DFLOW
Dflow is a Python framework for constructing scientific computing workflows (e.g. concurrent learning workflows) employing Argo Workflows as the workflow engine. (arXiv: https://arxiv.org/abs/2404.18392)
For dflow's users (e.g. ML application developers), dflow offers user-friendly functional programming interfaces for building their own workflows. Users need not be concerned with process control, task scheduling, observability and disaster tolerance. Users can track workflow status and handle exceptions by APIs as well as from frontend UI. Thereby users are enabled to concentrate on implementing operations (OPs) and orchestrating workflows.
For dflow's developers, dflow wraps on argo SDK, keeps details of computing and storage resources from users, and provides extension abilities. While argo is a cloud-native workflow engine, dflow uses containers to decouple computing logic or scheduling logic, and leverages Kubernetes to make workflows observable, reproducible and robust. Dflow is designed to be based on a distributed, heterogeneous infrastructure. With high-performance computing (HPC) clusters being a common resources in scientific computing, users can either use executor to manage HPC jobs using DPDispatcher plugin, or use virtual node technique to uniformly manage HPC resources in the framework of Kubernetes (wlm-operator).
OP template (abbr. OP) in dflow can be reused across workflows and shared among users. Dflow provides a cookie cutter recipe dflow-op-cutter for template a new OP package. Start developing an OP package at once from
pip install cookiecutter
cookiecutter https://github.com/deepmodeling/dflow-op-cutter.git
Dflow provides a debug mode for running workflows bare-metally whose backend is implemented in dflow itself in pure Python, independent of Argo/Kubernetes. The debug mode utilizes local environment to execute OPs instead of containers by default. It implements most APIs of the default (Argo) mode in order to ensure a consistent user experience. The debug mode offer convenience for debugging or testing without container. In cases where clusters face difficulties deploying Docker or Kubernetes or have limited external access, the debug mode may also be used for production, despite less robustness and observability.
1. Overview
1.1. Basics
1.1.1. Parameters and artifacts
Parameters and artifacts are data stored and transferred within a workflow. Parameters are saved as text which can be displayed in the UI, while artifacts are stored as files. Parameters are passed to an OP with their values, while artifacts are passed as paths.
1.1.2. OP template
OP template (abbr. OP) serves as a fundamental building block of a workflow. It defines a particular operation to be executed given the input and the expected output. Both the input and output can be parameters and/or artifacts. The most common OP template is the container OP template. Two types of container OP templates are supported: ShellOPTemplate
, PythonScriptOPTemplate
. ShellOPTemplate
defines an operation by shell script and a container image where the script runs. PythonScriptOPTemplate
defines an operation by Python script and a container image.
As a more Python-native category of OP templates provided by dflow, PythonOPTemplate
defines OPs in the form of Python classes or Python functions (termed class OP or function OP respectively). As Python is a weak typed
language, we impose strict type checking to Python OPs to minimize ambiguity and unexpected behaviors.
For a class OP, the input and output structures of an OP are declared in the static methods get_input_sign
and get_output_sign
. Each of them returns a dictionary mapping from the name of a parameter/artifact to its type. The execution of the OP is defined in the execute
method. The types of the parameter values passed in and out should be in accord with those declared in the sign. Type checking is implemented before and after the execute
method. For an input/output artifact, its sign should be like Artifact(type)
where type
can be Path
, List[Path]
, Dict[str, Path]
or dflow.python.NestedDict[Path]
. For input artifact, the OP receives a path, a list of paths or a dictionary of paths according to its sign . OP developer can directly process the file or directory at the path. For output artifact, the OP should also return a path, a list of paths or a dictionary of paths according to its sign.
