Hi, Alan: Thanks for you suggestion. I will take a look about MultipleOutput. But in this case, I still need to specify the schema in my driver, right? You mean I should use union schema in this case? But in my mapper, should I use SpecificRecord or GenericRecord? I can use (K,V) in my reducer, but in the mapper, I need the concrete Record object to serialize my data, right? Yong
From: [email protected] To: [email protected] Subject: RE: Question related to AvroJob.setMapOutputSchema(org.apache.hadoop.mapred.JobConf job, Schema s) Date: Mon, 30 Sep 2013 22:21:57 -0500 Hi Yong, It sounds like you might need to use AvroMultipleOutputs here. You can set all five of your output schemas in your driver, then route your message to the appropriate output in your reducer. See the following for mapred: http://avro.apache.org/docs/1.7.5/api/java/org/apache/avro/mapred/AvroMultipleOutputs.html And the following for mapreduce: http://avro.apache.org/docs/1.7.5/api/java/org/apache/avro/mapreduce/AvroMultipleOutputs.html If your mapper is generating the Avro records, then you will probably have to set AvroJob.setMapOutputSchema to a union of all five of your schemas. Thanks, Alan From: java8964 java8964 [mailto:[email protected]] Sent: Monday, September 30, 2013 9:37 PM To: [email protected] Subject: Question related to AvroJob.setMapOutputSchema(org.apache.hadoop.mapred.JobConf job, Schema s) Hi, I am new to user Avro. Currently, I am working on an existing project, and I want to see if using Avro makes sense. The project is to do some ETL around 5 data sets' data. The ETL logic is not complex, it will do different transferring logic for different data sets, and partition the data daily in reducer. There was one MR job to handle all 5 data sets originally. The data files have the name convention to distinguish the data sets. So in the mapper, it bases on the file name to understand what data set it is, and generate the key as "datase_name + date" to partition the data set first by data set, then daily. Now if I want to store the data in Avro format, it is straight-forward to write MR job for only one data set following a lot of online examples. I have no problem to change the MR job to store the data as Avro format for one data set. But if I still want to use one MR job for all 5 data sets, I got a problem. I tried both "SpecificRecord" and "GenericRecord", but I don't know how to solve this problem. For example, I created 5 avsc files for 5 data sets, and generate the Record object for all of them. But in the mapper/reducer, I don't want to specify any Record class, and this same mapper/reducer should be able to handle all data sets. So I try to put SpecificRecord class in my mapper/reducer, but in this case, I don't have the SpecificRecord.SCHEMA$ to use in my driver of AvroJob.setMapOutputSchema(conf, Schema), even though in my case, I really prefer the "SpecificRecord". So that makes me to try "GenericRecord". I change all my mapper and reducer to use "GenericRecord" class. But still, I don't know what schema I should use in my driver class for AvroJob.setMapOutputSchema(conf, Schema). The problem is that is there a generic abstract schema class I can use in AvroJob.setMapOutputSchema or AvroJob.setOutputSchema? My mapper class will correctly generate either "GenericRecord" or "SpecificRecord" class at runtime based on the file name, and reducer will write the correct "GenericRecord" or "SpecificRecord" object to the right output location without knowing the concrete Record object. But what stops me now is what kind of schema object I can use in AvroJob. I don't know during the driver stage what is my output schema, but the mapper/reducer will figure that out at runtime. Can I do this in Avro? Thanks Yong
