adhesive fork with support for Zeebe XML extension
Project description
This is a fork of Adhesive to handle Zeebe Modeler BPMN workflows.
Changes are:
Support for Zeebe Modeler BPMN xml files.
Can be used in a thread.
Should be used as a library from another app.
The following documentation is the original documentation and might not work as excepted when adhesive is used as a cli command.
Adhesive is a micro BPMN runner written in Python.
You can easily model complex logic in BPMN, and Adhesive will execute it for you taking care of parallelism, joining, etc. using the standard BPMN notation.
Since it’s small, it can easily be embedded in containers, or replace complex scripts.
Installation
pip install adhesive
Getting Started
Simple Builds
To create a basic build you just create a file in your project named _adhesive.py. In it you then declare some tasks. For example:
import adhesive
@adhesive.task("Checkout Code")
def checkout_code(context):
adhesive.scm.checkout(context.workspace)
@adhesive.task("Run Build")
def run_build(context):
context.workspace.run("mvn clean install")
adhesive.build()
Since no process was defined, adhesive takes the defined tasks, stitches them in order, and has a process defined as <start> → Checkout Code → Run Build → <end>.
To run it simply call adhesive in the terminal:
adhesive
This is the equivalent of Jenkins stages. But we can do better:
Programmatic Builds
In order to use the full programmatic functionalities that adhesive offers, you are able to stitch your BPM process manually. You have sub processes, branching and looping available:
import adhesive
import uuid
@adhesive.task("Run in parallel item {loop.value}")
def context_to_run(context):
if not context.data.executions:
context.data.executions = set()
context.data.executions.add(str(uuid.uuid4()))
data = adhesive.process_start()\
.branch_start()\
.sub_process_start() \
.task("Run in parallel",
loop="items") \
.sub_process_end()\
.branch_end() \
.branch_start() \
.sub_process_start() \
.task("Run in parallel item {loop.value}",
loop="items") \
.sub_process_end() \
.branch_end() \
.process_end()\
.build(initial_data={"items": [1, 2, 3, 4, 5]})
assert len(data.executions) == 10
Here you see the full BPMN power starting to unfold. We create a process that branches out, creates sub processes (sub processes can be looped as a single unit). Loops are creating execution tokens that also run in parallel in the same pool.
Note that you can pass initial_data into the process, and you can also get the context.data from the last execution token.
BPMN Process
Last but not least, adhesive reads BPMN files, and builds the process graph from them. This is particularly good if the process is complex and has a lot of dependencies:
The build of adhesive is modeled as a BPMN process itself, so we load it from the file directly using: adhesive.build_bpmn("adhesive-self.bpmn")
import adhesive
@adhesive.task("Read Parameters")
def read_parameters(context) -> None:
context.data.run_mypy = False
context.data.test_integration = True
@adhesive.task(re=r"^Ensure Tooling:\s+(.+)$")
def gbs_ensure_tooling(context, tool_name) -> None:
ge_tooling.ensure_tooling(context, tool_name)
# ...
adhesive.build_bpmn("adhesive-self.bpmn")
As you see steps are parametrizable, and use the data from the task name into the step definition.
Defining BPMN Tasks
For example here, we define an implementation of tasks using regex matching, and extracting values:
@adhesive.task(re=r"^Ensure Tooling:\s+(.+)$")
def gbs_ensure_tooling(context, tool_name) -> None:
# ...
Or a user task (interactive form):
@adhesive.usertask('Publish to PyPI?')
def publish_to_pypi_confirm(context, ui):
ui.add_checkbox_group(
"publish",
title="Publish",
values=(
("nexus", "Publish to Nexus"),
("pypitest", "Publish to PyPI Test"),
("pypi", "Publish to PyPI"),
),
value=("pypitest", "pypi")
)
Don’t forget, the @adhesive.task and @adhesive.usertask are just defining mappings for implementations of the task names available in the process. Only the adhesive.build() creates a linear process out of the declaration of the tasks.
As you notice, there’s always a first parameter named context. The context parameter contains the following information:
task - the Task in the graph that’s currently matched against this execution.
task_name - The resolved name, with the variables interpolated. Matching is attempted after the name is resolved.
data - Data that the current execution token contains. This data is always cloned across executions, and `set`s and `dict`s are automatically merged if multiple execution tokens are merged. So you have a modifiable copy of the data that you’re allowed to change, and is propagated into the following execution tokens.
loop - if the current task is in a loop, the entry contains its index, the key and value of the items that are iterating, and the expression that was evaluated. Note that loop execution happens in parallel since these are simple execution tokens.
lane - the current lane where the tasks belongs. Implicitly it’s default.
workspace - a way to interact with a system, and execute commands, create files, etc.
adhesive runs all the tasks on a parallel process pool for better performance. This happens automatically.
