Batch and run your code in parallel. Simply.
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
README.md
runforrest.py
setup.py
test_runforrest.py

README.md

Run, Forrest, Run!

Because sometimes a single Python is just not fast enough.

You have some code that looks like this:

for item in long_list_of_stuff:
    intermediate = prepare(item)
    result = dostuff(intermediate[1], intermediate.thing)

And the problem is, both prepare and dostuff take too long, and long_list_of_stuff is just too damn long. Running the above script takes ages.

What can you do? You could use IPyParallel! Or Dask.distributed! Or multiprocessing!

Well, but IPyParallel has kind of a weird API, and Dask.distributed is hard to set up, and multiprocessing is kind of limited... I've been there, I've done that, and was not satisfied.

Furthermore, none of the above can handle seriously buggy programs that throw segfaults, corrupt your file system, leak memory, and exhaust your process limits. All of this has happened to me when I tried to run scientific code.

So why is RunForrest better?

  1. Understandable. Just short of 200 lines of source code is manageable.
  2. No complicated dependencies. A recent version of Python with dill is all you need.
  3. Simple. The above call graph will now look like this:
  4. Robust. Runforrest survives errors, crashes, and even reboots, without losing data.
from runforrest import TaskList, defer
tasklist = TaskList('tasklist')

for item in long_list_of_stuff:
    task = defer(prepare, item)
    task = defer(dostuff, task[1], task.thing)
    tasklist.schedule(task)

for task in tasklist.run(nprocesses=4):
    result = task.returnvalue # or task.errorvalue

Wrap your function calls in defer, schedule them to a TaskList, then run. That's all there is to it. Deferred functions return objects that can be indexed and getattr'd as much as you'd like and all of that will be resolved once they are run (a single return object will never execute more than one time).

But the best thing is, each schedule will just create a file in a new directory tasklist/todo. Then run will execute those files, and put them in tasklist/done or tasklist/fail depending on whether there were errors or not.

This solves so many problems. Maybe you want to try to re-run a failed item? Just copy them back over to tasklist/todo, and run again. Or python runforrest.py --do_raise infile outfile them manually, and observe the error first hand!

Yes, it's simple. Stupid simple even, you might say. But it is debuggable. It doesn't start a web server. It doesn't set up fancy kernels and messaging systems. It doesn't fork three ways and choke on its own memory consumption. It's just one simple script.

Then again, maybe this won't run so well on more than a couple of computers, and it probably still contains bugs and stuff.

Finally, writing this is not my main job, and I won't be able to answer and fix every pull request at enterprise speed. Please be civil if I can't respond quickly.

More Details:

When creating a TaskList, it is assumed that you create a new, empty list, and will keep that list around after you are done. By default, TaskList therefore throws an error if the chosen directory already exists (exist_ok=False) and will keep all items after you are done (post_clean=False).

# new, empty, unique list:
>>> tasklist = TaskList('new_directory')

If you are experimenting, and will create and re-create a TaskList over and over, set exist_ok=True. This will automatically delete any tasks in the directory when you instantiate your TaskList (pre_clean=True):

>>> # new, empty, pre-existing list:
>>> tasklist = TaskList('existing_directory', exist_ok=True)

If you want to add to a pre-existing TaskList, set pre_clean=False:

>>> # existing list:
>>> tasklist = TaskList('existing_directory', exist_ok=True, pre_clean=False)

If your computer crashed, and you want to continue where you last ran, set noschedule_if_exist=True. This way, all your schedules will be skipped, and the existing TaskList will run all remaining TODO items:

tasklist = TaskList('tasklist', noschedule_if_exist=True)

for item in long_list_of_stuff:
    task = defer(prepare, item)
    task = defer(dostuff, task[1], task.thing)
    tasklist.schedule(task) # will not schedule!

for task in tasklist.run(nprocesses=4): # will run remaining TODOs
    result = task.returnvalue # or task.errorvalue

Tasks

Now you are ready to add tasks. A task is any deferred return value from a function or method call, or plain data:

