Details
-
New Feature
-
Status: Open
-
Major
-
Resolution: Unresolved
-
0.20.3
-
None
-
None
-
Incompatible change
-
Job Submission
Description
Currently the job submission protocol requires the job provider to put every bit of information inside an instance of JobConf. The submitted information includes the input data (hdfs path) , suspected resource requirement, number of reducers etc. This information is read by JobTracker as part of job initialization. Once initialized, job is moved into a running state. From this point, there is no mechanism for any additional information to be fed into Hadoop infrastructure for controlling the job execution.
The execution pattern for the job looks very much static from this point. Using the size of input data and a few settings inside JobConf, number of mappers is computed. Hadoop attempts at reading the whole of data in parallel by launching parallel map tasks. Once map phase is over, a known number of reduce tasks (supplied as part of JobConf) are started.
Parameters that control the job execution were set in JobConf prior to reading the input data. As the map phase progresses, useful information based upon the content of the input data surfaces and can be used in controlling the further execution of the job. Let us walk through some of the examples where additional information can be fed to Hadoop subsequent to job submission for optimal execution of the job.
I) "Process a part of the input , based upon the results decide if reading more input is required "
In a huge data set, user is interested in finding 'k' records that satisfy a predicate, essentially sampling the data. In current implementation, as the data is huge, a large no of mappers would be launched consuming a significant fraction of the available map slots in the cluster. Each map task would attempt at emitting a max of 'k' records. With N mappers, we get N*k records out of which one can pick any k to form the final result.
This is not optimal as:
1) A larger number of map slots get occupied initially, affecting other jobs in the queue.
2) If the selectivity of input data is very low, we essentially did not need scanning the whole of data to form our result.
we could have finished by reading a fraction of input data, monitoring the cardinality of the map output and determining if
more input needs to be processed.
Optimal way: If reading the whole of input requires N mappers, launch only 'M' initially. Allow them to complete. Based upon the statistics collected, decide additional number of mappers to be launched next and so on until the whole of input has been processed or enough records have been collected to for the results, whichever is earlier.
II) "Here is some data, the remaining is yet to arrive, but you may start with it, and receive more input later"
Consider a chain of 2 M-R jobs chained together such that the latter reads the output of the former. The second MR job cannot be started until the first has finished completely. This is essentially because Hadoop needs to be told the complete information about the input before beginning the job.
The first M-R has produced enough data ( not finished yet) that can be processed by another MR job and hence the other MR need not wait to grab the whole of input before beginning. Input splits could be supplied later , but ofcourse before the copy/shuffle phase.
III) " Input data has undergone one round of processing by map phase, have some stats, can now say of the resources
required further"
Mappers can produce useful stats about of their output, like the cardinality or produce a histogram describing distribution of output . These stats are available to the job provider (Hive/Pig/End User) who can
now determine with better accuracy of the resources (memory requirements ) required in reduction phase, and even the number of reducers or may even alter the reduction logic by altering the reducer class parameter.
In a nut shell, certain parameters about a job are governed by the input data and the intermediate results produced and hence need to be overridden as job progresses. Hadoop does not allow such information to be fed dynamically. Hence job execution may not always be optimal.
I would like to get feedback from the Hadoop community about the above proposal and if any similar effort is already underway.
If we agree, as a next step I would like to discuss the implementation details that I have worked out end-to-end.
