hadoop 迭代消耗大 每次迭代启动一个完整的MapReduce作业
spark 首要目标就是避免运算时 过多的网络和磁盘IO开销
Resilient Distributed Datasets
http://www.cs.cmu.edu/~pavlo/courses/fall2013/static/slides/spark.pdf
Resilient Distributed Datasets
Presented by Henggang Cui
15799b Talk
1
Why not MapReduce
• Provide fault-tolerance, but:
• Hard to reuse intermediate results across
multiple computations
– stable storage for sharing data across jobs
• Hard to support interactive ad-hoc queries
2
Why not Other In-Memory Storage
• Examples: Piccolo
– Apply fine-grained updates to shared states
• Efficient, but:
• Hard to provide fault-tolerance
– need replication or checkpointing
3
Resilient Distributed Datasets (RDDs)
• Restricted form of distributed shared memory
– read-only, partitioned collection of records
– can only be built through coarse‐grained
deterministic transformations
• data in stable storage
• transformations from other RDDs.
• Express computation by
– defining RDDs
4
Fault Recovery
• Efficient fault recovery using lineage
– log one operation to apply to many elements
(lineage)
– recompute lost partitions on failure
5
Example
lines = spark.textFile("hdfs://...")
errors = lines.filter(_.startsWith("ERROR"))
hdfs_errors = errors.filter(_.contains(“HDFS"))
6
Advantages of the RDD Model
• Efficient fault recovery
– fine-grained and low-overhead using lineage
• Immutable nature can mitigate stragglers
– backup tasks to mitigate stragglers
• Graceful degradation when RAM is not
enough
7
Spark
• Implementation of the RDD abstraction
– Scala interface
• Two components
– Driver
– Workers
8
• Driver
– defines and invokes actions on RDDs
– tracks the RDDs’ lineage
• Workers
– store RDD partitions
– perform RDD
transformations
Spark Runtime
9
Supported RDD Operations
• Transformations
– map (f: T->U)
– filter (f: T->Bool)
– join()
– ... (and lots of others)
• Actions
– count()
– save()
– ... (and lots of others)
10
Representing RDDs
• A graph-based representation for RDDs
• Pieces of information for each RDD
– a set of partitions
– a set of dependencies on parent RDDs
– a function for computing it from its parents
– metadata about its partitioning scheme and data
placement
11
RDD Dependencies
• Narrow dependencies
– each partition of the parent RDD is used by at
most one partition of the child RDD
• Wide dependencies
– multiple child partitions may depend on it
12
RDD Dependencies
13
RDD Dependencies
• Narrow dependencies
– allow for pipelined execution on one cluster node
– easy fault recovery
• Wide dependencies
– require data from all parent partitions to be
available and to be shuffled across the nodes
– a single failed node might cause a complete reexecution.
14
Job Scheduling
• To execute an action on an RDD
– scheduler decide the stages from the RDD’s
lineage graph
– each stage contains as many pipelined
transformations with narrow dependencies as
possible
15
Job Scheduling
16
Memory Management
• Three options for persistent RDDs
– in-memory storage as deserialized Java objects
– in-memory storage as serialized data
– on-disk storage
• LRU eviction policy at the level of RDDs
– when there’s not enough memory, evict a
partition from the least recently accessed RDD
17
Checkpointing
• Checkpoint RDDs to prevent long lineage
chains during fault recovery
• Simpler to checkpoint than shared memory
– Read-only nature of RDDs
18
Discussions
19
Checkpointing or Versioning?
20
• Frequent checkpointing, or
Keep all versions of ranks?