1-Introduction MapReduce has emerged as a popular way in working on large clusters The Google MapReduce framework and open-source Hadoop system Data-centric fashion: transformation to data sets distributed execution, network communication and fault tolerance are handled by MR
1-Introduction Pipelining provides several important advantages to a MapReduce framework: -map->reduce => generate & refine an approximation of their ﬁnal answer during the course of execution =online aggregation -MapReduce jobs(that run continuosly) accepting new data, as it arrives, analyse it immediately. This allows MR to be used in applications such as event monitoring and stream processing -pipelining can reduce job completion times by up to 25% in some scenarios.
1-Introduction We present a modiﬁed version of the Hadoop MapReduce framework that supports on-line aggregation, which allows users to see “early returns”
2-Hadoop Background 2.1 Programming Model -mapping, [intermediate pairs.] reducing -Optionally, user can use Combiner Function (map-side pre-aggregation). Reduces network traffic btw map-reduce
2-Hadoop Background 2.2 Hadoop Architecture -Hadoop MapReduce + Hadoop Distributed File System (HDFS) -HDFS is used to store both inputs to map and reduce steps(Intermediate results are stored in node’s local file system) -Hadoop installation =single master node and many worker nodes -Master Node - JobTracker (accept, divide, give) -Worker Node - TaskTracker (manages execution of tasks. By default is has 2 maps & 2 reduces slots)
2-Hadoop Background 2.3 Map Task Execution -Each map task is assigned a portions of input files (splits) -By default split contains a single HDFS block (in 64MB). number of file blocks = number of map tasks. -Execution of map tasks is divided into 2 phases: -1. The map phase : : Reads the task’s splits from HDFS, parses it, and applies map function to each record -2. The commit phase : After that the commit phase registers the final output with the TaskTracker, (informs JobTracker about completing of a task) -outputCollector function (in the map phase ) stores map output in a format that is easy for reduce tasks to consume.
2-Hadoop Background 2.4 Reduce Task Execution is divided into three phases: -Shufﬂe, sort, reduce The output of the reduce function is written to a temporary location on HDFS. After comlpeting, the task’s HDFS output ﬁle is atomically renamed from its temporary location to its ﬁnal location. If there is a failure in map or reduce task execution??
2-Hadoop Background Map tasks write their output to local disk – Output available after map task has completed Reduce tasks write their output to HDFS – Once job is finished, next job’s map tasks can be scheduled, and will read input from HDFS Therefore, fault tolerance is simple: simply re- run tasks on failure – No consumers see partial operator output
Dataflow in Hadoop Submit job schedule map reduce
3- Pipelined MapReduce 3.1 Pipelining within a job. 3.1.1 Naïve Pipelining -we modiﬁed Hadoop to send data directly from map to reduce tasks. -Client submit jobs-> JobTracker assigns map&reduce tasks to the available TaskTracker slots.. -TCP socket will be used to pipeline the output of the map function. As soon as map-output is produced, Mapper determines where(which partition in reduce task) the record should be sent to.
3- Pipelined MapReduce 3.1.2 Refinements Naïve Pipelining may suffer from several practical problems: -Prblm1. There may not be enough slots available to schedule every task in a new job and large number of TCP connection is needed. => Map Task write output to the disk. Once the reduce task assign a slot, then it can pull the records from the map task. For TCP problems: each reducer can be configurable. (pull the data from a certain number of mappers at once) -Prblm2. The map function was invoked by the same thread that wrote output records to the pipeline sockets. i.e. If the network is truncated the mapper will be prevented from doing useful work. => separate thread: stores its output in an in-memory buffer, then another one sends these data to the connected reducers
3- Pipelined MapReduce 3.1.3 Granularity of Map Output Another problem with the naïve design is that it eagerly sends each record as soon as it is produced, which prevents the use of map- side combiners. => Instead of sending the buffer contents to reducers directly, we wait for the buffer to grow to a threshold size. The mapper then applies the combiner function and writes the buffer to disk using the spill ﬁle format. -When a map task generates a new spill ﬁle, it ﬁrst queries the TaskTracker for the number of unsent spill ﬁles. If this number grows beyond a certain threshold. mapper will accumulate multiple spill ﬁles. -Once the queue of unsent spill ﬁles exceeds the threshold, the map task merges and combines the accumulated spill ﬁles into a single ﬁle, and then registers its output with the TaskTracker.
