One more thing that would be good to have is a set of command line tools to parse the JSON dump. I think something like a python wrapper script that given a (job-id, stage-id) can parse the JSON and print out the stage / task timings, with histograms etc. From my own perspective, It'll be super useful for generating data that can be used in benchmarks and experiment results etc.
Thanks Shivaram On Mon, Jan 13, 2014 at 1:32 PM, Patrick Wendell <pwend...@gmail.com> wrote: > Pillis - I agree we need to decouple the representation from a > particular history server. But why not provide as simple history > server people can (optionally) run if they aren't using Yarn or Mesos? > For people running the standalone cluster scheduler this seems > important. Giving them only a JSON dump isn't super consumable for > most users. > > - Patrick > > On Mon, Jan 13, 2014 at 10:43 AM, Pillis Work <pillis.w...@gmail.com> wrote: >> The listeners in SparkUI which update the counters can trigger saves along >> the way. >> The save can be on a 500ms delay after the last update, to batch changes. >> This solution would not require save on stop(). >> >> >> >> On Mon, Jan 13, 2014 at 6:15 AM, Tom Graves <tgraves...@yahoo.com> wrote: >> >>> So the downside to just saving stuff at the end is that if the app crashes >>> or exits badly you don't have anything. Hadoop has taken the approach of >>> saving events along the way. But Hadoop also uses that history file to >>> start where it left off at if something bad happens and it gets restarted. >>> I don't think the latter really applies to spark though. >>> >>> Does mesos have a history server? >>> >>> Tom >>> >>> >>> >>> On Sunday, January 12, 2014 9:22 PM, Pillis Work <pillis.w...@gmail.com> >>> wrote: >>> >>> IMHO from a pure Spark standpoint, I don't know if having a dedicated >>> history service makes sense as of now - considering that cluster managers >>> have their own history servers. Just showing UI of history runs might be >>> too thin a requirement for a full service. Spark should store history >>> information that can later be exposed in required ways. >>> >>> Since each SparkContext is the logical entry and exit point for doing >>> something useful in Spark, during its stop(), it should serialize that >>> run's statistics into a JSON file - like "sc_run_[name]_[start-time].json". >>> When SparkUI.stop() is called, it in turn asks its UI objects (which should >>> implement a trait) to provide either a flat or hierarchical Map of String >>> key/value pairs. This map (flat, hierarchical) is then serialized to a >>> configured path (default being "var/history"). >>> >>> With regards to Mesos or YARN, their applications during shutdown can have >>> API to import this Spark history into their history servers - by making API >>> calls etc. >>> >>> This way Spark's history information is persisted independent of cluster >>> framework, and cluster frameworks can import the history when/as needed. >>> Hope this helps. >>> Regards, >>> pillis >>> >>> >>> >>> On Thu, Jan 9, 2014 at 6:13 AM, Tom Graves <tgraves...@yahoo.com> wrote: >>> >>> > Note that it looks like we are planning on adding support for application >>> > specific frameworks to YARN sooner rather then later. There is an initial >>> > design up here: https://issues.apache.org/jira/browse/YARN-1530. Note >>> > this has not been reviewed yet so changes are likely but gives an idea of >>> > the general direction. If anyone has comments on how that might work >>> with >>> > SPARK I encourage you to post to the jira. >>> > >>> > As Sandy mentioned it would be very nice if the solution could be >>> > compatible with that. >>> > >>> > Tom >>> > >>> > >>> > >>> > On Wednesday, January 8, 2014 12:44 AM, Sandy Ryza < >>> > sandy.r...@cloudera.com> wrote: >>> > >>> > Hey, >>> > >>> > YARN-321 is targeted for the Hadoop 2.4. The minimum feature set doesn't >>> > include application-specific data, so that probably won't be part of 2.4 >>> > unless other things delay the release for a while. There are no APIs for >>> > it yet and pluggable UIs have been discussed but not agreed upon. I >>> think >>> > requirements from Spark could be useful in helping shape what gets done >>> > there. >>> > >>> > -Sandy >>> > >>> > >>> > >>> > On Tue, Jan 7, 2014 at 4:13 PM, Patrick Wendell <pwend...