Sam Ritchie archive about projects race results

Nov 15 2011 - Washington DC

Cascalog-Contrib

I've had the pleasure of working with Cascalog for about ten months now, and have seen the community produce some fantastic work. A number of businesses are using Cascalog in production; I use Cascalog at Twitter every day to write MapReduce queries for the new Twitter Web Analytics product.

One thing Cascalog doesn't yet have is a community repository for generic queries and operations. To fill this gap we've created cascalog-contrib.

Cascalog-contrib will be home to any higher-level abstractions over Cascalog that the community is willing to submit. If you have an idea for a module, file a pull request on GitHub or bring it up on the mailing list for discussion.

The first cascalog-contrib modules are now live on clojars. To include them in your leiningen or cake project, add any of the following to project.clj:

[cascalog-checkpoint "0.1.1"]
[cascalog-incanter "0.1.0"]
[cascalog-math "0.1.0"]

Contrib currently has three modules: cascalog.math, cascalog.incanter and cascalog.checkpoint. math and incanter are still fairly sparse, but checkpoint is quite interesting and battle-tested at Twitter. Read on if you're interested in the details of the checkpoint module; otherwise, I'll see you on the mailing list!

cascalog.contrib.checkpoint

The workflow macro in the checkpoint module allows you to break complicated workflows out into small, checkpointed steps. If one of these steps causes a job to fail and you restart the job, the workflow macro will skip every step up to the previous point of failure. Fault-tolerant MapReduce topologies ftw!

Let's look at the workflow macro in action. The following function takes an input-path to some existing Twitter data and an output-path, and executes a tweet-processing workflow with five steps:

(defn -main
  [input-path output-path]
  (workflow ["/tmp/example-checkpoint"]
            step-1     ([:tmp-dirs [staging-path]]
                          (transfer-tweets input-path staging-path))

            step-2     ([:deps :last :tmp-dirs user-path]
                          (harvest-users staging-path user-path))

            step-3a    ([:deps step-2 :tmp-dirs [cluster-path friend-path]]
                          (cluster-users user-path cluster-path)
                          (count-friends user-path friend-path))

            step-3b    ([:deps step-2 :tmp-dirs age-path]
                          (examine-ages user-path age-path))

            final-step ([:deps :all]
                          (big-analysis cluster-path
                                        friend-path
                                        age-path
                                        output-path))))

Let's look at this one piece at a time. The first argument to workflow is a vector with some path that the workflow can use to stage temporary files.

(workflow ["/tmp/example-checkpoint"] ...)

It doesn't matter what path you choose; just make sure that Hadoop has access and can write data to the folder.

Following this vector, workflow expects pairs of the form

step-name ([:deps <optional-deps, defaults to :last>]
             :tmp-dirs [<optional, creates temp-dirs if supplied>]
             ...<body, same as inside let>...)

Steps can identify other steps as dependencies by referencing their step-names with the :deps keyword argument.

The first step creates a temporary directory by supplying the symbol staging-path to the :tmp-dirs keyword argument. It then transfers tweets from the input directory into this staging directory, where they will remain available for future steps to consume.

step-1 ([:tmp-dirs [staging-path]]
          (transfer-tweets input-path staging-path))

Step 2 marks :last as a dependency. :last is the default, and marks the step as dependent only on the step directly above. A step will not execute until all of its dependencies have completed successfully.

step-2 uses staging-path defined in step-1 and creates a new temp directory (user-path) for its results.

If step-2 fails for any reason and you restart the workflow, the workflow macro will skip step-1, destroy any temporary directories created in the previous run of step-2, and start step-2 afresh.

step-2 ([:deps :last :tmp-dirs user-path]
          (harvest-users staging-path user-path))

The next two steps, step-3a and step-3b, each mark step-2 as a dependency. Once step-2 completes, step-3a and step-3b will run in parallel.

step-3a ([:deps step-2 :tmp-dirs [cluster-path friend-path]]
           (cluster-users user-path cluster-path)
           (count-friends user-path friend-path))

step-3b ([:deps step-2 :tmp-dirs age-path]
           (examine-ages user-path age-path))

The final step marks its dependencies as :all. This signifies that the step must wait for every step defined above it to complete before running. Again, if final-step fails and the workflow restarts, all previous successful steps will be skipped.

final-step ([:deps :all]
              (big-analysis cluster-path
                            friend-path
                            age-path
                            output-path))

In Conclusion

I'm quite excited about the Cascalog-contrib project and hope you all make heavy use of it as its functionality grows. In the short-term, I'm planning on hooking Cascalog in to Incanter's amazing visualization suite through the cascalog.incanter module.

blog comments powered by Disqus
Fork me on GitHub