Error handling in hadoop map reduce

Question

Based on the documentation, there are a few ways, how the error handling is performed in map reduce. Below are the few:

a. Custom counters using enum - increment for every failed record.

b. Log error and analyze later.

Counters give the number of failed records. However to get the identifier of the failed record(may be its unique key), and details of the exception occurred, node on which the error occurred - we need to perform centralized log analysis and there are many nodes running. Logstash is on which is available.

Apart from these, are there any other ways to handle the error scenarios, without manual intervention. Any tools, references, best practices are welcome.

I think the same technique applies to any distributed applications, with minor changes.

score 3 · Accepted Answer · answered Nov 03 '15 at 21:15

Few questions to ask, when working with error handling:

Should the job be stopped if an error occurred in data validation. Most of the Big data use cases might be ok to leave few bad records. But if your usecase wants all the records to be good enough, you should take that decision and move to the below steps.

Some times its better to let the job run by skipping the bad records or and in parallel, get the issues(errors) using below techniques, rectify and modify as you move along.
You want the errors to be occurred, but only limited times. Then how many times an exception can be thrown, before the entire job gets stopped is as below

For Map tasks: mapreduce.map.maxattempts property

For reducer tasks: mapreduce.reduce.maxattempts

Default is 4
Handling malformed data.

So we decided to handle the malformed data. Then define the condition or which the record is bad. You can use counters, to quickly give you the number of bad records.

In Mapper class,
```
enum Temperature { OVER_10 }
```
Inside map method,

//parse the record
```
if(value > 10) {
    System.err.println("Temperature over 100 degrees for input: " + value);
    context.setStatus("Detected possibly corrupt record: see logs.");
    context.getCounter(Temperature.OVER_10).increment(1);      
}
```
With the above method, all records get processed, and the counters get added based on the bad records. You can see the counter value, at the end of the job, after job statistics or through web UI or from shell command.
```
$mapred job -counter <job_id> '${fully_qualified_class_name}' ${enum_name}
$mapred job -counter job_1444655904448_17959 'com.YourMapper$Temperature' OVER_10
```
Once you know the impact of the problem i.e number of bad records, we need to know "why is it bad". For this, we need to go to the logs and search for the error messages.

Yarn provide log aggregation and combines all the logs for a job id and stores in hdfs. It can be get using
```
yarn logs -applicationId <application ID>
```

Error handling in hadoop map reduce

1 Answers1

Linked