Skip to main content

bisos.airflow: Apache Airflow platform management via BISOS Capability Bundles

Overview

bisos.airflow is a BISOS package for managing an Apache Airflow platform (webserver, scheduler, triggerer, database, and DAG development/testing) as a set of BISOS Capability Bundles (CBS/CBM) and systemd-planted Command Services. It is under active initial development; this README is a starting point and will be expanded as the package matures.

The remainder of this document is carried over from bisos.dockerProc *(the starting-point template this package was copied from) and still describes docker/podman container lifecycle management. It will be rewritten to describe bisos.airflow as the package develops.*

Capability Materialization

The figure below shows the top-down layering of airflow.here’s Capability Materialization: the planted CBM pointer leaf (/bisos/platform/sys/cbm/collective/hereWeb/airflow/cbmProc.spcs) points to (and does not duplicate) the CBS (py3/bin/airflow-cbs.pcs), which declares the bundle of airflow-sbom.pcs, the 4 airflow-*-sysd.pcs units, and airflow-assemble.cs. py3/bin/airflowAdmin.cs sits below all of that as the day-to-day admin-facing CS an operator runs against the materialized installation.

Regenerate this figure (after editing py3/images/airflow-graphviz.pcs):

cd py3/images
./airflow-graphviz.pcs --format=all -i ngProcess all
The companion image-specifications repo lives at
https://github.com/bxObjects/bro_dockerfiles — a set of Dockerfiles for XFCE4 desktops over VNC / noVNC / SSH on Debian 12 and 13, organised by init/isolation profile.

bisos.dockerProc is a python package that uses the PyCS-Framework. It is a BISOS-Capability and a Standalone-BISOS-Package.

Architectural context. bisos.dockerProc is one realization of a general BISOS facility for building walkable trees on top of .spcs (Spread Planted Command Services) — currently the first realization, with bisos.lcnt (LaTeX content), static-web publication, and others planned. The .spcs mechanism itself lives in bisos.csSeed; the tree walker and WalkExampleSpec dataclass live in bisos.fileObj; the architecture that composes them (three-box layered figure, authoring recipe, canonical references) is documented at `bisos.pycs/spcs/README.org </bisos/git/auth/bxRepos/bisos-pip/pycs/spcs/README.org>`__. This README describes what makes this realization specific — path-anchoring on debian/, per-profile ports, cgroup variants, container-in-container safety, backup workflow, the six-image matrix of the companion bro_dockerfiles tree.

Package Documentation At Github

The information below is a subset of the full of documentation for this bisos-pip package. More complete documentation is available at: https://github.com/bisos-pip/capability-cs

Use Cases

bisos.dockerProc addresses several deployment scenarios in which the usual “one Dockerfile, one docker run line” pattern falls short. If any of the following describes your situation, this package is built for you.

RawBisos-Reconstruction — BISOS’s Equivalent of the Common CI/CD Pipeline

BISOS’s equivalent of the common CI/CD pipeline is called RawBisos-Reconstruction. Two execution substrates exist:

  • Vagrant-based variant (canonical, in bxGenesis/start) — provisions a full VM from scratch, runs raw-bisos.sh -i installUnsitedBisos against it, validates the resulting environment. Heavyweight; suitable for full-release validation.

  • Container-based variant (this package + bro_dockerfiles) — provisions a fresh Debian container, runs installRawBisos.sh inside it, validates the resulting environment. Lightweight; suitable for on-demand validation during development.

The container-based variant runs entirely from planted .spcs files. No CI service, no VM lifecycle. The six bro_dockerfiles leaves (deb12/deb13 × confined/privileged/rootless-sysd) are the test matrix.