from dflow.python import OP, OPIO, OPIOSign, Artifact
from pathlib import Path
import shutil
class SimpleExample(OP):
def __init__(self):
pass
@classmethod
def get_input_sign(cls):
return OPIOSign(
{
"msg": str,
"foo": Artifact(Path),
}
)
@classmethod
def get_output_sign(cls):
return OPIOSign(
{
"msg": str,
"bar": Artifact(Path),
}
)
@OP.exec_sign_check
def execute(
self,
op_in: OPIO,
) -> OPIO:
shutil.copy(op_in["foo"], "bar.txt")
out_msg = op_in["msg"]
op_out = OPIO(
{
"msg": out_msg,
"bar": Path("bar.txt"),
}
)
return op_out
The above example illustrates an OP SimpleExample
. This operation involves copying the input artifact foo
to the output artifact bar
and duplicating the input parameter msg
to the output parameter msg
.
For an function OP, the input and output structures are declared more succinctly using type annotations and execution process is defined in the function body. Type checking is implemented before and after the function as well. We recommend python>=3.9
to use this syntactic sugar. To learn more about Python Annotation, refer to the Python official howtos.
from dflow.python import OP, Artifact
from pathlib import Path
import shutil
@OP.function
def SimpleExample(
msg: str,
foo: Artifact(Path),
) -> {"msg": str, "bar": Artifact(Path)}:
shutil.copy(foo, "bar.txt")
out_msg = msg
return {"msg": out_msg, "bar": Path("bar.txt")}
To create an OP template based on the above class or function, we need to specify the container image and any optional arguments to PythonOPTemplate
. pydflow
need not to be installed in this image because local pydflow
package will automatically be uploaded into the container by default
from dflow.python import PythonOPTemplate
simple_example_templ = PythonOPTemplate(SimpleExample, image="python:3.8")
An example is here
1.1.3. Step
Step
serves as the central component for establishing data flow rules. A step is created by instantiating an OP template, which requires specifying the values for all input parameters and sources for all input artifacts declared in the OP template. The input parameters/artifacts of a step may be either static at the time of submission, or dynamically from outputs of another step.
from dflow import Step
simple_example_step = Step(
name="step0",
template=simple_example_templ,
parameters={"msg": "HelloWorld!"},
artifacts={"inp_art": foo},
)
Note that foo
here is an artifact either uploaded locally or output of another step.
1.1.4. Workflow
Workflow
connects steps together to build a workflow. A simple serial workflow is created by adding steps in sequence. Adding a list of steps to a workflow means these steps will run concurrently.
from dflow import Workflow
wf = Workflow(name="hello-world")
wf.add(simple_example_step)
Submit a workflow by
wf.submit()
An example is here
2. Quick Start
2.1. Setup Argo Server
If you have an Argo server already, you can skip this step. Otherwise you can follow the installation guide.
2.2. Install dflow
Make sure your Python version is not less than 3.6 and install dflow
pip install pydflow
2.3. Run an example
There are several notebook tutorials that can help you start to use dflow. Besides, you can submit a simple workflow from the terminal
python examples/test_python.py
Then you can check the submitted workflow through argo's UI.
3.1. Common
3.1.1. Workflow management
After submitting a workflow by wf.submit()
, or retrieving a history workflow by wf = Workflow(id="xxx")
, you can track its real-time status with APIs
wf.id
: workflow ID in argowf.query_status()
: query workflow status, return "Pending"
, "Running"
, "Succeeded"
, etc.wf.query_step(name=None, key=None, phase=None, id=None, type=None)
: query step by name (support for regex)/key/phase/ID/type, return a list of argo step objects
step.phase
: phase of a step, "Pending"
, "Running"
, Succeeded
, etc.step.outputs.parameters
: a dictionary of output parameters mapping parameter names to parameter objects, get the value of a parameter by step.outputs.artifacts["foo"].value
step.outputs.artifacts
: a dictionary of output artifacts mapping artifact names to artifact objects, download an artifact by download_artifact(step.outputs.artifacts["bar"])
wf.terminate()
: terminate a workflow
3.1.2. Upload artifact
Dflow offers tools for uploading files to the artifact repository and downloading files from it (default artifact repository is Minio set up in the quick start). User can upload a file/directory, a list of files/directories or a dictionary of files/directories and obtain an artifact object, which can be used as argument of a step
artifact = upload_artifact([path1, path2])
step = Step(
...
artifacts={"foo": artifact}
)
Modify dflow.s3_config
to configure artifact repository settings globally.
