concurrently

A clojure library for making concurrent process-engine backed by core.async. Application can share the engine and can put requests onto the engine. Engine will return a channel containing all processed results for each request.

All verbose processes for splitting or collecting results for each request and putting the results onto a correct channel are handled by this library. All things that programmer must do is 1) creating a shared pipeline, 2) put all data you want to process onto the input-channel of the shared pipeline, 3) handle the results by calling get-results on the output-channel returned by the shared pipeline.

This library safely handles pipelines backed by core.async so that pipelines never stack by abuse of channels like stopping data-retrieving from channels before fully retrieving all data or forgetting to close channels.

This library is backed by databox (https://github.com/tyano/databox) that wraps data into a box and make data-handling in channel-pipeline exception-safe. All exceptions occured in databox safely are wrapped and be passed-through channels until the end of channel-pipeline, so that users can handle exceptions at the end of pipeline.

get-results function handles all actions needed for protecting pipeline from accidental stacking. get-results try to retrieve all data from an output-channel and unwrap databoxes. If an exception were occurred in the pipeline, unwrapping the failure-databox will throw the exception, but get-results handles all data remaining in then pipeline for protecting it from accidental stacking.

Usage

Install

Leiningen:

[concurrently "0.2.3"]

Clojure CLI:

concurrently/concurrently {:mvn/version "0.2.3"}

Create an process-engine by concurrent-process function

You can make an engine by calling concurrent-process or concurrent-process-blocking functions with a transducer, input channel, output channels and max parallel count. All data supplied into then input-channel will be handled by a supplied transducer in parallel by go-blocks or threads of same number with the parallel count.

But you should not put data into the input-channel directly. You must use the concurrently function (Explanation of the concurrently fn is following).

Transducer must accept a map with :data and :options keys. :data is a value from input channel. :options is a option-map supplied to concurrently function.

example:

(def shared-process (concurrent-process
                        8
                        (chan 1)
                        ;; transducer must accept a map with :data and :options keys
                        (map #(fn [{:keys [data options]}] (my-great-function data options))
                        (chan 1)))

unordered pipeline

All of pipeline functions of core.async only support ordered pipeline, on which all output data are produced in same order of input. But in real usecase you would want to get results as soon as possible just after the input data is processed.

This library supports 'unordered-pipeline'.

With onordered pipeline, output data are not in same order with input values. all results will be pushed to output channel as soon as possible after a parallel task finished.

For making unordered pipeline, just supply {:ordered false} as option map to make-concurrent-process function.

example:

(def shared-process (concurrent-process
                        8
                        (chan 1)
                        ;; transducer must accept a map with :data and :options keys
                        (map #(fn [{:keys [data options]}] (my-great-function data options))
                        (chan 1)
                        {:ordered? false})) ;; <<- This line

Supply data to the created engine

concurrent-process and concurrent-process-blocking functions return a process-context which contains an input channel. But you should not use the input channel directly. Instead, you should use the concurrently function with arguments of the returned process-context, a channel which all input data can be read from, and an option map.

concurrently wraps all input-data by databox and concurrent-process handles the databoxes so that if some exceptions occurred in a concurrent-process, the exception will safely be converted to failure-boxes and be passed-through the pipeline.

getting calculated results from a Job

concurrently returns a Job. You can cancel the job by calling (cancel job) function. Job contains a field :channel. You can read all calculated results for all input-data supplied to concurrently function from the channel, but should not read it directly. Use (get-results (:channel job)) function for safe-reading from channels.

get-results handles all databoxes from a channel and create a vector which contains all values of databoxes. If a failure databox is found while handling databoxes, get-results will throw the exception and handle all remaining data in a channel in background for protecting the channel from accidental stacking caused by never-read data in a channel.

get-results accepts arguments:

ch - a channel for reading. option-map - optional. a map containing the following keys.

keys: :catch - is a funciton called when an exception occurs. This fn is for closing related channels certainly. In many cases, if an exception occurred, no following channel-processsings are acceptable, So all read-channels must be closed at this time. It is recommended to supply this function always, but should not be used for handling application exceptions. Only for channel handling. Application exceptions should be handled in application code by a try-catch block which wraps this 'get-result' call.