Typical one-shot per variant (from any leaf):

cd bro_dockerfiles/debian/12/confined/vnc/xfce/bisos_deb12-fresh
./dockerProc.spcs -i containerProc_imageBuild
./dockerProc.spcs -i containerProc_instanceUp
./dockerProc.spcs -i containerProc_exec_installRawBisos    # bxGenesis bootstrap inside container
./dockerProc.spcs -i containerProc_instanceVerify
./dockerProc.spcs -i containerProc_fullClean

Or across the whole matrix from bro_dockerfiles/debian/:

./ftoBranchProc.spcs -i fto_forwardToLeaves --cmndName=containerProc_exec_installRawBisos

Each of the six leaves subprocess-runs the bootstrap independently (via bisos.fileObj Stage 2 Deliverable 6 subprocess-per-branch). Results stream to stdout; a failing bootstrap on one variant does not mask successful bootstraps on the others.

The installRawBisos.sh script (baked into each image at ~/raw-bisos/) refreshes raw-bisos.sh from bxGenesis/start via wget and invokes it non-interactively with -v -n showRun -i installUnsitedBisos. This means every reconstruction is against the current head of bxGenesis/start, not a pinned snapshot — catches regressions as they land.

Reproducible XFCE Desktops on Debian 12 and 13

You want an XFCE4 desktop, accessible via VNC / noVNC / SSH, built the same way on Debian 12 and Debian 13 so team members on either release get a consistent environment. Instead of maintaining two divergent Dockerfiles, bro_dockerfiles has one leaf per (release × profile × image), and dockerProc.spcs builds and runs any of them from its own directory. Same commands, different leaves.

Choosing Among Three Init/Isolation Profiles for the Same Image

The same desktop content can be shipped under three fundamentally different container postures — and which one to use depends on the host’s security policy, not the image content:

  • confined — unprivileged docker; no systemd; entrypoint script launches the services. Runs anywhere docker runs.

  • privileged — docker with --privileged; systemd as PID 1; services managed as systemd units. Full systemd fidelity, at the cost of a privileged container.

  • rootless-sysd — rootless podman; systemd as PID 1; unprivileged posture. Requires cgroup v2 with controller delegation on the host. Delivers systemd fidelity and an unprivileged security posture.

The three profiles sit side-by-side in the bro_dockerfiles tree — a single leaf-path change switches profile. This is unusual: most projects pick one and hide the others.

Multi-Tenant: N Engineers on One Powerful VM

One VM, many users, each user running one or more container instances of their own. This is the scenario that motivates the Platform Registrar (design-only in this phase; see Full Design at a Glance): per-user, per-image, per-instance port assignments so users don’t collide on the fixed 2222/5901/6901 host-side port table.

The rootless-sysd profile is what makes this scenario safe — each user’s containers run in their own user namespace with no --privileged flag, so isolation between users is real and not a policy promise.

Corporate Firewall: Only Port 22 Outbound

A common corporate deployment target: engineers’ workstations can reach the internet, but only outbound port 22 is permitted. VNC, noVNC, and SSH-alt ports to the container platform are blocked. The Remote User’s .ssh/config is populated (from the registrar) with LocalForward entries, so all container-side ports are tunnelled through the single port-22 ssh connection. The user then points local ssh / vnc / browser clients at localhost:<forwarded-port>.

Path-Derived Configuration Instead of Per-Leaf Config Files

You want to add a new image without writing a new configuration file. In bro_dockerfiles the leaf’s path (debian/<release>/<profile>/vnc/xfce/<imageName>/) supplies engine, profile, ports, base image, and cgroup support — no config.yaml, no environment file, no shell variables to set. Copy an existing leaf to a new path, drop in a Dockerfile, plant the appropriate .spcs, and the seed does the rest. Configuration lives in filesystem structure, not in files.

Full Design at a Glance

The figure below captures the full design of bisos.dockerProc — the three execution platforms it spans, the pieces on each, and the flows between them. Part of the design (the Platform Registrar) is not yet implemented and is shown with a dashed border. Scan the legend first, then read the figure.

Legend:

Element

Meaning

Light-blue cluster (left)

Remote User — corporate host, only port 22 outbound.

