borealis-fireworks 0.2.0

Run FireWorks workflows in Google Cloud

Borealis

Runs FireWorks workflows on Google Compute Engine (GCE).

See the repo Borealis.

• Borealis is the git repo name.
• borealis-fireworks is the PyPI package name.
• borealis-fireworker.service is the name of the systemd service.
• fireworker is the recommended process username and home directory name.

Background

You can launch as many Fireworker nodes as you want as Google Compute Engine (GCE) VM instances, and/or run local workers, as long as they can all connect to the LaunchPad server running MongoDB. Metadata parameters and the worker's gce_my_launchpad.yaml file (if that file doesn't exist, then my_launchpad.yaml) configure the MongoDB host, port, and DB name. Users can have their own DB names on a shared MongoDB server, and each user can have multiple DB names -- each an independent launchpad space for workflows and their Fireworker nodes.

Workers get Fireworks from the LaunchPad, run them in "rapidfire" mode, and eventually time out and shut themselves down.

Workers can run any Firetasks that are loaded on their disk images, but the best fit is to run the DockerTask Firetask. DockerTask pulls task input files from Google Cloud Storage (GCS), runs a payload task as a shell command within a Docker container, and pushes task output files to GCS.

DockerTask parameters include the Docker image to pull, the command shell tokens to run in the Docker container, and its input and output files and directories.

DockerTask pulls the inputs from and pushes the outputs to Google Cloud Storage (GCS). This avoids needing a shared NFS file service which costs 10x as much as GCS storage and doesn't scale as well.

Using a Docker image lets you bundle up the payload task with its entire runtime, e.g. Python version, pips, Linux apts, and config files. Your workflow can use one or more Docker images, and they're isolated from the Fireworker.

Team Setup

TODO: Install & configure dev tools, create a GCP project, auth stuff, install MongoDB on a GCE VM or set up Google-managed MongoDB, create a Fireworker disk image & image family, ...

Individual Developer Setup

TODO: Install & configure dev tools, make a storage bucket with a globally-unique name, build a Docker image to run, ...

Run

TODO

Change Log

v0.2.0 - 2020-02-13

• Read launchpad config info from gce_my_launchpad.yaml if possible, falling back to my_launchpad.yaml for compatibility. This lets people use one launchpad config file for their GCE workflows and another one for their other workflows.
• Improve the server installation steps and augment the fireworker --help text to display its directory.

v0.1.1 - 2020-02-13

• Correct the pip name in startup.sh.
• Use print() instead of logging in gce.py so the messages aren't filtered by the log level.
• Refine the installation instructions.

v0.1.0 - 2020-02-10

• Initial dev build.