Configuration

Cubed uses Donfig for managing configuration of things like the directory used for temporary files, or for setting executor properties.

This page covers how to specify configuration properties, and a reference with all the configuration options that you can use in Cubed.

Specification

There are three main ways of specifying configuration in Cubed:

1. by instantiating a Spec object,

2. by using a YAML file and setting the CUBED_CONFIG environment variable, or

3. by setting environment variables for individual properties.

We look at each in turn.

Spec object

This is how you configure Cubed directly from within a Python program - by instantiating a Spec object:

import cubed

spec = cubed.Spec(
    work_dir="s3://cubed-tomwhite-temp",
    allowed_mem="2GB",
    executor_name="lithops",
    executor_options=dict(use_backups=False, runtime="cubed-runtime", runtime_memory=2000)
)

The spec instance is then passed to array creation functions as follows:

import cubed.array_api as xp

a = cubed.random.random((50000, 50000), chunks=(5000, 5000), spec=spec)
b = cubed.random.random((50000, 50000), chunks=(5000, 5000), spec=spec)
c = xp.add(a, b)

All arrays in any given computation must share the same spec instance.

YAML file

A YAML file is a good way to encapsulate the configuration in a single file that lives outside the Python program. It’s a useful way to package up the settings for running using a particular executor, so it can be reused. The Cubed examples use YAML files for this reason.

spec:
  work_dir: "s3://cubed-$USER-temp"
  allowed_mem: "2GB"
  executor_name: "lithops"
  executor_options:
    use_backups: False
    runtime: "cubed-runtime"
    runtime_memory: 2000

Note that YAML files can use environment variables - in this example $USER will be expanded appropriately.

To use the YAML file, set the CUBED_CONFIG environment variable to the file path before invoking your Python program:

CUBED_CONFIG=/path/to/cubed.yaml python ...

Donfig will actually look for YAML files in a variety of locations, see the Donfig docs for details.

Environment variables

You can also set Donfig-style environment variables to set individual properties. Notice how double underscores are used to indicate nesting.

export CUBED_SPEC__WORK_DIR='s3://cubed-$USER-temp'
export CUBED_SPEC__ALLOWED_MEM=2GB
export CUBED_SPEC__EXECUTOR_NAME=lithops
export CUBED_SPEC__EXECUTOR_OPTIONS__USE_BACKUPS=False
export CUBED_SPEC__EXECUTOR_OPTIONS__RUNTIME=cubed-runtime
export CUBED_SPEC__EXECUTOR_OPTIONS__RUNTIME_MEMORY=2000

This can be handy if you only have a couple of properties to set:

CUBED_SPEC__ALLOWED_MEM=2GB CUBED_SPEC__EXECUTOR_NAME=processes python ...

Reference

Spec options

These properties can be passed directly to the Spec constructor. Or, equivalently, they are directly under spec in a YAML file.

Property	Default	Description
work_dir	None	The directory path (specified as an fsspec URL) used for storing intermediate data. If not set, the user’s temporary directory is used.
allowed_mem	2GB	The total memory available to a worker for running a task. This includes any reserved_mem that has been set.
reserved_mem	100MB	The memory reserved on a worker for non-data use when running a task
executor_name	single-threaded	The executor for running computations. One of single-threaded, threads, processes, beam, coiled, dask, lithops, modal.
executor_options	None	Options to pass to the executor on construction. See below for possible options for each executor.

Executor options

Different executors support different options. They are listed here for each executor.

These properties are keys in the executor_options passed to the Spec constructor. Or, equivalently, they are directly under spec.executor_options in a YAML file.

single-threaded

The single-threaded executor is a simple executor mainly used for testing. It doesn’t support any configuration options since it is deliberately designed not to have anything except the most basic features.

threads

Property	Default	Description
retries	2	The number of times to retry a task if it fails.
use_backups		Whether to use backup tasks for mitigating stragglers. Defaults to True only if work_dir is a filesystem supporting atomic writes (currently a cloud store like S3 or GCS).
batch_size	None	Number of input tasks to submit to be run in parallel. The default is not to batch.
compute_arrays_in_parallel	False	Whether arrays are computed one at a time or in parallel.
max_workers	None	The maximum number of workers to use in the ThreadPoolExecutor. Defaults to number of CPU cores.

processes

Property	Default	Description
use_backups		Whether to use backup tasks for mitigating stragglers. Defaults to True only if work_dir is a filesystem supporting atomic writes (currently a cloud store like S3 or GCS).
batch_size	None	Number of input tasks to submit to be run in parallel. None means don’t batch.
compute_arrays_in_parallel	False	Whether arrays are computed one at a time or in parallel.
max_workers	None	The maximum number of workers to use in the ProcessPoolExecutor. Defaults to number of CPU cores.
max_tasks_per_child	None	The number of tasks to run in each child process. See the Python documentation for concurrent.futures.ProcessPoolExecutor. (Python 3.11)

Note that retries is not currently supported for the processes executor.

beam

The beam executor doesn’t currently expose any configuration options. When running on Google Cloud Dataflow, four retry attempts are made for failing tasks.

coiled

Property	Default	Description
coiled_kwargs	None	Keyword arguments to pass to coiled.function.

Note that there is currently no way to set retries or a timeout for the Coiled executor.

dask

Property	Default	Description
retries	2	The number of times to retry a task if it fails.
use_backups		Whether to use backup tasks for mitigating stragglers. Defaults to True only if work_dir is a filesystem supporting atomic writes (currently a cloud store like S3 or GCS).
batch_size	None	Number of input tasks to submit to be run in parallel. The default is not to batch.
compute_arrays_in_parallel	False	Whether arrays are computed one at a time or in parallel.
compute_kwargs	None	Keyword arguments to pass to Dask’s distributed.Client constructor.

Note that there is currently no way to set a timeout for the Dask executor.

lithops

Property	Default	Description
retries	2	The number of times to retry a task if it fails.
timeout	None	Tasks that take longer than the timeout will be automatically killed and retried. Defaults to the timeout specified when deploying the lithops runtime image. This is 180 seconds in the examples.
use_backups		Whether to use backup tasks for mitigating stragglers. Defaults to True only if work_dir is a filesystem supporting atomic writes (currently a cloud store like S3 or GCS).
compute_arrays_in_parallel	False	Whether arrays are computed one at a time or in parallel.
Other properties	N/A	Other properties will be passed as keyword arguments to the lithops.executors.FunctionExecutor constructor.

Note that batch_size is not currently supported for Lithops.

modal

Property	Default	Description
cloud	aws	The cloud to run on. One of aws or gcp.
use_backups	True	Whether to use backup tasks for mitigating stragglers.
batch_size	None	Number of input tasks to submit to be run in parallel. The default is not to batch.
compute_arrays_in_parallel	False	Whether arrays are computed one at a time or in parallel.

Currently the Modal executor in Cubed uses a hard-coded value of 2 for retries and 300 seconds for timeouts, neither of which can be changed through configuration.