>>> task = defer(numpy.zeros, 5)
>>> task = defer(dict, alpha='a', beta='b')
>>> task = defer(lambda x: x, 42)
>>> task = defer(42) # behaves like the previous line

Tasks can also be used as arguments to deferred function calls:

>>> task1 = defer(dict, value=42)
>>> task2 = dever(lambda x: x['value'], task1)

You can test-run your tasks using evaluate:

>>> evaluate(task1)
{'value': 42}
>>> evaluate(task2)
42

Tasks can also access attributes or indexes of arguments of tasks. Each task attribute or task index is itself a task, that can be indexed or attributed further:

>>> task1 = defer(dict, value=42)
>>> task2 = defer(lambda x: x, task1['value'])
>>> evaluate(task2)
42
>>> task1 = defer({'value': 42})
>>> task2 = task1['value'] # also returns a Task!
>>> evaluate(task2)
42
>>> task1 = defer(list, [23, 42])
>>> task2 = defer(lambda x: x, task1[1]) # or task2 = task1[1]
>>> evaluate(task2)
42
>>> task1 = defer(Exception, 'an error message')
>>> task2 = defer(lambda x: x, task1.args) # or task2 = task1.args
>>> evaluate(task2)
('an error message',)
>>> task2 = defer(lambda x: x, task1.args[0]) # or task2 = task1.args[0]
'an error message'

This way, you can build arbitrary, deep trees of functions and methods that process your data.

Scheduling and Running

When you finally have a task that evaluates to the result you want, you can schedule it on a TaskList:

>>> tasklist.schedule(task2)

If you want, you can attach some metadata that you can retrieve later. This can be very useful if you are running many many tasks, and want to sort and organize them later:

>>> tasklist.schedule(task2, metadata={'date': date.date()})

Once you have scheduled all the tasks you want, you can run them:

>>> for task in tasklist.run(nprocesses=10):
>>>     do_something_with(task)

This will run each of your tasks in its own process, with ten processes active at any time. In its default, this will yield every finished task, with either the return value in task.returnvalue, or the error in task.errorvalue if there was an error, and the aforementioned metadata in task.metadata.

If you want to get more feedback for failing tasks, you can run them with print_errors=True, which will print the full stack trace of every error the moment it occurs.

Sometimes, dill won't catch some local functions or globals, and your tasks will fail. In that case, set save_session=True and try again.

Accessing Tasks

At any time, you can inspect all currently-scheduled tasks with

>>> for task in tasklist.todo_tasks():
>>>     do_something_with(task)

or, accordingly for done and failed tasks:

>>> for task in tasklist.done_tasks():
>>>     do_something_with(task)
>>> for task in tasklist.fail_tasks():
>>>     do_something_with(task)

These three task lists correspond to *.pkl files in the {tasklist}/todo, {tasklist}/done, and {tasklist}/fail directory, respectively. For example, you can manually re-schedule failed tasks by moving them from the {tasklist}/fail to {tasklist}/todo.

Cleaning

If you want to get rid of tasks, you can use

>>> tasklist.clean()

You can decide manually if you want to remove only selected tasks with clean_todo=True, clean_done=True, and clean_fail=True.

Running Tasks Manually

If one of your tasks is failing, and you can't figure out why, it is useful to manually run just a single task. For this, you can use tasklist as an executable library:

$ python -m runforrest path/to/task.pkl path/to/result.pkl -r

where -r raises any error; Alternatively, -p would print them just like run does with print_errors=True. If needed, you can load a session file with -s path/to/session.pkl.

FAQ:

  • Code that calls Numpy crashes in pthread_create: By default, Numpy creates a large number of threads. If you run many Numpies in parallel, this can exhaust your thread limit. You can either raise your thread limit (resources.setrlimit(resources.RLIMIT_NPROC, ...)) or force Numpy to use only one thread per process (os.putenv('OMP_NUM_THREADS', '1')).

  • Tasks fail because "name '<module>' is not defined": Sometimes, serialization can fail if run is called from a if __name__ == "__main__" block. Set save_session=True in run to fix this problem (for a small performance overhead).