rx_exec

[アーティクル]
07/15/2019

Usage

revoscalepy.rx_exec(function: typing.Callable, args: typing.Union[dict,
    typing.List[dict]] = None, times_to_run: int = -1,
    task_chunk_size: int = None, compute_context=None,
    local_state: dict = {}, callback: typing.Callable = None,
    continue_on_failure: bool = True, **kwargs)

Description

Allows the distributed execution of an arbitrary R function in parallel, across nodes (computers) or cores of a “compute context”, such as a cluster. For example, you could use the function to execute multiple instances of a model concurrently. When used with base R functions, rxexec does not lift the memory constraints of the underlying functions, or make a single-threaded process a multi-threaded one. The fundamentals of the function still apply; what changes is the execution framework of those functions.

Arguments

function

The function to be executed; the nodes or cores on which it is run are determined by the currently-active compute context and by the other arguments of rx_exec.

args

Arguments passed to the function ‘function’ each time it is executed. for multiple executions with different parameters, args should be a list where each element is a dict. each dict element contains the set of named parameters to pass the function on each execution.

times_to_run

Specifies the number of times ‘function’ should be executed. in case of different parameters for each execution, the length of args must equal times_to_run. in case where length of args equals 1 or args is not specified, function will be called times_to_run number of times with the same arguments.

task_chunk_size

Specifies the number of tasks to be executed per compute element. by submitting tasks in chunks, you can avoid some of the overhead of starting new python processes over and over. for example, if you are running thousands of identical simulations on a cluster, it makes sense to specify the task_chunk_size so that each worker can do its allotment of tasks in a single python process.

compute_context

A RxComputeContext object.

local_state

Dictionary with variables to be passed to a RxInSqlServer compute context not supported in RxSpark compute context.

callback

An optional function to be called with the results from each execution of ‘function’

continue_on_failure

None or logical value. If True, the default, then if an individual instance of a job fails due to a hardware or network failure, an attempt will be made to rerun that job. (Python errors, however, will cause immediate failure as usual.) Furthermore, should a process instance of a job fail due to a user code failure, the rest of the processes will continue, and the failed process will produce a warning when the output is collected. Additionally, the position in the returned list where the failure occurred will contain the error as opposed to a result.

**kwargs

Contains named arguments to be passed to function. Alternative to using args param. For a function foo that takes two arguments named ‘x’ and ‘y’, kwargs can be used like: rx_exec(function=foo, x=5, y=4) (for a single execution with x=5, y=4) rx_exec(function=foo, x=[1,2,3,4,5], y=6) (for multiple executions with different values of x and y=6) rx_exec(function=foo, x=[1,2,3], y=[4,5,6]) (for multiple executions with different values of x and y) Note: if you want to pass in a list as an argument, you must wrap it in another list. For a function foobar that takes an argument named ‘the_list’, kwargs can be used like: rx_exec(function=foobar, the_list=[ [0,1,2,3] ]) (single execution with the_list = [0,1,2,3]) rx_exec(function=foobar, the_list=[ [0,1,2],[3,4,5],[6,7,8] ]) (multiple executions)

Returns

If local compute context or a waiting compute context is active, a list containing the return value of each execution of ‘function’ with specified parameters. If a non-waiting compute context is active, a jobInfo object, or a list of a jobInfo objects for multiple executions.

Example

###
# Local parallel rx_exec with different parameters for each execution
###
import os
from revoscalepy import RxLocalParallel, rx_set_compute_context, rx_exec, rx_btrees, RxOptions, RxXdfData

sample_data_path = RxOptions.get_option("sampleDataDir")
in_mort_ds = RxXdfData(os.path.join(sample_data_path, "mortDefaultSmall.xdf"))
rx_set_compute_context(RxLocalParallel())
formula = "creditScore ~ yearsEmploy"
args = [
    {'data' : in_mort_ds, 'formula' : formula, 'n_tree': 3},
    {'data' : in_mort_ds, 'formula' : formula, 'n_tree': 5},
    {'data' : in_mort_ds, 'formula' : formula, 'n_tree': 7}
]

    models = rx_exec(rx_btrees, args = args)
    # Alternatively
    models = rx_exec(rx_btrees, data=in_mort_ds, formula=formula, n_tree=[3,5,7])

###
## SQL rx_exec
###

from revoscalepy import RxSqlServerData, RxInSqlServer, rx_exec
formula = "ArrDelay ~ CRSDepTime + DayOfWeek"
connection_string="Driver=SQL Server;Server=.;Database=RevoTestDB;Trusted_Connection=TRUE"
query="select top 100 [ArrDelay],[CRSDepTime],[DayOfWeek] FROM airlinedemosmall"

ds = RxSqlServerData(sql_query = query, connection_string = connection_string)
cc = RxInSqlServer(
    connection_string = connection_string,
    num_tasks = 1,
    auto_cleanup = False,
    console_output = True,
    execution_timeout_seconds = 0,
    wait = True
    )

def remote_call(dataset):
    from revoscalepy import rx_data_step
    df = rx_data_step(dataset)
    return len(df)

results = rx_exec(function = remote_call, args={'dataset': ds}, compute_context = cc)
print(results)

###
## Run rx_exec in RxSpark compute context
###
from revoscalepy import *

# start Spark app
spark_cc = rx_spark_connect()

# define function to conditional check
def my_check_fun(id_param):
    if id_param == 13:
        raise Exception("Ensure to fail")
    return id_param

# prepare args for each run with total 20 runs
elem_arg = []
for i in range(0,20):
    elem_arg.append({"id_param": i})

# get result and print
results = rx_exec(my_check_fun, elem_arg, continue_on_failure=True)
print(results)

# stop Spark app
rx_spark_disconnect(spark_cc)

## End(Not run)

次の方法で共有