Pink cluster (middle)

Container Platform — the powerful VM; hosts engines, .spcs, registrar.

Green cluster (right)

Container Instances — one image, many running instances per user.

Dashed border (registrar)

Design-only in this phase — shape reserved, not yet implemented.

Yellow cylinder

Persistent store / path-derived identity source.

Salmon rarrow shape

CS / .spcs dispatcher.

Blue rectangle

Engine / CLI parameter / port.

Purple edge

ssh tunnel path (Remote User → Platform via port 22 → Instances).

Blue dashed edge

Registrar read (design-only; two consumers: .spcs and Remote User).

Dotted “planted in” edge

Which leaves in bro_dockerfiles plant which .spcs.

Reading the figure, left to right:

  • Container Platform (middle) — the anchor. This is the powerful VM where everything runs. Image identity comes from the leaf directory path in bro_dockerfiles (debian/<release>/<profile>/vnc/xfce/<imageName>/): dockerProc.spcs is planted in confined and privileged leaves; podmanProc.spcs is planted in rootless-sysd leaves. Both delegate to the engine-agnostic seed containerProc-seed.cs, which drives docker or podman according to the leaf’s profile.

  • Platform Registrar (middle, dashed) — design-only in this phase. The current implementation derives host-side ports from the leaf path (fixed 2222/2223/… table). This works for one user with one instance per profile. The moment you have multiple users on one VM — or multiple instances of the same image per user — fixed ports collide. The Registrar is the reserved slot for that: a per-user, per-image, per-instance store at /bisos/platform/registrar/dockerProc/<user>/<imageName>/instance0..N/ that assigns and tracks host-side ports. Path gives the image identity; registrar gives the instance identity.

  • Two consumers of the registrar (both drawn as blue dashed edges). .spcs reads assigned ports on the platform side to do run / composeUp / verify for the right host-side ports. Remote User reads the same assignments (over ssh) to generate its local .ssh/config LocalForward entries so its ssh client tunnels the right host ports to the right instances.

  • Remote User (left) — corporate host, port-22-only. Typical deployment target: a workstation behind corporate networking where everything but outbound port 22 is blocked. The user cannot open VNC/noVNC/SSH-alt ports directly to the platform. Instead, the Remote User’s .ssh/config is populated with LocalForward entries that tunnel the assigned host-side ports through the single port-22 ssh connection. The user then points local ssh/vnc/browser clients at localhost:<forwarded-port>. The .ssh/config entries are a generated artefact: their content comes from the registrar.

  • Container Instances (right) — many per user. Each running container exposes the same fixed container-side ports (22 for ssh, 5901 for VNC, 6901 for noVNC). What varies per instance is the host-side mapping. The figure shows image-A / inst0, inst1, image-B / inst0, ... instN as representative — one image can have many instances; a user can have many images.

What is implemented today vs what the figure describes:

  • Implemented: image path → .spcs → seed → engine → running container. The dockerProc.spcs and podmanProc.spcs planting, the seed’s build / composeUp / run / verify commands, and podmanHostVerify.cs for host readiness are all in place. Verified on both docker (deb13 privileged sysd, cgroup v1 + v2) and podman (deb13 rootless-sysd).

  • Not yet implemented: everything downstream of Platform Registrar. Its shape is fixed in the figure so subsequent multi-tenant work can add it without redesign.

Table of Contents TOC

Part of BISOS — ByStar Internet Services Operating System

Layered on top of Debian, BISOS (By* Internet Services Operating System) is a unified and universal framework for developing both internet services and software-service continuums that use internet services. See Bootstrapping ByStar, BISOS and Blee for information about getting started with BISOS.

