1. What is schedula?

Schedula implements a intelligent function scheduler, which selects and executes functions. The order (workflow) is calculated from the provided inputs and the requested outputs. A function is executed when all its dependencies (i.e., inputs, input domain) are satisfied and when at least one of its outputs has to be calculated.

Note

Schedula is performing the runtime selection of the minimum-workflow to be invoked. A workflow describes the overall process - i.e., the order of function execution - and it is defined by a directed acyclic graph (DAG). The minimum-workflow is the DAG where each output is calculated using the shortest path from the provided inputs. The path is calculated on the basis of a weighed directed graph (data-flow diagram) with a modified Dijkstra algorithm.

2. Installation

To install it use (with root privileges):

$ pip install schedula

Or download the last git version and use (with root privileges):

$ python setup.py install

2.1. Install extras

Some additional functionality is enabled installing the following extras:

• plot: enables the plot of the Dispatcher model and workflow (see plot()).
• web: enables to build a dispatcher Flask app (see web()).
• sphinx: enables the sphinx extension directives (i.e., autosummary and dispatcher).
• parallel: enables the parallel execution of Dispatcher model.

To install schedula and all extras, do:

$ pip install schedula[all]

Note

plot extra requires Graphviz. Make sure that the directory containing the dot executable is on your systems’ path. If you have not you can install it from its download page.

3. Why may I use schedula?

Imagine we have a system of interdependent functions - i.e. the inputs of a function are the output for one or more function(s), and we do not know which input the user will provide and which output will request. With a normal scheduler you would have to code all possible implementations. I’m bored to think and code all possible combinations of inputs and outputs from a model.

3.1. Solution

Schedula allows to write a simple model (Dispatcher) with just the basic functions, then the Dispatcher will select and execute the proper functions for the given inputs and the requested outputs. Moreover, schedula provides a flexible framework for structuring code. It allows to extract sub-models from a bigger one and to run your model asynchronously or in parallel without extra coding.

Note

A successful application is CO₂MPAS, where schedula has been used to model an entire vehicle.

4. Very simple example

Let’s assume that we have to extract some filesystem attributes and we do not know which inputs the user will provide. The code below shows how to create a Dispatcher adding the functions that define your system. Note that with this simple system the maximum number of inputs combinations is 31 (\((2^n - 1)\), where n is the number of data).

>>> import schedula as sh
>>> import os.path as osp
>>> dsp = sh.Dispatcher()
>>> dsp.add_data(data_id='dirname', default_value='.', initial_dist=2)
'dirname'
>>> dsp.add_function(function=osp.split, inputs=['path'],
...                  outputs=['dirname', 'basename'])
'split'
>>> dsp.add_function(function=osp.splitext, inputs=['basename'],
...                  outputs=['fname', 'suffix'])
'splitext'
>>> dsp.add_function(function=osp.join, inputs=['dirname', 'basename'],
...                  outputs=['path'])
'join'
>>> dsp.add_function(function_id='union', function=lambda *a: ''.join(a),
...                  inputs=['fname', 'suffix'], outputs=['basename'])
'union'

Tip

You can explore the diagram by clicking on it.

Note

For more details how to created a Dispatcher see: add_data(), add_func(), add_function(), add_dispatcher(), SubDispatch, SubDispatchFunction, SubDispatchPipe, DispatchPipe, and DFun.

The next step to calculate the outputs would be just to run the dispatch() method. You can invoke it with just the inputs, so it will calculate all reachable outputs:

>>> inputs = {'path': 'schedula/_version.py'}
>>> o = dsp.dispatch(inputs=inputs)
>>> o
Solution([('path', 'schedula/_version.py'),
          ('basename', '_version.py'),
          ('dirname', 'schedula'),
          ('fname', '_version'),
          ('suffix', '.py')])

or you can set also the outputs, so the dispatch will stop when it will find all outputs:

>>> o = dsp.dispatch(inputs=inputs, outputs=['basename'])
>>> o
Solution([('path', 'schedula/_version.py'), ('basename', '_version.py')])

5. Advanced example (circular system)

Systems of interdependent functions can be described by “graphs” and they might contains circles. This kind of system can not be resolved by a normal scheduler.

