While Hadoop fits well in most batch processing workloads, and is the primary choice of big data processing today, it is not optimized for other types of workloads due to its following limitation:
- Lack of iteration support
- High latency due to persisting intermediate data onto disk
Nevertheless, the Map/Reduce processing paradigm is a proven mechanism for dealing with large scale data. On the other hand, many of Hadoop's infrastructure piece such as HDFS, HBase has been mature over time.
In this blog post, we'll look at a different architecture called Spark, which has taken the strength of Hadoop and made improvements in a number of Hadoop's weaknesses, and provides a more efficient batch processing framework with a much lower latency. Spark has generated a lot of excitement in the big data community and represents a very promising parallel execution stack for big data analytics.
Berkeley SparkWithin the Spark cluster, there is a driver program where the application logic execution is started, with multiple workers which processing data in parallel. Although this is not mandated, data is typically collocated with the worker and partitioned across the same set of machines within the cluster. During the execution, the driver program will passed code/closure into the worker machine where processing of corresponding partition of data will be conducted. The data will undergoing different steps of transformation while staying in the same partition as much as possible (to avoid data shuffling across machines). At the end of the execution, actions will be executed at the worker and result will be returned to the driver program.
Underlying the cluster, there is an important Distributed Data Structure called RDD (Resilient Distributed Dataset), which is a logically centralized entity but physically partitioned across multiple machines inside a cluster based on some notion of key. Controlling how different RDD are co-partitioned (with the same keys) across machines can reduce inter-machine data shuffling within a cluster. Spark provides a "partition-by" operator which create a new RDD by redistributing the data in the original RDD across machines within the cluster.
RDD can optionally be cached in RAM and hence providing fast access. Currently the granularity of caching is done at the RDD level, either the whole or none of the RDD is cached. Cached is a hint but not a guarantee. Spark will try to cache the RDD if sufficient memory is available in the cluster, based on LRU (Least Recent Use) eviction algorithm.
RDD provides an abstract data structure from which application logic can be expressed as a sequence of transformation processing, without worrying about the underlying distributed nature of the data.
Typically an application logic are expressed in terms of a sequence of TRANSFORMATION and ACTION. "Transformation" specifies the processing dependency DAG among RDDs and "Action" specifies what the output will be (ie: the sink node of the DAG with no outgoing edge). The scheduler will perform a topology sort to determine the execution sequence of the DAG, tracing all the way back to the source nodes, or node that represents a cached RDD.
Notice that dependencies in Spark come in two forms. "Narrow dependency" means the all partitions of an RDD will be consumed by a single child RDD (but a child RDD is allowed to have multiple parent RDDs). "Wide dependencies" (e.g. group-by-keys, reduce-by-keys, sort-by-keys) means a parent RDD will be splitted with elements goes to different children RDDs based on their keys. Notice that RDD with narrow dependencies preserve the key partitioning between parent and child RDD. Therefore RDD can be co-partitioned with the same keys (parent key range to be a subset of child key range) such that the processing (generating child RDD from parent RDD) can be done within a machine with no data shuffling across network. On the other hand, RDD will wide dependencies involves data shuffling. The scheduler will examine the type of dependencies and group the narrow dependency RDD into a unit of processing called a stage. Wide dependencies will span across consecutive stages within the execution and require the number of partition of the child RDD to be explicitly specified.
A typical execution sequence is as follows ...
- RDD is created originally from external data sources (e.g. HDFS, Local file ... etc)
- RDD undergoes a sequence of TRANSFORMATION (e.g. map, flatMap, filter, groupBy, join), each provide a different RDD that feed into the next transformation.
- Finally the last step is an ACTION (e.g. count, collect, save, take), which convert the last RDD into an output to external data sources
In a classical distributed system, fault resilience is achieved by replicating data across different machines together with a active monitoring system. In case of any machine crashes, there is always another copy of data residing in a different machine from where recovery can take place.
Fault resiliency in Spark takes a different approach. First of all, as a large scale compute cluster, Spark is not meant to be a large scale data cluster at all. Spark makes two assumptions of its workload.
- The processing time is finite (although the longer it takes, the cost of recovery after fault will be higher)
- Data persistence is the responsibility of external data sources, which keeps the data stable within the duration of processing.
Notice that the re-execution of lost partition is exactly the same as the lazy evaluation of the DAG, which starts from the leaf node of the DAG, tracing back the dependencies on what parent RDD is needed and then eventually track all the way to the source node. Recomputing the lost partition is done is a similar way, but taking partition as an extra piece of information to determine which parent RDD partition is needed.
However, re-execution across wide dependencies can touch a lot of parent RDD across multiple machines and may cause re-execution of everything. To mitigate this, Spark persist the intermediate data output from a Map phase before it shuffle them to different machines executing the reduce phase. In case of machine crash, the re-execution (from another surviving machine) just need to trace back to fetch the intermediate data from the corresponding partition of the mapper's persisted output. Spark also provide a checkpoint API to explicitly persist intermediate RDD so re-execution (when crash) doesn't need to trace all the way back to the beginning. In future, Spark will perform check-pointing automatically by figuring out a good balance between the latency of recovery and the overhead of check-pointing based on statistical result.
Spark provides a powerful processing framework for building low latency, massively parallel processing for big data analytics. It supports API around the RDD abstraction with a set of operation for transformation and action for a number of popular programming language like Scala, Java and Python.
In future posts, I'll cover other technologies in the Spark stack including real-time analytics using streaming as well as machine learning frameworks.