I personally try to steer people away from multi-line input formats b/c of how tedious they are to write/maintain. To me, the question of supporting CSVs maps to a more general question about whether we should support some kind of named Record/Row type for processing data from CSV/Hive/Avro/PB/Thrift/etc. in a generic way. I could make arguments either way, which I'm happy to do if folks are interested, but I'd rather hear from other people first, esp. if anyone feels strongly about it.
J On Mon, Mar 18, 2013 at 3:14 AM, Christian Tzolov < christian.tzo...@gmail.com> wrote: > Hi, > > I am working on ETL projects that consume and produce data in the RFC4180 > [1] CSV format. Although unreliable IMO, this RFC is used as an exchange > format by several Dutch government agencies. > > The RFC4180 spec supports multi-line fields (e.g. fields with line > breaks) and escaping of double quotes and delimiters within fields. Because > of the multi-line feature one can't use directly the > FileInputFormat/TextInputFormat or LineRecordReader implementations. > Furthermore as I see it the input splitting must be disabled (not sure if > any efficient splitting strategy is possible at all). > > There are several java libraries that provide some RFC4180 support [3]. For > Pig a slightly modified CSVExcelStorage UDF [2] seems to do the job (not > sure about the input splitting though). Also the "Hadoop in Practice" > example [4] does not support the multi-line fields. > > Has someone used similar 'multi-line fields' formats? I wonder how common > is this use case. > > Also shall we provide support for it in Crunch? > > Cheers, > Chris > > [1] RFC 4180 - http://tools.ietf.org/html/rfc4180 > [2] PIG CVSExcelStorage UDF - > > http://svn.apache.org/repos/asf/pig/trunk/contrib/piggybank/java/src/main/java/org/apache/pig/piggybank/storage/CSVExcelStorage.java > [3] jCSV, OpenCSV, SuperCSV > [4] > > https://github.com/alexholmes/hadoop-book/blob/master/src/main/java/com/manning/hip/ch3/csv/CSVInputFormat.java > -- Director of Data Science Cloudera <http://www.cloudera.com> Twitter: @josh_wills <http://twitter.com/josh_wills>