:finally - is a function that will be called always, but be called after the ch is CLOSED. If the ch is not read fully, it will be read fully automatically. When the ch is read fully or be closed manually, this :finally fn will be called. So SHOULD NOT DO APPLICATION's FINALLY-PROCESS here. This function is for actions which must be occurred after the ch is closed. Application's finally-process must be handled in application code by a try-catch block which wraps this 'get-result' call.

:context-name - optional. a information which is used at logging for distincting processes.

:timeout-ms - optional. default is 120000 (ms). The time to give up reading the ch. An exception will be thrown after this timeout and the exception will be wrapped by a failure box and be returned.

Connecting channels

You should call get-results at last for safe-processing of channels, but before calling it, you can connect the job to other transducers. Although you can do it with pipe of core.async, but there is a safe utility function for doing it.

(chain job-or-channel transducer exception-handler)

You can apply a transducer to a job-or-channel by chain function (with an exception handler if you want) and can get a next job-or-channel. You can chain the calling of chain fns like:

(-> job
    (chain xf1)
    (chain xf2)
    (chain xf3))

This will return a last job-or-channel and you can get the results by calling get-results on the last job-or-channel.

Difference of chain and pipe is that pipe stops reading from an input channel if the output channel is closed, but chain never stop reading input so that data never remain in an input channel. Remaining data in a channel might cause accidental channel-stacking. All data should be read fully.

You can use chain on all instances that satisfies the Chainable protocol. Currently channels of core.async and ConcurrentJob returned from the concurrently function satisfy the protocol. So you can:

;; You can apply a transducer to a ConcurrentJob.
;; The transducer will be applied to the :channel of the ConcurrentJob.
(let [job (-> (concurrently ...)
              (chain xf)]
  ..do something..)

chain always returns a same type of object that is passed to the chain function. So if you call chain on a channel, a channel will be returned, if you call chain on a ConcurrentJob, a ConcurrentJob will be returned.

Note that the data from a channel in a job always are databoxes. Transducer supplied to chain must handle databox and must return databox. Use databox.core/map, databox.core/mapcat or databox.core/filter transducers for handling databoxes for safe. Because functions of databox safely handle exceptions that occurred while processing data and always return a databox, and most of the functions construct transducers, you should use them for handling databoxes.

For example:

(let [job (-> (concurrently ...)
              (chain (databox.core/map #(your-great-mapper %))
              (chain (databox.core/filter #(your-great-filter %)))]
  ..do something..)

WORK THROUGH

(defn my-great-function
  [data options]
  ;; do something and return a string
  )

;;; create an engine handle data by 8 threads in parallel.
(def shared-process (concurrent-process-blocking 8
                                                 (chan 1)
                                                 ;; transducer must accept a map with :data and :options keys
                                                 (map #(fn [{:keys [data options]}] (my-great-function data options))
                                                 (chan 1)))

;;; pass data
(let [{:keys [channel] :as job} (-> (concurrenty shared-process (to-chan! [:a :b :c]) {:option-to-function true})
                                    (chain (databox.core/map #(upper-case %))))
      results (get-results channel
                           {:catch (fn [ex] (cancel job))
                            :finally (fn [] ...)
                            :timeout-ms 5000})]
  ;; you can handle results here
  )

License

Copyright © 2019 Tsutomu YANO

This program and the accompanying materials are made available under the terms of the Eclipse Public License 2.0 which is available at http://www.eclipse.org/legal/epl-2.0.

This Source Code may also be made available under the following Secondary Licenses when the conditions for such availability set forth in the Eclipse Public License, v. 2.0 are satisfied: GNU General Public License as published by the Free Software Foundation, either version 2 of the License, or (at your option) any later version, with the GNU Classpath Exception which is available at https://www.gnu.org/software/classpath/license.html.