3- Pipelined MapReduce 3.2 Pipelining Between Jobs In the traditional Hadoop architecture, the output of each job is written to HDFS. (j1, j2) Furthermore, the JobTracker cannot schedule a consumer job until the producer job has completed, because scheduling a map task requires knowing the HDFS block locations of the map’s input split. => In our modiﬁed version of Hadoop, the reduce tasks of one job can optionally pipeline their output directly to the map tasks of the next job. And we’ll introduce ‘snapshot’ outputs that is publishd by online aggregation and continuous queries.
3- Pipelined MapReduce FAULT TOLERANCE Traditional fault tolerance algorithms for pipelined dataflow systems are complex HOP approach: write to disk and pipeline – Producers write data into in-memory buffer – In-memory buffer periodically spilled to disk – Spills are also sent to consumers – Consumers treat pipelined data as “tentative” until producer is known to complete – Fault tolerance via task restart, tentative output discarded
4- Online Aggregation Although MapReduce was originally designed as a batch-oriented system, it is often used for interactive data analysis. (examle) an interactive user would prefer a “quick and dirty” approximation over a correct answer that takes much longer to compute. How we extended our pipelined Hadoop implementation to support online aggregation within a single job (Section 4.1) and between multiple jobs (Section 4.2).
4- Online Aggregation 4.1 Single-Job Online Aggregation In HOP, the data records produced by map tasks are sent to reduce tasks shortly after each record is generated. Snapshot is an output of the reduce task at a certain time. It is important for us, how correct a snapshot is. & how does snapshot coincide with the correct data. It is a hard problem..
4- Online Aggregation Snapshots are computed periodically, as new data arrives at each reducer. The user may -speciﬁy how often snapshots should be computed - specify whether to include data from tentative (unﬁnished) map tasks if there are not enough free slots to allow all the reduce tasks in a job to be scheduled, snapshots will not be available for reduce tasks that are still waiting to be executed Within a single job: periodically invoke reduce function at each reduce task on available data Between jobs: periodically send a “snapshot” to consumer jobs
4- Online Aggregation 4.1 Multi-Job Online Aggregation Similar to the single-job Online Aggregation, but approximate answers are pipelined to map tasks of next job. (j1, j2, …) Unfortunately output of the reduce function is not monotonic. Why that co-scheduling a sequence of jobs is required. Consumer job computes an approximation
4- Online Aggregation Fault Tolerance for Multi-Job Online Aggregation: Let’s assume we have 2 jobs (j1,j2)… We suppose 3 cases: 1. Task in j1 fails (we discussed earlier) 2. Task in j2 fails (system restarts the failed task) 3. To handle failures in j1, we replace the most recent snapshot from j2 to j1, with the failed one. If tasks from both jobs fail, a new task in j2 recovers the most recent snapshot from j1.
Stream Processing MapReduce is often applied to streams of data that arrive continuously – Click streams, network traffic, web crawl data, … Traditional approach: buffer, batch process 1.Poor latency 2.Analysis state must be reloaded for each batch Instead, run MR jobs continuously, and analyze data as it arrives
Monitoring The thrashing host was detected very rapidly—notably faster than the 5-second TaskTracker- JobTracker heartbeat cycle that is used to detect straggler tasks in stock Hadoop. We envision using these alerts to do early detection of stragglers within a MapReduce job.
5- Conclusion HOP extends the applicability of the model to pipelining behaviors, while preserving the simple programming model and fault tolerance of a full- featured MapReduce framework. Future topics -Scheduling -using MapReduce-style programming for even more interactive applications.