BISOS is a foundation for The Libre-Halaal ByStar Digital Ecosystem which is described as a cure for losses of autonomy and privacy in a book titled: Nature of Polyexistentials

bisos.dockerProc is part of BISOS. It is a standalone package that can be used independently of the full BISOS environment.

bisos.dockerProc is a Command-Only PyCS Facility

bisos.dockerProc is a command-line tool. It is a PyCS multi-unit command service. PyCS is a framework that converges development of CLI tools and services. PyCS is an alternative to FastAPI, Typer and Click.

bisos.dockerProc uses the PyCS-Framework to:

  1. Provide a seed (containerProc-seed.cs) and two planted command services (dockerProc.spcs for docker leaves, podmanProc.spcs for rootless-sysd leaves) that manage container image build, compose, run, verify, status, and clean.

  2. Derive all operating parameters from the leaf directory path in which each .spcs is planted — there are no per-leaf parameter files.

  3. Provide standalone helpers: podmanHostVerify.cs (host readiness check for rootless-sysd), dockerProc-sbom.pcs / podman-sbom.pcs (install docker / podman via bisos.sbom), and dockerCmnds.cs / podmanCmnds.cs (cheat sheet of direct-engine invocations).

The core of PyCS-Framework is the bisos.b package (the PyCS-Foundation).

The spcs Pattern: Directory Path as Configuration

/=.spcs= is a general BISOS facility, not a bisos.dockerProc invention. For the architectural picture — three-box layered figure, the WalkExampleSpec typed-data mechanism, the branch-side walker Cmnds (fto_forwardToLeaves, fto_walkRunExternal), and how to build a new consumer — see `bisos.pycs/spcs/README.org </bisos/git/auth/bxRepos/bisos-pip/pycs/spcs/README.org>`__. This section summarises how bisos.dockerProc uses .spcs./

A .spcs file (Spread Planted Command Service) is a thin Python file whose behaviour is entirely context-dependent on the directory in which it is planted. The same file content is spread across many directories; each instance is contextualised by its location. The leaf directory path encodes all parameters — there are no per-leaf configuration files.

Two .spcs variants exist here:

  • dockerProc.spcs — planted in leaves that use docker.

  • podmanProc.spcs — planted in leaves that use rootless podman.

Both delegate to the same engine-agnostic seed, containerProc-seed.cs, which parses the leaf path (via bisos.dockerProc.containerProc_seedInfo.paramsFromPlantPath()) to derive engine, profile, ports, base image, and cgroup support.

Companion Repo: brodockerfiles

The image specifications live in a separate repo,
https://github.com/bxObjects/bro_dockerfiles. Its directory hierarchy encodes the configuration:
debian/
  <majorRelease>/           12 or 13
    confined/vnc/xfce/      unprivileged, entrypoint.sh init
      <imageName>/
    privileged/vnc/xfce/    --privileged (docker), systemd PID 1
      <imageName>/
    rootless-sysd/vnc/xfce/ rootless podman, systemd PID 1
      <imageName>/

Clone it alongside your other repos — in BISOS the canonical location is /bisos/git/bxRepos/bxObjects/:

cd /bisos/git/bxRepos/bxObjects
git clone https://github.com/bxObjects/bro_dockerfiles.git

Path-to-parameter Mapping

Each segment of the leaf path debian/<release>/<profile>/vnc/xfce/<imageName>/ encodes one dimension of the container’s configuration:

Path segment

Parameter derived

debian

distro (anchor segment: paramsFromPlantPath() anchors on this)

<release>

12 or 13 → base OS, base image tag

<profile>

confined / privileged / rootless-sysd → engine, init, privilege

vnc/xfce

desktop type → VNC/noVNC ports, xstartup variant

<imageName>

image name → DockerHub name, container name

From these, paramsFromPlantPath() derives: engine (docker vs podman), base image name, host port assignments (ssh / vnc / novnc), --privileged flag, --isolation=chroot for rootless podman builds, --systemd=always for podman run, and cgroup-variant support.

