I am also thinking about this for my current project, so here I share some of my thoughts, but maybe some of them are not correct.
1) In my previous projects years ago, we store a lot of data as plain text, as at that time, people thinks the Big data can store all the data, no need to worry about space issues, until we run out of space in the cluster very fast :-). So lesson number 1, don't store them as text file.2) To compress text file, we need a container for the file, like 'Seq file', 'Avro' or 'proto buf'. I didn't use RC/ORC before, so it will be interested to know more about them later.3) I did a benchmark before, for the data sets we are using, there is also a webpage about the benchmark result. You can google it. I believe the performance from them are close. The real questions are: a) Language supports b) Flexible of serialization format c) How easy it can be used in tools like 'pig/hive' etc. d) How good it supported in hadoop.
>From my experience, sequence file is not good supported outside of Java language, and it is just a key/value storage, if your data have nest structure data, like your XML/JSON data, you still need a serialization format like google protobuf or Avro to handle it. Store directly XML/JSON in HDFS is really not a good idea. As any InputFormat to support split for them them all requires strict format of the data, and compression won't work very nicely on these kind of data.
We originally used google protobuf a lot, as twitter releases the elephant-bird as open source to support it in hadoop. It is a big plus for it at that time. But recently, we also start to consider Avro seriously now, as it is better supported directly in hadoop. I also like its schema-less vs schema objects both options design. It gives us some flexibility in designing MR jobs.
> From: [EMAIL PROTECTED]
> To: [EMAIL PROTECTED]
> Subject: File formats in Hadoop: Sequence files vs AVRO vs RC vs ORC
> Date: Mon, 30 Sep 2013 09:40:44 +0200
> the file format topic is still confusing me and I would appreciate if you
> could share your thoughts and experience with me.
> From reading different books/articles/websites I understand that
> - Sequence files (used frequently but not only for binary data),
> - AVRO,
> - RC (was developed to work best with Hive -columnar storage) and
> - ORC (a successor of RC to give Hive another performance boost - Stinger
> are all container file formats to solve the "small files problem" and all
> support compression and splitting.
> Additionally, each file format was developed with specific features/benefits
> in mind.
> Imagine I have the following text source data
> - 1 TB of XML documents (some millions of small files)
> - 1 TB of JSON documents (some hundred thousands of medium sized files)
> - 1 TB of Apache log files (some thousands of bigger files)
> How should I store this data in HDFS to process it using Java MapReduce and
> Pig and Hive?
> I want to use the best tool for my specific problem - with "best"
> performance of course - i.e. maybe one problem on the apache log data can be
> best solved using Java MapReduce, another one using Hive or Pig.
> Should I simply put the data into HDFS as the data comes from - i.e. as
> plain text files?
> Or should I convert all my data to a container file format like sequence
> files, AVRO, RC or ORC?
> Based on this example, I believe
> - the XML documents will be need to be converted to a container file format
> to overcome the "small files problem".
> - the JSON documents could/should not be affected by the "small files
> - the Apache files should definitely not be affected by the "small files
> problem", so they could be stored as plain text files.
> So, some source data needs to be converted to a container file format,
> others not necessarily.
> But what is really advisable?
> Is it advisable to store all data (XML, JSON, Apache logs) in one specific