Hadoop in SIGMOD 2011
Nova: Continuous Pig/Hadoop Workflows
Apache Hadoop Goes Realtime at Facebook
Emerging Trends in the Enterprise Data Analytics
A Hadoop Based Distributed Loading Approach to
Parallel Data Warehouses
Industrial Session in Sigmod 2011
Data Management for
Feeds and Streams(2)
Applying Hadoop
Dynamic Optimization and
Unstructured Content (4)
Support for Business Analytics
and Warehousing (4)
Nova: Continuous Pig/Hadoop Workflows
By Yahoo!
Nova Overview
 Ingesting and analyzing user behavior logs
 Building and updating a search index from a stream of crawled web
 Processing semi-structured data
Two-layer programming model (Nova over Pig)
 Continuous processing
 Independent scheduling
 Cross-module optimization
 Manageability features
Workflow Model
 Two kinds of vertices: tasks (processing
steps) and channels (data containers)
 Edges connect tasks to channels and channels
to tasks
Four common patterns of processing
 Non-incremental (template detection)
 Stateless incremental (shingling)
 Stateless incremental with lookup table
(template tagging)
 Stateful incremental (de-duping)
Workflow Model (Cont.)
Data and Update Model
 Blocks: A channel’s data is divided into blocks
Base block
Contains a complete snapshot of data on a
channel as of some point in time
Base blocks are assigned increasing sequence
Delta block
Used in conjunction with incremental
Contains instructions for transforming a base
block into a new base block( i  j  Bi  B j (i  j ) )
Workflow Model (Cont.)
Task/Data Interface
 Consumption mode: all or new
 Production mode: B or Δ
Workflow Model (Cont.)
Workflow Programming and Scheduling
 Data-based trigger.
 Time-based trigger
 Cascade trigger.
Data Compaction and Garbage Collection
 If a channel has blocks B0,01 ,12 ,  23 ,the
compaction operation computes and adds B3 to the channel
 After compaction is used to add B3 to the channel,and current
cursor is at sequence number 2, then B0,01 ,
can be garbage-collected.
Nova System Architecture
Apache Hadoop Goes Realtime at Facebook
By Facebook
Workload Types
 Facebook Messaging
 High Write Throughput
 Large Tables
 Data Migration
 Facebook Insights
 Realtime Analytics
 High Throughput Increments
 Facebook Metrics System (ODS)
 Automatic Sharding
 Fast Reads of Recent Data and Table Scans
Why Hadoop & HBase
High write throughput
Efficient and low-latency strong consistency semantics within
a data center
Efficient random reads from disk
High Availability and Disaster Recovery
Fault Isolation
Atomic read-modify-write primitives
Range Scans
Tolerance of network partitions within a single data center
Zero Downtime in case of individual data center failure
Active-active serving capability across different data centers
Realtime HDFS
High Availability - AvatarNode
Realtime HDFS (Cont.)
Hadoop RPC compatibility
Block Availability: Placement Policy
 a pluggable block placement policy
Realtime HDFS (Cont.)
Performance Improvements for a Realtime Workload
 RPC Timeout
 Reads from Local Replicas
New Features
 HDFS sync
 Concurrent Readers
Production HBase
ACID Compliance (RWCC: Read Write Consistency Control)
 Atomicity (WALEdit)
 Consistency
Availability Improvements
 HBase Master Rewrite,Region assignment in memory -> ZooKeeper
 Online Upgrades
 Distributed Log Splitting
Performance Improvements
 Compaction(minor and major)
 Read Optimizations
Emerging Trends in the Enterprise Data Analytics:
Connecting Hadoop and DB2 Warehouse
1.Increasing volumes of data
2. Hadoop-based solutions in conjunction with
data warehouses
A Hadoop Based Distributed Loading Approach to
Parallel Data Warehouses
By Teradata
 ETL(Extraction Transformation Loading) is a critical
part of data warehouse
 While data are partitioned and replicated across all
nodes in a parallel data warehouse, load utilities reside
on a single node(bottleneck)
Why Hadoop for Teradata EDW(Enterprise Data Warehouse)?
 More disk space can be easily added
 Use as a intermediate storage
 MapReduce for transformation
 Load data in parallel
Block Assignment Problem
HDFS file F on a cluster of P nodes (each node is uniquely
identified with an integer i where 1 ≤ i ≤ P)
The problem is defined by: assignment(X, Y, n,m, k, r)
 X is the set of n blocks (X = {1, . . . , n}) of F
 Y is the set of m nodes running PDBMS (called PDBMS nodes)
(Y⊆ {1, . . . , P })
 k copies, m nodes
 r is the mapping recording the replicated block locations of
each block. r(i) returns the set of nodes which has a copy of the
block i.
Block Assignment Problem(Cont.)
• An assignment g from the blocks in X to the nodes in Y is
denoted by a mapping from X = {1, . . . , n} to Y where g(i)
= j (i ∈ X, j ∈ Y ) means that the block i is assigned to the
node j.
An even assignment g is an assignment such that ∀ i ∈ Y ∀
j ∈ Y | |{ x | ∀ 1 ≤ x ≤ n&&g(x) = i}| - |{y | ∀ 1 ≤ y ≤ n&&g(y)
= j}| | ≤ 1.
The cost of an assignment g is defined to be cost(g) = |{i |
g(i) r(i) ∀ 1 ≤ i ≤ n}|, which is the number of blocks
assigned to remote nodes.
Thank You!