neighbor whose
logical identifier is the most similar
to id(O)

Replica Management in Tapestry
• Each node represents a peer and contains
the peer’s logical identifier in hexadecimal
format
• Two replicas O1 and O2 of object O (e.g., a
book file) are inserted into distinct peers
(O1 at peer 4228 and O2 at peer AA93). The
identifier of O1 is equal to that of O2 (i.e.,
4378)
• When O1 is inserted into its server node
(peer 4228), the mapping (4378; 4228) is
routed from peer 4228 to peer 4377 (the
root node for O1’s identifier)
• As the message approaches the root node,
the object and the node identifiers become
increasingly similar
• In addition, the mapping (4378; 4228) is
stored at all peers along the message path

OceanStore
• OceanStore is a data management system designed to provide continuous
access to persistent information
• Relies on Tapestry and assumes untrusted powerful servers connected by
high-speed links
• To improve performance, data are allowed to be cached
anywhere, anytime
• Allows concurrent updates on replicated objects; it relies on reconciliation
to assure data consistency

Reconciliation in OceanStore
• Ri and ri denote, respectively, a primary and a
secondary copy of object R
• Nodes n1 and n2 are concurrently updating R
as follows
➡ Nodes that hold primary copies of R, called the
master group of R, are responsible for ordering
updates
➡ (a) n1 and n2 perform tentative updates on their
local secondary replicas and send these updates
to the master group of R as well as to other
random secondary replicas
➡ (b) The tentative updates are ordered by the
master group based on timestamps assigned by
n1 and n2, and epidemically propagated among
secondary replicas
➡ (c) Once the master group obtains an agreement,
the result of updates is multicast to secondary
replicas

Conclusion
• Advanced P2P applications will need high-level data management services
• Various P2P networks will improve
➡ Network-independence crucial to exploit and combine them
• Many technical issues
➡ Decentralized schema management, complex query processing, transaction
support and replication, and data privacy
• Important to characterize applications that can most benefit from P2P with
respect to other distributed architectures

Database ,16 P2P

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Database ,16 P2P (20)

More from Ali Usman (15)

Recently uploaded (20)

Database ,16 P2P