Suppose to have a system of sequential functions in circle - i.e., the input of a function is the output of the previous function. The maximum number of input and output permutations is \((2^n - 1)^2\), where n is the number of functions. Thus, with a normal scheduler you have to code all possible implementations, so \((2^n - 1)^2\) functions (IMPOSSIBLE!!!).

Schedula will simplify your life. You just create a Dispatcher, that contains all functions that link your data:

>>> import schedula as sh
>>> dsp = sh.Dispatcher()
>>> increment = lambda x: x + 1
>>> for k, (i, j) in enumerate(sh.pairwise([1, 2, 3, 4, 5, 6, 1])):
...     dsp.add_function('f%d' % k, increment, ['v%d' % i], ['v%d' % j])
'...'

Then it will handle all possible combination of inputs and outputs (\((2^n - 1)^2\)) just invoking the dispatch() method, as follows:

>>> out = dsp.dispatch(inputs={'v1': 0, 'v4': 1}, outputs=['v2', 'v6'])
>>> out
Solution([('v1', 0), ('v4', 1), ('v2', 1), ('v5', 2), ('v6', 3)])

5.1. Sub-system extraction

Schedula allows to extract sub-models from a model. This could be done with the shrink_dsp() method, as follows:

>>> sub_dsp = dsp.shrink_dsp(('v1', 'v3', 'v5'), ('v2', 'v4', 'v6'))

Note

For more details how to extract a sub-model see: get_sub_dsp(), get_sub_dsp_from_workflow(), SubDispatch, SubDispatchFunction, DispatchPipe, and SubDispatchPipe.

5.2. Iterated function

Schedula allows to build an iterated function, i.e. the input is recalculated. This could be done easily with the DispatchPipe, as follows:

>>> func = sh.DispatchPipe(dsp, 'func', ('v1', 'v4'), ('v1', 'v4'))
>>> x = [[1, 4]]
>>> for i in range(6):
...     x.append(func(*x[-1]))
>>> x
[[1, 4], [7, 4], [7, 10], [13, 10], [13, 16], [19, 16], [19, 22]]

6. Asynchronous and Parallel dispatching

When there are heavy calculations which takes a significant amount of time, you want to run your model asynchronously or in parallel. Generally, this is difficult to achieve, because it requires an higher level of abstraction and a deeper knowledge of python programming and the Global Interpreter Lock (GIL). Schedula will simplify again your life. It has four default executors to dispatch asynchronously or in parallel:

• async: execute all functions asynchronously in the same process,
• parallel: execute all functions in parallel excluding SubDispatch functions,
• parallel-pool: execute all functions in parallel using a process pool excluding SubDispatch functions,
• parallel-dispatch: execute all functions in parallel including SubDispatch.

Note

Running functions asynchronously or in parallel has a cost. Schedula will spend time creating / deleting new threads / processes.

The code below shows an example of a time consuming code, that with the concurrent execution it requires at least 6 seconds to run. Note that the slow function return the process id.

>>> import schedula as sh
>>> dsp = sh.Dispatcher()
>>> def slow():
...     import os, time
...     time.sleep(1)
...     return os.getpid()
>>> for o in 'abcdef':
...     dsp.add_function(function=slow, outputs=[o])
'...'

while using the async executor, it lasts a bit more then 1 second:

>>> import time
>>> start = time.time()
>>> sol = dsp(executor='async').result()  # Asynchronous execution.
>>> (time.time() - start) < 2  # Faster then concurrent execution.
True

all functions have been executed asynchronously, but in the same process:

>>> import os
>>> pid = os.getpid()  # Current process id.
>>> {sol[k] for k in 'abcdef'} == {pid}  # Single process id.
True

if we use the parallel executor all functions are executed in different processes:

>>> sol = dsp(executor='parallel').result()  # Parallel execution.
>>> pids = {sol[k] for k in 'abcdef'}  # Process ids returned by `slow`.
>>> len(pids) == 6  # Each function returns a different process id.
True
>>> pid not in pids  # The current process id is not in the returned pids.
True
>>> sorted(sh.shutdown_executors())
['async', 'parallel']

7. Next moves

Things yet to do: utility to transform a dispatcher in a command line tool.