I think this is coming together! I like the idea of a client-side handler method that allows us to look at all errors in aggregate and make a decisions based on proportions. How can we guard against catching the wrong mistakes -- say, letting a mapper that's running on a bad node and fails all local disk writes finish "successfully" even though properly, the task just needs to be rerun on a different mapper and normally MR would just take care of it? Let's put this on a wiki for wider feedback.
P.S. What's a "rror" and why do we only want one of them? On Wed, Jan 19, 2011 at 3:07 PM, Julien Le Dem <led...@yahoo-inc.com> wrote: > Some more thoughts. > > * Looking at the existing keywords: > http://pig.apache.org/docs/r0.8.0/piglatin_ref2.html#Reserved+Keywords > It seems ONERROR would be better than ON_ERROR for consistency. There is an > existing ONSCHEMA but no _ based keyword. > > * The default behavior should be to die on error and can be overridden as > follows: > DEFAULT ONERROR <error handler>; > > * Built in error handlers: > Ignore() => ignores errors by dropping records that cause exceptions > Fail() => fails the script on error. (default) > FailOnThreshold(threshold) => fails if number of errors above threshold > > * The error handler interface needs a method called on client side after > the relation is computed to decide what to do next. > Typically FailOnThreshold will throw an exception if > (#errors/#input)>threshold using counters. > public interface ErrorHandler<T> { > > // input is not the input of the UDF, it's the tuple from the relation > T handle(IOExcetion ioe, EvalFunc evalFunc, Tuple input) throws > IOException; > > Schema outputSchema(Schema input); > > // called afterwards on the client side > void collectResult() throws IOException; > > } > > * SPLIT is optional > > example: > DEFAULT ONERROR Ignore(); > ... > > DESCRIBE A; > A: {name: chararray, age: int, gpa: float} > > -- fail it more than 1% errors > B1 = FOREACH A GENERATE Foo(age, gpa), Bar(name) ONERROR > FailOnThreshold(0.01) ; > > -- need to make sure the twitter infrastructure can handle the load > C1 = FOREACH A GENERATE Foo(age, gpa), Bar(name) ONERROR Tweet() ; > > -- custom handler that counts errors and logs on the client side > D1 = FOREACH A GENERATE Foo(age, gpa), Bar(name) ONERROR CountMyErrors() ; > > -- uses default handler and SPLIT > B2 = FOREACH A GENERATE Foo(age, gpa), Bar(name) ONERROR SPLIT INTO > B2_ERRORS; > > -- B2_ERRORS can not really contain the input to the UDF as it would have a > different schema depending on what UDF failed > DESCRIBE B_ERRORS; > B2_ERRORS: {input: (name: chararray, age: int, gpa: float), udf: chararray, > error:(class: chararray, message: chararray, stacktrace: chararray) } > > -- example of filtering on the udf > C2 = FOREACH A GENERATE Foo(age, gpa), Bar(name) ONERROR SPLIT INTO > C2_FOO_ERRORS IF udf='Foo', C2_BAR_ERRORS IF udf='Bar'; > > Julien > > On 1/18/11 3:24 PM, "Dmitriy Ryaboy" <dvrya...@gmail.com> wrote: > > We should think more about the interface. > For example, "Tuple input" argument -- is that the tuple that was passed to > the udf, or the whole tuple that was being processed? I can see wanting > both. > Also the Handler should probably have init and finish methods in case some > accumulation is happening, or state needs to get set up... > > not sure about "splitting" into a table. Maybe more like > > A = FOREACH FOO GENERATE Bar(*) ON_ERROR [use] MyHandler SPLIT [into] > A_ERRORS; > > "use" and "into" are optional syntactic sugar. > > This allows us to do any combination of: > - die > - put original record into a table > - process the error using a custom handler (which can increment counters, > write to dbs, send tweets... definitely send tweets...) > > D > > On Tue, Jan 18, 2011 at 10:27 AM, Julien Le Dem <led...@yahoo-inc.com > >wrote: > > > That would be nice. > > Also letting the error handler output the result to a relation would be > > useful. > > (To let the script output application error metrics) > > For example it could (optionally) use the keyword INTO just like the > SPLIT > > operator. > > > > FOO = LOAD ...; > > A = FOREACH FOO GENERATE Bar(*) ON_ERROR SPLIT MyHandler INTO A_ERRORS; > > > > ErrorHandler would look a little more like EvalFunc: > > > > public interface ErrorHandler<T> { > > > > public T handle(IOExcetion ioe, EvalFunc evalFunc, Tuple input) throws > > IOException; > > > > public Schema outputSchema(Schema input); > > > > } > > > > There could be a built-in handler to output the skipped record (input: > > tuple, funcname:chararray, errorMessage:chararray) > > > > A = FOREACH FOO GENERATE Bar(*) ON_ERROR SPLIT INTO A_ERRORS; > > > > Julien > > > > On 1/16/11 12:22 AM, "Dmitriy Ryaboy" <dvrya...@gmail.com> wrote: > > > > I was thinking about this.. > > > > We add an optional ON_ERROR clause to operators, which allows a user to > > specify error handling. The error handler would be a udf that would > > implement an interface along these lines: > > > > public interface ErrorHandler { > > > > public void handle(IOExcetion ioe, EvalFunc evalFunc, Tuple input) > throws > > IOException; > > > > } > > > > I think this makes sense not to make a static method so that users could > > keep required state, and for example have the handler throw its own > > IOException of it's been invoked too many times. > > > > D > > > > > > On Sat, Jan 15, 2011 at 11:53 PM, Santhosh Srinivasan <s...