Professional Documents
Culture Documents
Hadoop
Hadoop
All the completed jobs will be maintained in HS cache. If the cache is full, it is unable to load the
new jobs/applications in the cache.
Solution:
3. Why Jobs fail in COMMIT stage with COMMIT_SUCCESS file exists exception?
This appears to be issue with speculative execution where different tasks create same
COMMIT_SUCCESS file for a single task.
Solution:
Rerun job with below properties
mapreduce.map.speculative=FALSE
mapreduce.reduce.speculative=FALSE
Tasks may fail if they don’t have enough memory to store its input data.
General scenario is either to increase number of Reducers. At times, even though we increase
number of Reducers, the data may be skewed to only few Reducers.
1
For ex:
If the total map output keys are 10000 and you are using 5 Reducers to process this data, and if
your data is not skewed uniformly like below, only 1st Reducer is taking complete load to process
this data and failing due to out of memory.
Solution:
Try to increase the Reducer memory and Java Opts property only for that particular job as
shown below. This will launch each Reducer with 6GB memory and tasks will be succeeded
without memory issue.
- Dmapreduce.reduce.memory.mb=6144
-Dmapreduce.reduce.java.opts=-Xmx4915m
At times tasks will take more time to complete due to disk latency.
Possible checks:
Check for any defunc processes running from long time. If found any, kill them.
ps -aef |grep -i defunc
7. No LoginModules Exception?
8. How to enable verbose property to know which jars are getting picked from where?
<property>
<name>mapreduce.map.java.opts</name>
<value> -Xmx512M -verbose:class </value>
</property>
export SPARK_SUBMIT_OPTS=-verbose:class
spark.driver.userClassPathFirst=true
spark.executor.userClassPathFirst=true
11. .
12. A
13. A
14. A
15. A
16. A
17. A