Datasplash

Clojure API for a more dynamic Google Cloud Dataflow and (not really battle tested) any other Apache Beam backend.

Usage

You can also see ports of the official Dataflow examples in the datasplash.examples namespace.

Here is the classic word count example.

ℹ️ You will need to run (compile 'datasplash.examples) every time you make a change.

(ns datasplash.examples
  (:require [clojure.string :as str]
            [datasplash.api :as ds]
            [datasplash.options :refer [defoptions]])
  (:gen-class))

(defn tokenize
  [^String l]
  (remove empty? (.split (str/trim l) "[^a-zA-Z']+")))

(defn count-words
  [p]
  (ds/->> :count-words p
          (ds/mapcat tokenize {:name :tokenize})
          (ds/frequencies)))

(defn format-count
  [[k v]]
  (format "%s: %d" k v))

(defoptions WordCountOptions
  {:input {:default "gs://dataflow-samples/shakespeare/kinglear.txt"
           :type String}
   :output {:default "kinglear-freqs.txt" :type String}
   :numShards {:default 0 :type Long}})

(defn -main
  [& str-args]
  (let [p (ds/make-pipeline WordCountOptions str-args)
        {:keys [input output numShards]} (ds/get-pipeline-options p)]
    (->> p
         (ds/read-text-file input {:name "King-Lear"})
         (count-words)
         (ds/map format-count {:name :format-count})
         (ds/write-text-file output {:num-shards numShards})
         (ds/run-pipeline))))

Run it from the repl

Locally on your machine using a DirectRunner:

(in-ns 'datasplash.examples)
(clojure.core/compile 'datasplash.examples)
(-main "--input=sometext.txt" "--output=out-freq.txt" "--numShards=1")

Remotely on Google Cloud using a DataflowRunner:

You should have properly configured your Google Cloud account and Dataflow access from your machine.

(in-ns 'datasplash.examples)
(clojure.core/compile 'datasplash.examples)
(-main "--project=my-project"
       "--runner=DataflowRunner"
       "--gcpTempLocation=gs://bucket/tmp"
       "--input=gs://apache-beam-samples/shakespeare/kinglear.txt"
       "--output=gs://bucket/outputs/kinglear-freq.txt"
       "--numShards=1")

Run it as a standalone program

Datasplash needs to be AOT compiled, so you should prepare an uberjar and run from your main entry like so:

java -jar my-dataflow-job-uber.jar [beam-args]

Caveats

Due to the way the code is loaded when running in distributed mode, you may get some exceptions about unbound vars, especially when using instances with a high number of cpu. They will not however cause the job to fail and are of no consequences. They are caused by the need to prep the Clojure runtime when loading the class files in remote instances and some tricky business with locks and require.
If you have to write your own low-level ParDo objects (you shouldn't), wrap all your code in the safe-exec macro to avoid issues with unbound vars. Any good idea about finding a better way to do this would be greatly appreciated!
If some of the UserCodeException as seen in the cloud UI are mangled and missing the relevant part of the Clojure source code, this is due to a bug with the way the sdk mangles stacktraces in Clojure. In this case look for ClojureRuntimeException in the logs to find the original unaltered stacktrace.
Beware of using Clojure 1.9: proxy results are not Serializable anymore, so you cannot use anywhere in your pipeline Clojure code that uses proxy. Use Java shim for these objects instead.
If you see something like java.lang.ClassNotFoundException: Options you probably forgot to compile your namespace.
Whenever you need to check some spec in user code, you will have to first require those specs because they may not be loaded in your Clojure runtime. But don't use (require) because it's not thread safe. See [this issue] for a workaround.
If you see a java.io.IOException: No such file or directory when invoking compile, make sure there is a directory in your project root that matches the value of *compile-path* (default classes).

About compression libraries

The Beam Java SDK does not pull in the Zstd library by default, so it is the user's responsibility to declare an explicit dependency on zstd-jni. Attempts to read or write .zst files without this library loaded will result in NoClassDefFoundError at runtime.

The Beam Java SDK does not pull in the required libraries for LZOP compression by default, so it is the user's responsibility to declare an explicit dependency on io.airlift:aircompressor and com.facebook.presto.hadoop:hadoop-apache2. Attempts to read or write .lzo files without those libraries loaded will result in a NoClassDefFoundError at runtime.

See Apache Beam Compression enum for details.

License

Distributed under the Eclipse Public License either version 1.0 or (at your option) any later version.