@gmail.com> >>> > wrote: >>> > >>> > > Hey Sandy, >>> > > >>> > > Do you know what the status is for YARN-321 and what version of YARN >>> > > it's targeted for? Also, is there any kind of documentation or API for >>> > > this? Does it control the presentation of the data itself (e.g. it >>> > > actually has its own UI)? >>> > > >>> > > @Tom - having an optional history server sounds like a good idea. >>> > > >>> > > One question is what format to use for storing the data and how the >>> > > persisted format relates to XML/HTML generation in the live UI. One >>> > > idea would be to add JSON as an intermediate format inside of the >>> > > current WebUI, and then any JSON page could be persisted and rendered >>> > > by the history server using the same code. Once a SparkContext exits >>> > > it could dump a series of named paths each with a JSON file. Then the >>> > > history server could load those paths and pass them through the second >>> > > rendering stage (JSON => XML) to create each page. >>> > > >>> > > It would be good if SPARK-969 had a good design doc before anyone >>> > > starts working on it. >>> > > >>> > > - Patrick >>> > > >>> > > On Tue, Jan 7, 2014 at 3:18 PM, Sandy Ryza <sandy.r...@cloudera.com> >>> > > wrote: >>> > > > As a sidenote, it would be nice to make sure that whatever done here >>> > will >>> > > > work with the YARN Application History Server (YARN-321), a generic >>> > > history >>> > > > server that functions similarly to MapReduce's JobHistoryServer. It >>> > will >>> > > > eventually have the ability to store application-specific data. >>> > > > >>> > > > -Sandy >>> > > > >>> > > > >>> > > > On Tue, Jan 7, 2014 at 2:51 PM, Tom Graves <tgraves...@yahoo.com> >>> > wrote: >>> > > > >>> > > >> I don't think you want to save the html/xml files. I would rather >>> see >>> > > the >>> > > >> info saved into a history file in like a json format that could then >>> > be >>> > > >> re-read and the web ui display the info, hopefully without much >>> change >>> > > to >>> > > >> the UI parts. For instance perhaps the history server could read >>> the >>> > > file >>> > > >> and populate the appropriate Spark data structures that the web ui >>> > > already >>> > > >> uses. >>> > > >> >>> > > >> I would suggest making it so the history server is an optional >>> server >>> > > and >>> > > >> could be run on any node. That way if the load on a particular node >>> > > becomes >>> > > >> to much it could be moved, but you also could run it on the same >>> node >>> > as >>> > > >> the Master. All it really needs to know is where to get the history >>> > > files >>> > > >> from and have access to that location. >>> > > >> >>> > > >> Hadoop actually has a history server for MapReduce which works very >>> > > >> similar to what I mention above. One thing to keep in minds here >>> is >>> > > >> security. You want to make sure that the history files can only be >>> > > read by >>> > > >> users who have the appropriate permissions. The history server >>> itself >>> > > >> could run as a superuser who has permission to server up the files >>> > > based >>> > > >> on the acls. >>> > > >> >>> > > >> >>> > > >> >>> > > >> On Tuesday, January 7, 2014 8:06 AM, "Xia, Junluan" < >>> > > junluan....@intel.com> >>> > > >> wrote: >>> > > >> >>> > > >> Hi all >>> > > >> Spark job web ui will not be available when job is over, >>> but >>> > it >>> > > >> is convenient for developer to debug with persisting job web ui. I >>> > just >>> > > >> come up with draft for this issue. >>> > > >> >>> > > >> 1. We could simply save the web page with html/xml >>> > > >> format(stages/executors/storages/environment) to certain location >>> when >>> > > job >>> > > >> finished >>> > > >> >>> > > >> 2. But it is not easy for user to review the job info with #1, >>> > we >>> > > >> could build extra job history service for developers >>> > > >> >>> > > >> 3. But where will we build this history service? In Driver >>> node >>> > or >>> > > >> Master node? >>> > > >> >>> > > >> Any suggestions about this improvement? >>> > > >> >>> > > >> regards, >>> > > >> Andrew >>> > > >> >>> > > >>> > >>>
