A system to integrate data from multiple workflows.
Project description
FlowCept
FlowCept is a system for runtime data integration of data processed by multiple workflows, allowing users (scientists, engineers) to understand, at runtime, complex, heterogeneous, large-scale data coming from various sources.
FlowCept is intended to address scenarios where multiple workflows in a science campaign or in an enterprise run and generate important data to be analyzed in an integrated manner. Since these workflows may use different data generation tools or can be executed within different parallel computing systems (e.g., Dask, Spark, workflow management systems), its key differentiator is the capability to seamless integrate data from various sources. By using provenance data management techniques, it builds an integrated data view at runtime of these multi-workflow data following W3C PROV recommendations for its data schema. By using data observability, it does not require changes in user codes or systems (i.e., instrumentation). All users need to do is to create adapters for their systems or tools, if one is not available yet.
Currently, FlowCept provides adapters for: Dask, MLFlow, TensorBoard, and Zambeze.
See the Jupyter Notebooks for utilization examples.
See the Contributing file for guidelines to contribute with new adapters. Note that we may use the term 'plugin' in the codebase as a synonym to adapter. Future releases should standardize the terminology to use adapter.
Install and Setup:
- Install FlowCept:
pip install .[full]
in this directory (or pip install flowcept[full]
).
For convenience, this will install all dependencies for all adapters. But it can install
dependencies for adapters you will not use. For this reason, you may want to install
like this: pip install .[adapter_key1,adapter_key2]
for the adapters we have implemented, e.g., pip install .[dask]
.
See extra_requirements if you want to install the dependencies individually.
- Start MongoDB and Redis:
To enable the full advantages of FlowCept, the user needs to run Redis, as FlowCept's message queue system, and MongoDB, as FlowCept's main database system. The easiest way to start Redis and MongoDB is by using the docker-compose file for its dependent services: MongoDB and Redis. You only need RabbitMQ if you want to observe Zambeze messages as well.
-
Define the settings (e.g., routes and ports) accordingly in the settings.yaml file.
-
Start the observation using the Controller API, as shown in the Jupyter Notebooks.
-
To use FlowCept's Query API, you need to start the flask webserver:
python flowcept/flowcept_webserver/app.py
. Query API utilization examples are available in the notebooks.
Performance Tuning for Performance Evaluation
In the settings.yaml file, the following variables might impact interception performance:
main_redis:
buffer_size: 50
insertion_buffer_time_secs: 5
plugin:
enrich_messages: false
And other variables depending on the Plugin. For instance, in Dask, timestamp creation by workers add interception overhead.
Plugins-specific info
You can run pip install flowcept[plugin_name]
to install requirements for a specific plugin, instead of installing the
whole package.
RabbitMQ for Zambeze plugin
$ docker run -it --rm --name rabbitmq -d -p 5672:5672 -p 15672:15672 rabbitmq:3.11-management
Tensorboard
If you're on mac, pip install
may not work out of the box because of Tensorflow library.
You may need to pip install tensorflow-macos
instead of the tensorflow
lib available in the tensorboard-requirements.
See also
Acknowledgement
This research uses resources of the Oak Ridge Leadership Computing Facility at the Oak Ridge National Laboratory, which is supported by the Office of Science of the U.S. Department of Energy under Contract No. DE-AC05-00OR22725.
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.