Note: during the upload process, dflow retains the relative path of the uploaded file/directory with respect to the current directory. If file/directory outside current directory is uploaded, its absolute path is used as the relative path in the artifact. If you want a different directory structure in the artifact compared to the local one, you can create soft links and then upload.
3.1.3. Steps
Steps
is another type of OP template which is defined by its constituent steps instead of a container. It can be viewed as a sub-workflow or a super OP template composed of smaller OPs. A steps
includes an array of arrays of step
s, e.g. [[s00,s01],[s10,s11,s12]]
, where the inner array represents concurrent steps while the outer array indicates sequential steps. Input/output parameters/artifacts for a steps can be declared as follows:
steps.inputs.parameters["msg"] = InputParameter()
steps.inputs.artifacts["foo"] = InputArtifact()
steps.outputs.parameters["msg"] = OutputParameter()
steps.outputs.parameters["bar"] = OutputArtifact()
Adding a step
to a steps
is similar to adding one to a workflow
steps.add(step)
steps
can be used as the template to instantiate a bigger step
just like script OP templates. This allows for constructing complex workflows of nested structure. It is also possible to recursively use a steps
as the template of a building block within itself to achieve dynamic loop.
To set an output parameter of steps
from a step within it
steps.outputs.parameters["msg"].value_from_parameter = step.outputs.parameters["msg"]
Here, step
must be contained within steps
. Similarly, to assign an output artifact for steps
, use
steps.outputs.artifacts["foo"]._from = step.outputs.parameters["foo"]
3.1.4. DAG
DAG
is another type of OP template which is defined by its constituent tasks and their dependencies. The usage of DAG
is similar to that of Steps
. To incorporate a task
into a dag
, use
dag.add(task)
The usage of task
is also similar to that of step
. Dflow will automatically detect dependencies among tasks of a dag
based on their input/output relations. Additional dependencies can be specified by
task_3 = Task(..., dependencies=[task_1, task_2])
3.1.5. Conditional step, parameters and artifacts
Set a step to be conditional by Step(..., when=expr)
where expr
is an boolean expression in string format. Such as "%s < %s" % (par1, par2)
. The step will be executed if the expression is evalutated to be true, otherwise skipped. The when
argument is often used as the breaking condition of recursive steps. The output parameter of a steps
(similar to dag
) can be assigned as conditional by
steps.outputs.parameters["msg"].value_from_expression = if_expression(
_if=par1 < par2,
_then=par3,
_else=par4
)
Similarly, the output artifact of a steps
can be assigned as conditional by
steps.outputs.artifacts["foo"].from_expression = if_expression(
_if=par1 < par2,
_then=art1,
_else=art2
)
3.1.6. Produce parallel steps using loop
In scientific computing, it is often required to produce a list of parallel steps which share a common OP template, and only differ in the input parameters. with_param
and with_sequence
are 2 arguments of Step
for automatically generating a list of parallel steps. These steps share a common OP template, and only differ in the input parameters.