Three Init/Isolation Profiles

The <profile> path segment selects one of three init/isolation profiles:

Profile

Engine

Init

Privilege

Host cgroup

confined

docker

ent rypoint.sh (no systemd)

unprivileged

v1 or v2

`` privileged``

docker

`` /sbin/init`` (systemd PID 1)

-- privileged

v1 or v2 (v1 needs cgv1 overlay)

roo tless-sysd

podman

`` /sbin/init`` (systemd PID 1)

rootless

v2 only

  • confined: unprivileged, no systemd; services (VNC / noVNC / sshd) launched manually by entrypoint.sh.

  • privileged: full systemd as PID 1; runs --privileged on docker. Requires the docker-compose.cgv1.yml overlay on cgroup-v1 hosts.

  • rootless-sysd: full systemd as PID 1, but unprivileged via rootless podman. Requires cgroup v2 with controller delegation. Motivated by the multi-tenant “N engineers on one shared VM” use case where --privileged is a security dealbreaker.

Rootless-sysd delivers the systemd fidelity of the privileged profile with the security posture of the confined profile. It is the target model for multi-tenant use once the Platform Registrar (see the Full Design at a Glance figure) is implemented.

Installation

The sources for the bisos.dockerProc pip package are maintained at:
The bisos.dockerProc pip package is available at PYPI as

You can install bisos.dockerProc with pip or pipx.

Installation With pip

If you need access to bisos.dockerProc as a python module, install it with pip:

pip install bisos.dockerProc

Installation With pipx

If you only need access to bisos.dockerProc on the command line, install it with pipx:

pipx install bisos.dockerProc

The following commands are made available:

  • containerProc-seed.cs — the engine-agnostic seed both .spcs files delegate to.

  • dockerProc.spcs — Spread Planted CS for docker leaves (confined + privileged).

  • podmanProc.spcs — Spread Planted CS for rootless-sysd leaves.

  • dockerCmnds.cs / podmanCmnds.cs — cheat-sheet of direct-engine invocations (podmanCmnds.cs is a symlink to dockerCmnds.cs; the CS dispatches on argv[0]).

  • podmanHostVerify.cs — host readiness check for rootless-sysd containers.

  • dockerProc-sbom.pcs / podman-sbom.pcs — install docker / podman via bisos.sbom.

Usage

The primary workflow is: clone bro_dockerfiles, cd into a leaf directory, and run dockerProc.spcs or podmanProc.spcs there. The .spcs file is already planted in each leaf as part of the companion repo.

Planting a .spcs File in a Leaf Directory

For most users this is already done — the bro_dockerfiles repo ships with dockerProc.spcs planted in each of the four docker leaves and podmanProc.spcs planted in each of the two rootless-sysd leaves. If you create a new leaf, plant the appropriate .spcs:

cp $(which dockerProc.spcs) .

cp $(which podmanProc.spcs) .

chmod +x dockerProc.spcs   # or podmanProc.spcs

The leaf’s path (debian/<release>/<profile>/vnc/xfce/<imageName>/) supplies all parameters.

The Seed Commands: image, instance, verify+status, combined

From any planted leaf, run the .spcs file with no arguments to see the menu of common invocations — filtered to just the commands relevant to that leaf (docker leaves don’t show podman-only options, and vice versa).

cd /bisos/git/bxRepos/bxObjects/bro_dockerfiles/debian/13/privileged/vnc/xfce/bisos_deb13-sysd
./dockerProc.spcs

The Cmnd surface uses two noun prefixes — image* for commands that operate on the container image, and instance* for commands that operate on the running (or stopped) container instance. Plus a full* combined command.

Image commands (image = the built artefact):

Command

Purpose

containerProc_imageBuild

docker build or podman build. --noCache”true”= for clean build.

containerProc_imageDelete

rmi (image only; does NOT touch instances).

Instance commands (instance = a container built from the image):

Command

Purpose

containerProc_instanceUp

Docker: docker compose up -d. Podman: podman run --systemd=always.

containerProc_instanceDown

Docker: docker compose down. Podman: podman stop (does NOT rm).

containerProc_instanceDelete

Stop + remove instance (image preserved).

