crunch-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Brush,Ryan" <RBR...@CERNER.COM>
Subject Re: Multiple output channels from Crunch DoFn
Date Tue, 20 Aug 2013 16:49:13 GMT
I happen to have some context around this, so I wanted to expand on Brandon's question a bit.
 The use case here is we're dealing with a large volume of third-party input and expect a
certain percentage of bogus or malformed data. Rather than simply logging instances of bad
records, we want to treat it as a signal we can learn from, both for improving our processing
logic and for creating structured reports we can use to troubleshoot data sources.

This leads to the "standard out" and "standard error" metaphors Brandon mentions: in most
cases, our Crunch DoFns would emit a processed structure useful downstream. But we'd also
like to be able to emit a structured error -- probably as an Avro object in our case -- and
persist that as a byproduct of our main processing pipeline. 

Would it make sense for such DoFn's to emit something some form of "Option" object? We could
then attach two consuming functions to it: one that handles the "success" case, sending the
resulting Avro object downstream. Another DoFn attached to the "Option" object would be a
no-op unless the Option contained an "error" structure, at which point we persist it to some
well-known location for later analysis.

I think this is entirely achievable using existing mechanisms...but it seems like common enough
use case (at least for us) to establish some idioms for dealing it.

On Aug 20, 2013, at 11:13 AM, Inman,Brandon wrote:

> 
> We've been looking at ways to do multiple outputs in Crunch jobs,
> specifically writing out some kind of Status or Error Avro object, based
> on failures that occur processing individual records in various jobs. It
> had been suggested that, rather than logging these errors to traditional
> loggers, to consider them an output of the Crunch job.  After some
> internal discussion, it was suggested to run the ideas past the Crunch
> community.
> 
> 
> A major goal we have is to end with all the error output in a location
> that makes it easy to run Hive queries or perform other MapReduce-style
> analysis to quickly view all errors across the larger system without the
> need go to multiple facilities.  This means standardizing on the Avro
> object, but it also necessitates decoupling the storage of the object from
> the "standard output" of the job.
> 
> 
> As Crunch DoFns support a single Emitter per invocation of process(), the
> solution that gathered the most support would be to emit an object similar
> to Pair<>, where first would be the "standard out" and second would be the
> "standard error".  A DoFn would generally only populate one (nothing
> preventing it from populating both if appropriate, but not really intended
> as a part of general use), and separate DoFns would filter out the two
> components of the pair and write the values to the appropriate targets.
> 
> As far as the emitted pairing object; the concept of a tagged union was
> suggested although there currently isn't support in Java or Avro for the
> concept; it was noted that
> https://issues.apache.org/jira/browse/CRUNCH-239 might be a close
> candidate. Pair<> would meet the requirements, although it was suggested
> that a simple object dedicated to the task could make a cleaner approach.
> 
> Any general thoughts on this approach? Are there any other patterns that
> might serve us better, or anything on the Crunch roadmap that might be
> more appropriate?
> 
> 
> Brandon Inman
> Software Architect
> www.cerner.com
> 
> 
> CONFIDENTIALITY NOTICE This message and any included attachments are from Cerner Corporation
and are intended only for the addressee. The information contained in this message is confidential
and may constitute inside or non-public information under international, federal, or state
securities laws. Unauthorized forwarding, printing, copying, distribution, or use of such
information is strictly prohibited and may be unlawful. If you are not the addressee, please
promptly delete this message and notify the sender of the delivery error by e-mail or you
may call Cerner's corporate offices in Kansas City, Missouri, U.S.A at (+1) (816)221-1024.



Mime
View raw message