A step using with_param
option generates parallel steps based on a list (either a constant list or referring to another parameter, e.g. an output parameter of another step or an input parameter of the steps
or DAG
context), the parallelism equals to the length of the list. Each parallel step picks an item from the list by "{{item}}"
, such as
step = Step(
...
parameters={"msg": "{{item}}"},
with_param=steps.inputs.parameters["msg_list"]
)
A step using with_sequence
option generates parallel steps based on a numeric sequence. with_sequence
is usually used in coordination with argo_sequence
which returns an Argo's sequence. For argo_sequence
, the number at which to start the sequence is specified by start
(default: 0). You can either specify the number of elements in the sequence by count
or the ending number of the sequence by end
. The printf format string can be specified by format
to format the values in the sequence. Each argument can be passed with a parameter, argo_len
which returns the length of a list may be useful. Each parallel step picks an element from the sequence by "{{item}}"
, such as
step = Step(
...
parameters={"i": "{{item}}"},
with_sequence=argo_sequence(argo_len(steps.inputs.parameters["msg_list"]))
)
3.1.7. Timeout
Set the timeout of a step by Step(..., timeout=t)
. The unit is second.
3.1.8. Continue on failed
To allow a workflow to continue even if a step fails by Step(..., continue_on_failed=True)
.
3.1.9. Continue on success number/ratio of parallel steps
For a group of parallel steps generated by with_param
or with_sequence
, enable the workflow to continue when a specific number/ratio of parallel steps succeed by Step(..., continue_on_num_success=n)
or Step(..., continue_on_success_ratio=r)
.
3.1.10. Optional input artifacts
Set an input artifact to be optional by op_template.inputs.artifacts["foo"].optional = True
.
3.1.11. Default value for output parameters
Set default value for an output parameter by op_template.outputs.parameters["msg"].default = default_value
. The default value will be used when the expression in value_from_expression
fails or the step is skipped.
3.1.12. Key of a step
You can assign a key to a step by Step(..., key="some-key")
for the convenience of locating the step. The key can be regarded as an input parameter which may contain references to other parameters. For instance, the key of a step may change with iterations of a dynamic loop. Once a key is assigned to a step, the step can be query by wf.query_step(key="some-key")
. If the key is unique within the workflow, the query_step
method returns a singleton list (a list consist of only one element).
3.1.13. Resubmit a workflow
Workflows often include some computationally expensive steps. Outputs of previously run steps can be reused for submitting a new workflow. This allows, e.g. a failed workflow to be restarted from a specific point after modifying the workflow template or even the outputs of completed steps. For example, submit a workflow with reused steps by wf.submit(reuse_step=[step0, step1])
. Here, step0
and step1
are previously run steps returned by query_step
method. Before the new workflow runs a step, it will detect if there exists a reused step with a matching key. If a match is found, the workflow will skip the step and use the outputs from the reused step. To modify the outputs of a step before reusing, use step0.modify_output_parameter(par_name, value)
for parameters and step0.modify_output_artifact(art_name, artifact)
for artifacts.
3.1.14. Executor
For a container/script step (a step using script OP template), by default the Shell script or Python script runs in the container directly. Alternatively, you can modify the executor to run the script differently. Dflow offers an extension point for container/script step Step(..., executor=my_executor)
. Here, my_executor
should be an instance of class derived from the abstract class Executor
. An implementation class of Executor
should implement a method render
which converts original template to a new template.
class Executor(ABC):
@abc.abstractmethod
def render(self, template):
pass
A context is similar to an executor, but assigned to a workflow Workflow(context=...)
, affecting every step.
3.1.15. Submit HPC/Bohrium job via dispatcher plugin
DPDispatcher is a python package used to generate HPC scheduler systems (Slurm/PBS/LSF) or Bohrium jobs input scripts, submit these scripts and poke until they finish. Dflow provides a simple interface to utilize dispatcher as an executor to accomplish script steps. E.g.
from dflow.plugins.dispatcher import DispatcherExecutor
Step(
...,
executor=DispatcherExecutor(host="1.2.3.4",
username="myuser",
queue_name="V100")
)
For SSH authentication, you can either specify the path of a private key file locally, or upload an authorized private key to each node (or equivalently add each node to the authorized host list). For configuring additional machine, resources or task parameters for dispatcher, use DispatcherExecutor(..., machine_dict=m, resources_dict=r, task_dict=t)
.
