crunch-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Gabriel Reid <>
Subject Re: Multiple output channels from Crunch DoFn
Date Wed, 21 Aug 2013 06:50:18 GMT
I like the idea of having a generic way of writing to multiple collections,
or branching outputs. I think that internally it's as easy as just writing
to pairs or tuples (or maybe unions if they're available) as Chao
suggested, and I'm thinking that we could make even make it a bit more user
friendly so that client code doesn't even need to know about the fact that
things are being written to a pair in the underlying implementation.

The idea I've got in my head is of a client api something like this
(although I'm sure I'm forgetting some details):

    static enum Output { STDOUT, STDERR };

    PCollection<String> inputCollection = ....

    Branch branch = Branch.newBuilder()
        .addOutput(Output.STDOUT, Avros.strings())
        .addOutput(Output.STDERR, Avros.records(MyErrorClass.class))

    BranchResult branchResult = branch.execute(
        new BranchDoFn() {
            public void process(String input, BranchEmitter emitter) {
                try {
                    String transformed = transform(input);
                    emitter.emit(Output.STDOUT, transformed);
                } catch (Exception e) {
                    emitter.emit(Output.STDERR, new MyErrorClass(e));

    PCollection<String> output = branchResult.getOutput(Output.STDOUT);
    PCollection<MyErrorClass> errors =

The general idea is that the Output enumeration could be any enum, with any
number of values,
and the underlying implementation would use sparse tuples and then null
filters to create the output
PCollections. I'm sure there are a lot of details that would need to still
be worked out (particularly around
generics), but this is a use case that I often run into as well, and I
think it would be good to have
a pretty simple branching model like this so that it's (hopefully) easy to
see what's going on.

Any thoughts on this approach?

- Gabriel

On Wed, Aug 21, 2013 at 5:45 AM, Josh Wills <> wrote:

> I think that we try to disallow it in general, so it's probably okay.
> On Tue, Aug 20, 2013 at 8:43 PM, Chao Shi <> wrote:
>> Yes, I think so. Do we generally allow nulls in crunch APIs? I'm a afraid
>> that it would be confusing if some excludes null values while others don't.
>> 2013/8/21 Josh Wills <>
>>> ...with the assumption that we would exclude null values in the Pair<A,
>>> B>?
>>> On Tue, Aug 20, 2013 at 7:40 PM, Josh Wills <> wrote:
>>>> That does sound pretty clean...
>>>> On Tue, Aug 20, 2013 at 7:34 PM, Chao Shi <> wrote:
>>>>> Is it possible to provide a utility that transforms
>>>>> PCollection<Pair<A, B>> to Pair<PCollection<A>,
PCollection<B>>? So one can
>>>>> simply emit Pairs and then write them to two Targets. This could be
>>>>> generalized to Tuples.
>>>>> 2013/8/21 Josh Wills <>
>>>>>> On Tue, Aug 20, 2013 at 12:23 PM, Inman,Brandon <
>>>>>>> wrote:
>>>>>>>  I like the flexibility of this approach, although would the
>>>>>>> of having some official constants defined for a small set of
>>>>>>> channels be reasonable (the concepts of "out" and "error" are
>>>>>>> common, others may be warranted as well)?
>>>>>> So I think the way I would handle this would be having a main output
>>>>>> directory and an error output directory that was underneath it. Cascading
>>>>>> does this trick within their existing flows where you can throw exceptions
>>>>>> to "traps," which is essentially the same idea, though I'm not wild
>>>>>> control flow that relies on throwing exceptions.
>>>>>>>  Is this something that you would see being added to core Crunch
>>>>>>> APIs (for example, directly to Pipeline), or implemented on top
of Crunch
>>>>>>> with a filtering approach similar to my original post?  If it's
>>>>>>> on top, shouldn't materialization work as-is?
>>>>>> Yes, your model would be simpler. I think that mine would require
>>>>>> special kind of Target implementation, a custom implementation of
>>>>>> Emitter interface that would be used for routing the outputs of the
>>>>>> and possibly some post-processing code to move the data to a sensible
>>>>>> place. I don't know if that work is strictly necessary, and your
impl is
>>>>>> certainly much more straightforward than mine. :)
>>>>>>>  If the type was PTable<String, T>, could Union<S,U>
be a choice
>>>>>>> for T as appropriate? In our case, we would likely be looking
at a
>>>>>>> PTable<String, T extends SpecificRecordBase> and not necessarily
need Union
>>>>>>> with this approach.
>>>>>> Yeah, I think it would be fine, but we'd have to be cognizant of
>>>>>> when we were implementing the union type, and it would be up to the
>>>>>> to ensure that the right data type ended up in the right file, which
>>>>>> maybe less good?
>>>>>>>   From: Josh Wills <>
>>>>>>> Reply-To: "" <>
>>>>>>> Date: Tuesday, August 20, 2013 1:00 PM
>>>>>>> To: "" <>
>>>>>>> Subject: Re: Multiple output channels from Crunch DoFn
>>>>>>>   A related idea that has come up a few times has been the idea
>>>>>>> having a way of writing values to different files based on a
key: some kind
>>>>>>> of generalization of Target that would itself write multiple
outputs under
>>>>>>> the covers, with the name of the output file indicated by some
function of
>>>>>>> the key of the PTable.
>>>>>>> For this situation, we would have a PTable that was like
>>>>>>> PTable<String, Union<S, T>>, or just PTable<String,
T> if the output types
>>>>>>> were all the same, and the String would specify the name of an
>>>>>>> directory (that I suppose would live underneath some base output
>>>>>>> for the Target) that the record would be written to.
>>>>>>>  There are a couple of limitations to this approach, I think:
>>>>>>> couldn't consider this kind of PTable "materialized" w/o doing
an overhaul
>>>>>>> of the materialization logic-- it would act sort of like an HTableTarget
>>>>>>> that it would be write-only in flows. There are probably some
others I
>>>>>>> can't think of off the top of my head. What do you guys think?
>>>>>>>  J
>>>>>>> On Tue, Aug 20, 2013 at 9:49 AM, Brush,Ryan <>wrote:
>>>>>>>> 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
>>>>>>>> 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
>>>>>>>> object, based
>>>>>>>> > on failures that occur processing individual records
in various
>>>>>>>> jobs. It
>>>>>>>> > had been suggested that, rather than logging these errors
>>>>>>>> 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
>>>>>>>> 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
>>>>>>>> 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
>>>>>>>> (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
>>>>>>>> ><>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
>>>>>>>> patterns that
>>>>>>>> > might serve us better, or anything on the Crunch roadmap
>>>>>>>> might be
>>>>>>>> > more appropriate?
>>>>>>>> >
>>>>>>>> >
>>>>>>>> > Brandon Inman
>>>>>>>> > Software Architect
>>>>>>>> >
>>>>>>>> >
>>>>>>>> >
>>>>>>>> > CONFIDENTIALITY NOTICE This message and any included
>>>>>>>> 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,
>>>>>>>> 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
>>>>>>>> corporate offices in Kansas City, Missouri, U.S.A at (+1)
>>>>>>>> (816)221-1024.
>>>>>>>  --
>>>>>>> Director of Data Science
>>>>>>> Cloudera<>
>>>>>>> Twitter: @josh_wills<>
>>>> --
>>>> Director of Data Science
>>>> Cloudera <>
>>>> Twitter: @josh_wills <>
>>> --
>>> Director of Data Science
>>> Cloudera <>
>>> Twitter: @josh_wills <>
> --
> Director of Data Science
> Cloudera <>
> Twitter: @josh_wills <>

View raw message