Hi Ashutosh, Yes that hook will not actually fire if the client goes away in any way. One way around that (apart from the StorageHandler approach) is to have a PreExecute hook to generate a start of query entry as well. So the ones which do not have an end of query entry are the ones that failed.
About distinguishing between CTAS and create table - one simple method would be that the later would not have any ReadEntity associated with it. For the exception if you throw any Exception in the hook the driver should be able to catch it and report back a failure. I hope this helps. Ashish -----Original Message----- From: Ashutosh Chauhan [mailto:[email protected]] Sent: Thursday, May 27, 2010 11:34 PM To: [email protected] Subject: Re: job level output committer in storage handler Hi Ashish, Thanks for your suggestion. I am getting closer. Few questions I have while I was trying it: I presume this hook is executed by hive client. What will happen if hive client goes away after job is submitted to cluster. I guess this hook wont be executed. If so, this will hurt my use case as I cant afford to not update my external system on successful completion of job. Note that this may not be a problem if a similar hook is exposed via StorageHandler interface (via job level output committer) as then hook will become part of the job itself and will be invoked by MR framework in clean up task at the end. Also write entity is set to be of type table in case of create table query in addition to Insert overwrite and CTAS queries. I want to invoke the hook only in case of INSERT OVERWRITE and CTAS and not in create table statement. Is there way to know this in the hook ? Lastly, if my external system throws an exception or I cant update it, I want to tell hive to consider query as failure. In the hook how can I tell this to hive? Thanks, Ashutosh On Wed, May 26, 2010 at 18:50, Ashutosh Chauhan <[email protected]> wrote: > Oh cool.. I will try that out. > > Thanks, > Ashutosh > > On Wed, May 26, 2010 at 18:46, Ashish Thusoo <[email protected]> wrote: >> Actually if you want to do that then I believe you can check in the post >> execute hook that you have a valid write entity that is of the type table or >> partition. You should have that only in the case of an insert or a CTAS. >> >> Ashish >> >> -----Original Message----- >> From: Ashutosh Chauhan [mailto:[email protected]] >> Sent: Wednesday, May 26, 2010 6:36 PM >> To: [email protected] >> Subject: Re: job level output committer in storage handler >> >> Thanks everyone for the reply. I think its HIVE-1225 is really what I want. >> At this point I can implement PostExecute as I need to call the hook only at >> the end of query and not at the end of each job or task of query. If I >> register it through hive-site.xml then I guess it will get executed for each >> query which is where the complication starts. I want to execute this hook >> only for insert queries and not for all the queries. One workaround is to >> get the Cmd string from session and then parse it to find out if it actually >> is an insert query and only if it is then execute the remainder of code. >> But that looks hacky, I look forward to HIVE-1225. >> >> Thanks, >> Ashutosh >> >> On Wed, May 26, 2010 at 10:35, John Sichi <[email protected]> wrote: >>> I think we'll need to extend the StorageHandler interface so that it can >>> participate in the commit semantics (separate from the handler-independent >>> hooks Ashish mentioned). That was the intention of this followup JIRA >>> issue I logged as part of HBase integration work: >>> >>> https://issues.apache.org/jira/browse/HIVE-1225 >>> >>> To add this one, we need to determine what information needs to be passed >>> along to the storage handler now (and how to make it easy to pass along >>> more information as needed without having to change the interface in the >>> future). >>> >>> JVS >>> >>> ________________________________________ >>> From: Ning Zhang [[email protected]] >>> Sent: Wednesday, May 26, 2010 10:22 AM >>> To: [email protected] >>> Subject: Re: job level output committer in storage handler >>> >>> Hi Ashutosh, >>> >>> Hive doesn't use OutputCommitter explicitly because it handles commit and >>> abort by itself. >>> >>> If you are looking for task level committer where you want to do something >>> after a task successfully finished, you can take a look at the >>> FileSinkOperator.cloaseOp(). It renames tempFile to final file name which >>> implement the commit semantics. >>> >>> If you are looking for job level committer where you want to do something >>> after the job (including all task) finished successfully, you can take a >>> look at the MoveTask implementation. The MoveTask is generated as a follow >>> up task after a MR job for each insert overwrite statement. It moves the >>> directory that contains the results from all finished tasks to its >>> destination path (e.g. a directory specified in the insert statement or >>> inferred from the table's storage location property). The MoveTask >>> implements the commit semantics of the whole job. >>> >>> Ning >>> >>> On May 26, 2010, at 9:16 AM, Ashutosh Chauhan wrote: >>> >>>> Hi Kortni, >>>> >>>> Thanks for your suggestion. But we cant use it in our setup. We are >>>> not spinning hive jobs in a separate process which we can monitor >>>> rather I want to get the handle on when job finishes in my storage >>>> handler / serde. >>>> >>>> Ashutosh >>>> >>>> On Tue, May 25, 2010 at 12:25, Kortni Smith <[email protected]> wrote: >>>>> Hi Ashutosh , >>>>> >>>>> I'm not sure how to accomplish that on the hive side of things, >>>>> but in case it helps I am writing because it sounds like you to >>>>> know when your job is done so you can update something externally >>>>> and my company will also be implementing this in the near future. >>>>> Our plan is to have the process that kicks off our hive jobs in >>>>> the cloud, to monitor each job status periodically using amazon's >>>>> emr java library, and when their state changes to complete, update our >>>>> external systems accordingly. >>>>> >>>>> >>>>> Kortni Smith | Software Developer >>>>> AbeBooks.com Passion for books. >>>>> >>>>> [email protected] >>>>> phone: 250.412.3272 | fax: 250.475.6014 >>>>> >>>>> Suite 500 - 655 Tyee Rd. Victoria, BC. Canada V9A 6X5 >>>>> >>>>> www.abebooks.com | www.abebooks.co.uk | www.abebooks.de >>>>> www.abebooks.fr | www.abebooks.it | www.iberlibro.com >>>>> >>>>> -----Original Message----- >>>>> From: Ashutosh Chauhan [mailto:[email protected]] >>>>> Sent: Tuesday, May 25, 2010 12:13 PM >>>>> To: [email protected] >>>>> Subject: job level output committer in storage handler >>>>> >>>>> Hi, >>>>> >>>>> I am implementing my own serde and storage handler. Is there any >>>>> method in one of these interfaces (or any other) which give me a >>>>> handle to do some operation after all the records have been >>>>> written by all reducer. Something very similar to job level >>>>> output committer. I want to update some state in an external >>>>> system once I know job has completed successfully. Ideally, I >>>>> would do this kind of a thing in a job level output committer, but >>>>> since Hive is on old MR api, I dont have access to that. There is >>>>> a Hive's >>>>> RecordWriter#close() I tried that but it looks like its a task >>>>> level handle. So, every reducer will try to update the state of my >>>>> external system, which is not I want. >>>>> Any pointers on how to achieve this will be much appreciated. If >>>>> its unclear what I am asking for, let me know and I will provide >>>>> more details. >>>>> >>>>> Thanks, >>>>> Ashutosh >>>>> >>> >>> >> >