3.1.16. Submit Slurm job via virtual node
Following the installation steps in the wlm-operator project to add Slurm partitions as virtual nodes to Kubernetes (use manifests configurator.yaml, operator-rbac.yaml, operator.yaml in this project which modified some RBAC configurations)
$ kubectl get nodes
NAME STATUS ROLES AGE VERSION
minikube Ready control-plane,master 49d v1.22.3
slurm-minikube-cpu Ready agent 131m v1.13.1-vk-N/A
slurm-minikube-dplc-ai-v100x8 Ready agent 131m v1.13.1-vk-N/A
slurm-minikube-v100 Ready agent 131m v1.13.1-vk-N/A
Then you can assign a step to be executed on a virtual node (i.e. submit a Slurm job to the corresponding partition to complete the step)
step = Step(
...
executor=SlurmJobTemplate(
header="#!/bin/sh\n#SBATCH --nodes=1",
node_selector={"kubernetes.io/hostname": "slurm-minikube-v100"}
)
)
3.1.17. Use resources in Kubernetes
A step can also be completed using a Kubernetes resource (e.g. Job or custom resources). At the beginning, a manifest is applied to Kubernetes. Then the status of the resource is monitered until the success condition or the failure condition is satisfied.
class Resource(ABC):
action = None
success_condition = None
failure_condition = None
@abc.abstractmethod
pass
3.1.18. Important note: variable names
Dflow has following restrictions on variable names.
Variable name | Static/Dynamic | Restrictions | Example |
---|
Workflow/OP template name | Static | Lowercase RFC 1123 subdomain (must consist of lower case alphanumeric characters, '-' or '.', and must start and end with an alphanumeric character), not more than 57 characters long | my-name |
Step/Task name | Static | Must consist of alpha-numeric characters or '-', and must start with an alpha-numeric character | My-name1-2, 123-NAME |
Parameter/Artifact name | Static | Must consist of alpha-numeric characters, '_' or '-' | my_param_1, MY-PARAM-1 |
Key name | Dynamic | Lowercase RFC 1123 subdomain (must consist of lower case alphanumeric characters, '-' or '.', and must start and end with an alphanumeric character) | my-name |
3.1.19. Debug mode: dflow independent of Kubernetes
The debug mode is enabled by setting
from dflow import config
config["mode"] = "debug"
Before running a workflow locally, ensure that the dependencies of all OPs within the workflow are properly configured in the locally environment. This is not necessary when using the dispatcher executor to submit jobs to some remote environments. The debug mode uses the current directory as the working directory by default. Each workflow will generate a new directory there, whose structure will be like
python-lsev6
├── status
└── step-penf5
├── inputs
│ ├── artifacts
│ │ ├── dflow_python_packages
│ │ ├── foo
│ │ └── idir
│ └── parameters
│ ├── msg
│ └── num
├── log.txt
├── outputs
│ ├── artifacts
│ │ ├── bar
│ │ └── odir
│ └── parameters
│ └── msg
├── phase
├── script
├── type
└── workdir
├── ...
The top level contains the the workflow's status and all its steps. The directory name for each step will be its key if provided, or generated from its name otherwise. Each step directory contains the input/output parameters/artifacts, type and phase of the step. For a step of type "Pod", its directory also includes the script, log file and working directory for the step.
3.1.20. Artifact storage plugins
The default storage for artifacts in dflow is a Minio deployment in the Kubernetes cluster. While other artifact storages are supported (e.g. Aliyun OSS, Azure blob storage (ABS), Google cloud storage(GCS)). Dflow also offers an extension point for implementing customized storage in the artifact management. This can be achieved through a storage client StorageClient
, a class implementing 5 abstract methods, upload
, download
, list
, copy
and get_md5
(optional), which offer the functionality of uploading file, downloading file, listing files with specific prefix, copying file on the server side and getting the MD5 sum of file, respectively. Use a custom storage client object by configuring s3_config["storage_client"].