@yahoo-inc.com > > >wrote: > > > > > Thanks for the clarification Ashutosh. > > > > > > Implementing this in the user realm is tricky as Dmitriy states. > > > Sensitivity to error thresholds will require support from the system. > We > > can > > > probably provide a taxonomy of records (good, bad, incomplete, etc.) to > > let > > > users classify each record. The system can then track counts of each > > record > > > type to facilitate the computation of thresholds. The last part is to > > allow > > > users to specify thresholds and appropriate actions (interrupt, exit, > > > continue, etc.). A possible mechanism to realize this is the > > > ErrorHandlingUDF described by Dmitriy. > > > > > > Santhosh > > > > > > -----Original Message----- > > > From: Ashutosh Chauhan [mailto:hashut...@apache.org] > > > Sent: Friday, January 14, 2011 7:35 PM > > > To: u...@pig.apache.org > > > Subject: Re: Exception Handling in Pig Scripts > > > > > > Santhosh, > > > > > > The way you are proposing, it will kill the pig script. I think what > user > > > wants is to ignore few "bad records" and to process the rest and get > > > results. Problem here is how to let user tell Pig the definition of > "bad > > > record" and how to let him specify threshold for % of bad records at > > which > > > Pig should fail the script. > > > > > > Ashutosh > > > > > > On Fri, Jan 14, 2011 at 18:18, Santhosh Srinivasan <s...@yahoo-inc.com> > > > wrote: > > > > Sorry about the late response. > > > > > > > > Hadoop n00b is proposing a language extension for error handling, > > similar > > > to the mechanisms in other well known languages like C++, Java, etc. > > > > > > > > For now, can't the error semantics be handled by the UDF? For > > exceptional > > > scenarios you could throw an ExecException with the right details. The > > > physical operator that handles the execution of UDF's traps it for you > > and > > > propagates the error back to the client. You can take a look at any of > > the > > > builtin UDFs to see how Pig handles it internally. > > > > > > > > Santhosh > > > > > > > > -----Original Message----- > > > > From: Dmitriy Ryaboy [mailto:dvrya...@gmail.com] > > > > Sent: Tuesday, January 11, 2011 10:41 AM > > > > To: u...@pig.apache.org > > > > Subject: Re: Exception Handling in Pig Scripts > > > > > > > > Right now error handling is controlled by the UDFs themselves, and > > there > > > is no way to direct it externally. > > > > You can make an ErrorHandlingUDF that would take a udf spec, invoke > it, > > > trap errors, and then do the specified error handling behavior.. that's > a > > > bit ugly though. > > > > > > > > There is a problem with trapping general exceptions of course, in > that > > if > > > they happen 0.000001% of the time you can probably just ignore them, > but > > if > > > they happen in half your dataset, you want the job to tell you > something > > is > > > wrong. So this stuff gets non-trivial. If anyone wants to propose a > > design > > > to solve this general problem, I think that would be a welcome > addition. > > > > > > > > D > > > > > > > > On Tue, Jan 11, 2011 at 12:47 AM, hadoop n00b <new2h...@gmail.com> > > > wrote: > > > > > > > >> Thanks, I sometimes get a date like 0001-01-01. This would be a > valid > > > >> date format, but when I try to get the seconds between this and > > > >> another date, say 2011-01-01, I get an error that the value is too > > > >> large to be fit into int and the process stops. Do we have something > > > >> like ifError(x-y, null,x-y)? Or would I have to implement this as an > > > >> UDF? > > > >> > > > >> Thanks > > > >> > > > >> On Tue, Jan 11, 2011 at 11:40 AM, Dmitriy Ryaboy < > dvrya...@gmail.com> > > > >> wrote: > > > >> > > > >> > Create a UDF that verifies the format, and go through a filtering > > > >> > step first. > > > >> > If you would like to save the malformated records so you can look > > > >> > at them later, you can use the SPLIT operator to route the good > > > >> > records to your regular workflow, and the bad records some place > on > > > HDFS. > > > >> > > > > >> > -D > > > >> > > > > >> > On Mon, Jan 10, 2011 at 9:58 PM, hadoop n00b <new2h...@gmail.com> > > > wrote: > > > >> > > > > >> > > Hello, > > > >> > > > > > >> > > I have a pig script that uses piggy bank to calculate date > > > differences. > > > >> > > Sometimes, when I get a wierd date or wrong format in the input, > > > >> > > the > > > >> > script > > > >> > > throws and error and aborts. > > > >> > > > > > >> > > Is there a way I could trap these errors and move on without > > > >> > > stopping > > > >> the > > > >> > > execution? > > > >> > > > > > >> > > Thanks > > > >> > > > > > >> > > PS: I'm using CDH2 with Pig 0.5 > > > >> > > > > > >> > > > > >> > > > > > > > > > > > > >