The tasks perform the actual work for the build. But in order to have that, we need to be able to execute commands, and create files. For that we have the workspace.
Start Event Messages
Adhesive supports also start events with messages in the process. Each message start event, is being processed in its own thread and yield results:
@adhesive.message('Generate Event')
def message_generate_event(context):
for i in range(10):
yield i
@adhesive.task('Process Event')
def process_event(context):
print(f"event data: {context.data.event}")
Each yield generates a new event that fires up the connected tasks. The data yielded is present in the event attribute in the token, for the following tasks.
Callback Messages
The other option to push messages into a process is to use callback messages:
@adhesive.message_callback('REST: /rest/process-resource')
def message_rest_rest_process_resource(context, callback):
@app.route("/rest/resource/create")
def create_resource():
callback(Dict({
"type": "CREATE"
}))
return "Create event fired"
Using this we’re able to hook into other systems that have their own loop, such as in this case the Flask server, and push messages using the callback. This approach has also the advantage of not creating new threads for each message endpoint.
Connections
Tasks are linked using connections. In some cases, connections can have conditions. Conditions are expressions that when evaluated to True will allow the token to pass the connection. In the connection there is access to the task, task_name, data, loop, lane and context, as well as the variables defined in the context.data.
So if in a task there is defined a data field such as:
@adhesive.task('prepare data')
def prepare_data(context):
context.data.navigation_direction = "forward"
The navigation_direction can be validated in the condition with any of the following:
context.data.navigation_direction == "forward"
data.navigation_direction == "forward"
navigation_direction == "forward"
Workspace
Workspaces are just a way of interacting with a system, running commands, and writing/reading files. Currently there’s support for:
the local system
docker containers
kubernetes
remote SSH connections
When starting adhesive allocates a default workspace folder in the configured temp location (implicitly /tmp/adhesive). The Workspace API is an API that allows you to run commands, and create files, taking care of redirecting outputs, and even escaping the commands to be able to easily run them inside docker containers.
The workspace is available from the cotext directly from the context, by calling context.workspace.
For example calling context.workspace.run(…) will run the command on the host where adhesive is running:
@adhesive.task("Run Maven")
def build_project(context) -> None:
context.workspace.run("mvn clean install")
If we’re interested in the program output we simply do a run with a capture_stdout that returns the output as a string:
@adhesive.task("Test")
def gbs_test_linux(context) -> None:
content = context.workspace.run("echo yay", capture_stdout=True)
assert content == "yay"
or we can use the simplified call with run_output that guarantees a str as result, unlike the Optional[str] for run:
@adhesive.task("Test")
def gbs_test_linux(context) -> None:
content = context.workspace.run_output("echo yay")
assert content == "yay"
The run commands implicitly use /bin/sh, but a custom shell can be specified by passing the shell argument:
content = context.workspace.run_output("echo yay", shell="/bin/bash")
Docker Workspace
To create a docker workspace that runs inside a container with the tooling you just need to:
from adhesive.workspace import docker
Then to spin up a container that has the current folder mounted in, where you’re able to execute commands inside the container. You just need to:
@adhesive.task("Test")
def gbs_test_linux(context) -> None:
image_name = 'some-custom-python'
with docker.inside(context.workspace, image_name) as w:
w.run("python -m pytest -n 4")
This creates a container using our current context workspace, where we simply execute what we want, using the run() method. After the with statement the container will be teared down automatically.
SSH Workspace
In order to have ssh, make sure you installed adhesive with SSH support:
pip install -U adhesive[ssh]
To have a SSH Workspace, it’s again the same approach:
from adhesive.workspace import ssh
Then to connect to a host, you can just use the ssh.inside the same way like in the docker sample:
@adhesive.task("Run over SSH")
def run_over_ssh(context) -> None:
with ssh.inside(context.workspace,
"192.168.0.51",
username="raptor",
key_fileaname="/home/raptor/.ssh/id_rsa") as s:
s.run("python -m pytest -n 4")
The parameters are being passed to paramiko, that’s the implementation beneath the SshWorkspace.
Kubernetes Workspace
To run things in pods, it’s the same approach:
from adhesive.workspace import kube
Then we can create a workspace to run things in kubernetes pods. The workspace, as well as the API, will use the kubectl command internally.