class StorageClient(ABC):
@abc.abstractmethod
def upload(self, key: str, path: str) -> None:
pass
@abc.abstractmethod
def download(self, key: str, path: str) -> None:
pass
@abc.abstractmethod
def list(self, prefix: str, recursive: bool = False) -> List[str]:
pass
@abc.abstractmethod
def copy(self, src: str, dst: str) -> None:
pass
@abc.abstractmethod
def get_md5(self, key: str) -> str:
pass
3.2. Python OP
3.2.1. Slices
In coordination with parallel steps, Slices
assists user to slice input parameters/artifacts (which must be lists) to feed parallel steps and stack their output parameters/artifacts into lists following the same pattern. The Python OP only need to handle one slice. For example,
step = Step(name="parallel-tasks",
template=PythonOPTemplate(
...,
slices=Slices(
input_parameter=["msg"],
input_artifact=["data"],
output_artifact=["log"])
),
parameters = {
"msg": msg_list
},
artifacts={
"data": data_list
},
with_param=argo_range(5)
)
In this example, each item in msg_list
is passed to a parallel step as the input parameter msg
, each part in data_list
is passed to a parallel step as the input artifact data
. Finally, the output artifacts log
of all parallel steps are collected to a single artifact step.outputs.artifacts["log"]
.
This example is analogous to the following pseudocode
log = [None] * 5
for item in range(5):
log[item] = my_op(msg=msg_list[item], data=data_list[item])
where with_param
and slices
corresponds to the for
loop and the statement in loop in the pseudocode, respectively.
Note that this feature merges the sliced output artifacts by default, which means the relative path in the output artifact of each slice is remained in the merged artifact. This could potentially lead to overwriting issues due to path conflicts. To create a separate dir for each slice for saving output artifacts, set create_dir=True
.
It should be noticed that this feature, by default, passes full input artifacts to each parallel step which may only use some slices of these artifacts. In contrast, the subpath mode of slices only passes one single slice of the input artifacts to each parallel step. To use the subpath mode of slices,
step = Step(name="parallel-tasks",
template=PythonOPTemplate(
...,
slices=Slices(sub_path=True,
input_parameter=["msg"],
input_artifact=["data"],
output_artifact=["log"])
),
parameters = {
"msg": msg_list
},
artifacts={
"data": data_list
})
Here, each input parameter and artifact to be sliced must be of the same length, and the parallelism equals to this length. Another notable point is that in order to use the subpath of the artifacts, these artifacts must be stored without compression when they are generated. E.g. declare Artifact(..., archive=None)
in the output signs of Python OP, or specify upload_artifact(..., archive=None)
while uploading artifacts. Besides, you can use dflow.config["archive_mode"] = None
to set default archive mode to no compression globally.
3.2.2. Retry and error handling
Dflow catches TransientError
and FatalError
thrown from OP
. User can set maximum number of retries on TransientError
by PythonOPTemplate(..., retry_on_transient_error=n)
. Timeout error is regarded as fatal error for default. To treat timeout error as transient error, set PythonOPTemplate(..., timeout_as_transient_error=True)
. When a fatal error is raised or the retries on transient error reaches maximum retries, the step is considered as failed.
3.2.3. Progress
A OP
can update progress in the runtime so that user can moniter its real-time progress from the frontend UI
class Progress(OP):
progress_total = 100
...
def execute(op_in):
for i in range(10):
self.progress_current = 10 * (i + 1)
...
3.2.4. Upload python packages for development
To avoid frequently creating images during development, dflow offers an interface to upload local packages into the container and add them to $PYTHONPATH
, such as PythonOPTemplate(..., python_packages=["/opt/anaconda3/lib/python3.9/site-packages/numpy"])
. Additionally, you can globally specify packages to be uploaded, which will impact all OP
s
from dflow.python import upload_packages
upload_packages.append("/opt/anaconda3/lib/python3.9/site-packages/numpy")