I also prefer the exception. I recently found out about this behavior and am now facing a task of going back through my code base to explicitly add special error handling to detect this case.
On Wed, Jan 28, 2015 at 1:21 PM, David Whiting <davidwhit...@gmail.com> wrote: > I think "fail catastrophically" is probably exactly what should happen > here. You can always catch and use an empty iterable if it fails. A common > use case here is to do one step, materialize it into a collection or map, > then pass that into a DoFn to use as a small lookup table. This failure > mode means that future steps silently continue to execute with empty lookup > tables as part of their processing on the cluster. > > On 28 January 2015 at 13:45, Josh Wills <jwi...@cloudera.com> wrote: > > > Yeah, I think that before, we would just fail catastrophically by > throwing > > a CrunchRuntimeException, which I found annoying. Do you prefer that > > behavior? It's certainly something that could be configurable. > > > > J > > > > On Wed, Jan 28, 2015 at 10:36 AM, Jinal Shah <jinalshah2...@gmail.com> > > wrote: > > > > > I think it was intented from these commits I see here > > > > > > > > > https://github.com/apache/crunch/commit/3711cea61bded4c90b235a01163ae5f855089917 > > > and > > > > > > > > > https://github.com/apache/crunch/commit/ded504eb133fa0814e2d90ff2a662e72a67e04bb > > > . > > > Josh can enhance on this more. > > > > > > On Wed, Jan 28, 2015 at 9:26 AM, Mārtiņš Kalvāns < > > > martins.kalv...@gmail.com> > > > wrote: > > > > > > > Hi. > > > > > > > > When pipeline fails on cluster with some exception, materialize() > > returns > > > > empty collection and just logs error message. > > > > > > > > I'm (very, very) puzzled about this behaviour: > > > > > > > > > > > > > > https://github.com/apache/crunch/blob/master/crunch-core/src/main/java/org/apache/crunch/materialize/MaterializableIterable.java#L92 > > > > Is this really intended behaviour? > > > > > > > > If so, then some documentation for materialize() function about this > > > > behaviour would be really nice to have. :) > > > > > > > > > > > > -- > > > > Mārtiņš > > > > > > > > > > > > > > > -- > > Director of Data Science > > Cloudera <http://www.cloudera.com> > > Twitter: @josh_wills <http://twitter.com/josh_wills> > > >