@adhesive.task("Run things in the pod")
def run_in_the_pod(context) -> None:
with kube.inside(context.workspace,
pod_name="nginx-container") as pod:
pod.run("ps x") # This runs in the pod
Kubernetes API
Adhesive also packs a kubernetes api, that’s available on the adhesive.kubeapi:
from adhesive.kubeapi import KubeApi
To use it, we need to create an instance against a workspace.
@adhesive.gateway('Determine action')
def determine_action(context):
kubeapi = KubeApi(context.workspace,
namespace=context.data.target_namespace)
Let’s create a namespace:
kubeapi.create(kind="ns", name=context.data.target_namespace)
Or let’s create a service using the kubectl apply approach:
kubeapi.apply(f"""
apiVersion: v1
kind: Service
metadata:
name: nginx-http
labels:
app: {context.data.target_namespace}
spec:
type: ClusterIP
ports:
- port: 80
protocol: TCP
name: http
selector:
app: {context.data.target_namespace}
""")
Or let’s get some pods:
pod_definitions = kubeapi.getall(
kind="pod",
filter=f"execution_id={context.execution_id}",
namespace=context.data.target_namespace)
These returns objects that allow navigating properties as regular python attributes:
new_pods = dict()
for pod in pod_definitions:
if not pod.metadata.name:
raise Exception(f"Wrong definition {pod}")
new_pods[pod.metadata.name] = pod.status.phase
You can also navigate properties that are not existing yet, for example to wait for the status of a pod to appear:
@adhesive.task('Wait For Pod Creation {loop.key}')
def wait_for_pod_creation_loop_value_(context):
kubeapi = KubeApi(context.workspace,
namespace=context.data.target_namespace)
pod_name = context.loop.key
pod_status = context.loop.value
while pod_status != 'Running':
time.sleep(5)
pod = kubeapi.get(kind="pod", name=pod_name)
pod_status = pod.status.phase
To get the actual data from the wrappers that the adhesive API creates, you can simply call the _raw property.
Workspace API
Here’s the full API for it:
class Workspace(ABC):
"""
A workspace is a place where work can be done. That means a writable
folder is being allocated, that might be cleaned up at the end of the
execution.
"""
@abstractmethod
def write_file(
self,
file_name: str,
content: str) -> None:
pass
@abstractmethod
def run(self,
command: str,
capture_stdout: bool = False) -> Union[str, None]:
"""
Run a new command in the current workspace.
:param capture_stdout:
:param command:
:return:
"""
pass
@abstractmethod
def rm(self, path: Optional[str]=None) -> None:
"""
Recursively remove the file or folder given as path. If no path is sent,
the whole workspace will be cleared.
:param path:
:return:
"""
pass
@abstractmethod
def mkdir(self, path: str=None) -> None:
"""
Create a folder, including all its needed parents.
:param path:
:return:
"""
pass
@abstractmethod
def copy_to_agent(self,
from_path: str,
to_path: str) -> None:
"""
Copy the files to the agent from the current disk.
:param from_path:
:param to_path:
:return:
"""
pass
@abstractmethod
def copy_from_agent(self,
from_path: str,
to_path: str) -> None:
"""
Copy the files from the agent to the current disk.
:param from_path:
:param to_path:
:return:
"""
pass
@contextmanager
def temp_folder(self):
"""
Create a temporary folder in the current `pwd` that will be deleted
when the `with` block ends.
:return:
"""
pass
@contextmanager
def chdir(self, target_folder: str):
"""
Temporarily change a folder, that will go back to the original `pwd`
when the `with` block ends. To change the folder for the workspace
permanently, simply assing the `pwd`.
:param target_folder:
:return:
"""
pass
User Tasks
In order to create user interactions, you have user tasks. These define form elements that are populated in the context.data, and available in subsequent tasks.
When a user task is encountered in the process flow, the user is prompted to fill in the parameters. Note that the other started tasks continue running, proceeding forward with the build.
The name used in the method call defines the name of the variable that’s in the context.data.
For example in here we define a checkbox group that allows us to pick where to publish the package:
@adhesive.usertask("Read User Data")
def read_user_data(context, ui) -> None:
ui.add_input_text("user",
title="Login",
value="root")
ui.add_input_password("password",
title="Password")
ui.add_checkbox_group("roles",
title="Roles",
value=["cyborg"],
values=["admin", "cyborg", "anonymous"])
ui.add_radio_group("disabled", # title is optional
values=["yes", "no"],
value="no")
ui.add_combobox("machine",
title="Machine",
values=(("any", "<any>"),
("win", "Windows"),
("lin", "Linux")))
This will prompt the user with this form:
This data is also available for edge conditions, so in the BPMN modeler we can define a condition such as "pypi" in context.data.roles, or since data is also available in the edge scope: "pypi" in data.roles.
