What about communication errors and not corrupted files? Both when reading input and when writing output. We currently experience a failure of the entire process, if the last stage of writing the output (to Amazon S3) failed because of a very temporary DNS resolution issue (easily resolved by retrying).
*Romi Kuntsman*, *Big Data Engineer* http://www.totango.com On Wed, Apr 1, 2015 at 12:58 PM, Gil Vernik <g...@il.ibm.com> wrote: > I actually saw the same issue, where we analyzed some container with few > hundreds of GBs zip files - one was corrupted and Spark exit with > Exception on the entire job. > I like SPARK-6593, since it can cover also additional cases, not just in > case of corrupted zip files. > > > > From: Dale Richardson <dale...@hotmail.com> > To: "dev@spark.apache.org" <dev@spark.apache.org> > Date: 29/03/2015 11:48 PM > Subject: One corrupt gzip in a directory of 100s > > > > Recently had an incident reported to me where somebody was analysing a > directory of gzipped log files, and was struggling to load them into spark > because one of the files was corrupted - calling > sc.textFiles('hdfs:///logs/*.gz') caused an IOException on the particular > executor that was reading that file, which caused the entire job to be > cancelled after the retry count was exceeded, without any way of catching > and recovering from the error. While normally I think it is entirely > appropriate to stop execution if something is wrong with your input, > sometimes it is useful to analyse what you can get (as long as you are > aware that input has been skipped), and treat corrupt files as acceptable > losses. > To cater for this particular case I've added SPARK-6593 (PR at > https://github.com/apache/spark/pull/5250). Which adds an option > (spark.hadoop.ignoreInputErrors) to log exceptions raised by the hadoop > Input format, but to continue on with the next task. > Ideally in this case you would want to report the corrupt file paths back > to the master so they could be dealt with in a particular way (eg moved to > a separate directory), but that would require a public API > change/addition. I was pondering on an addition to Spark's hadoop API that > could report processing status back to the master via an optional > accumulator that collects filepath/Option(exception message) tuples so the > user has some idea of what files are being processed, and what files are > being skipped. > Regards,Dale. >