The other option is simply reading what the user has selected in a following task:
@adhesive.task("Register User")
def publish_items(context):
for role in context.data.roles:
# ...
User tasks support the following API, available on the ui parameter, the parameter after the context:
class UiBuilderApi(ABC):
def add_input_text(self,
name: str,
title: Optional[str] = None,
value: str = '') -> None:
def add_input_password(self,
name: str,
title: Optional[str] = None,
value: str = '') -> None:
def add_combobox(self,
name: str,
title: Optional[str] = None,
value: Optional[str]=None,
values: Optional[Iterable[Union[Tuple[str, str], str]]]=None) -> None:
def add_checkbox_group(
self,
name: str,
title: Optional[str]=None,
value: Optional[Iterable[str]]=None,
values: Optional[Iterable[Union[Tuple[str, str], str]]]=None) -> None:
def add_radio_group(self,
name: str,
title: Optional[str]=None,
value: Optional[str]=None,
values: Optional[List[Any]]=None) -> None:
def add_default_button(self,
name: str,
title: Optional[str] = None,
value: Optional[Any] = True) -> None:
Secrets
Secrets are files that contain sensitive information are not checked in the project. In order to make them available to the build, we need to define them in either ~/.adhesive/secrets/SECRET_NAME or in the current folder as .adhesive/secrets/SECRET_NAME.
In order to make them available, we just use the secret function that creates the file in the current workspace and deletes it when exiting. For example here’s how we’re doing the actual publish, creating the secret inside a docker container:
@adhesive.task('^PyPI publish to (.+?)$')
def publish_to_pypi(context, registry):
with docker.inside(context.workspace, context.data.gbs_build_image_name) as w:
with secret(w, "PYPIRC_RELEASE_FILE", "/germanium/.pypirc"):
w.run(f"python setup.py bdist_wheel upload -r {registry}")
Note the docker.inside that creates a different workspace.
Configuration
Adhesive supports configuration via its config files, or environment variables. The values are read in the following order:
environment variables: ADHESIVE_XYZ, then
values that are in the project config yml file: .adhesive/config.yml, then
values configured in the global config yml file: $HOME/.adhesive/config.yml.
Currently the following values are defined for configuration:
temp_folder
default value /tmp/adhesive, environment var: ADHESIVE_TEMP_FOLDER.
Is where all the build files will be stored.
plugins
default value [], environment var: ADHESIVE_PLUGINS_LIST.
This contains a list of folders, that will be added to the sys.path. So to create a reusable plugin that will be reused by multiple builds, you need to simply create a folder with python files, then point to it in the ~/.adhesive/config.yml:
plugins:
- /path/to/folder
Then in the python path you can simply do regular imports.
color
default value True, environment var: ADHESIVE_COLOR.
Marks if the logging should use ANSI colors in the terminal. Implicitly this is true, but if log parsing is needed, it can make sense to have it false.
log_level
default_value info, environment var: ADHESIVE_LOG_LEVEL.
How verbose should the logging be on the terminal. Possible values are trace, debug, info, warning, error and critical.
pool_size
default value is empty, environment var: ADHESIVE_POOL_SIZE.
Sets the number of workers that adhesive will use. Defaults to the number of CPUs if unset.
stdout
default value is empty, environment var: ADHESIVE_STDOUT.
Implicitly for each task, the log is redirected in a different file, and only shown if the task failed. The redirection can be disabled.
parallel_processing
default value is thread, environment var: ADHESIVE_PARALLEL_PROCESSING.
Implicitly tasks are scaled using multiple threads in order to alleviate waits for I/O. This is useful for times when remote ssh workspaces are defined in the lanes, so the same connection can be reused for multiple tasks.
This value can be set to process, in case the tasks are CPU intensive. This has the drawback of recreating the connections on workspaces’ each task execution.
Hacking Adhesive
Adhesive builds with itself. In order to do that, you need to checkout the adhesive-lib shared plugin, and configure your local config to use it:
plugins:
- /path/to/adhesive-lib
Then simply run the build using adhesive itself:
adhesive
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for adhesive_zeebe-2021.4.3-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 0da426eabdd068c85e5f3533d6418f55c6127d93ebd2fb46ee8960926b5eecd8 |
|
MD5 | 17d433000ef39d3da02f429446ee1af9 |
|
BLAKE2b-256 | 9f4bdb4a52fe0fe2b9c0235c9abec79d5a8e6967571e3ead70a10ee931bb0f95 |