` containerProc_instanceRestart`

Stop + start in place (state preserved).

containerProc_instancePs

ps -a filtered to this leaf’s container name.

containerProc_instanceLogs

logs (--follow”true”= to stream).

containerProc_instanceExec

exec -it <container> bash (or --execCmd…= for other command).

Verify + status:

Command

Purpose

containerProc_instanceVerify

Host-side smoke test: port + noVNC HTTP + SSH-based systemd checks.

containerProc_instanceStatus

Engine inspect + SSH systemd status.

Combined:

Command

Purpose

containerProc_fullClean

instanceDelete + imageDelete (from-scratch rebuild).

Additional notes:

  • imageBuild on a rootless-sysd leaf auto-builds the confined base image if it is missing from podman’s store.

  • instanceVerify is exec-free for rootless (uses SSH instead of podman exec, which is unreliable on old Podman for systemd containers).

  • For docker leaves, --cgroupVer”v1”= selects docker-compose.cgv1.yml when the host is on cgroup v1.

  • The old flat names (build, composeUp, composeDown, run, verify, status, clean) are kept as deprecated aliases for one release — they still work but emit a DeprecationWarning.

Example — build and start the deb13 privileged image on a cgroup-v2 host:

cd /bisos/git/bxRepos/bxObjects/bro_dockerfiles/debian/13/privileged/vnc/xfce/bisos_deb13-sysd
./dockerProc.spcs -i containerProc_imageBuild
./dockerProc.spcs -i containerProc_instanceUp
./dockerProc.spcs -i containerProc_instanceVerify

On a cgroup-v1 host (e.g. RHEL 8), pass cgroupVer=v1 to instance up/down:

./dockerProc.spcs -i containerProc_instanceUp   --cgroupVer="v1"
./dockerProc.spcs -i containerProc_instanceDown --cgroupVer="v1"

For rootless-sysd (podman) leaves:

cd /bisos/git/bxRepos/bxObjects/bro_dockerfiles/debian/13/rootless-sysd/vnc/xfce/bisos_deb13-rootless-sysd
./podmanProc.spcs -i containerProc_imageBuild
./podmanProc.spcs -i containerProc_instanceUp --detach="true"
./podmanProc.spcs -i containerProc_instanceVerify

Host Readiness Check: podmanHostVerify.cs

Before running rootless-sysd containers on a host for the first time, run podmanHostVerify.cs -i verify to check the host meets the requirements: non-root user, podman installed, cgroup v2, crun OCI runtime, subuid/subgid ranges, XDGRUNTIMEDIR + user systemd/D-Bus session, linger enabled, cgroup-v2 controller delegation, graphroot on local disk with free space.

podmanHostVerify.cs -i verify

Reports PASS / WARN / FAIL for each check and an overall GO / NO-GO exit status.

Docker / Podman Installation via -sbom.pcs Files

Two .pcs files use bisos.sbom to install docker or podman on a Debian host with all their supporting packages:

dockerProc-sbom.pcs -i sbom_apt_install   # installs docker-ce + friends
podman-sbom.pcs    -i sbom_apt_install    # installs podman + rootless deps

The docker sbom adds Docker’s official apt repository first (mirrors what dockerInstall.sh used to do); the podman sbom uses Debian’s own packages.

Cheat Sheet: dockerCmnds.cs / podmanCmnds.cs

For direct docker / podman command references (not routed through the seed — just a documented list of the underlying engine commands), run:

dockerCmnds.cs
podmanCmnds.cs

Both print an examples menu covering inspect / images / run / compose / exec / cleanup. podmanCmnds.cs is a symlink to dockerCmnds.cs; the CS dispatches on argv[0] to select the docker or podman variant. The podman menu also includes rootless-sysd-specific commands (--systemd=always, --isolation=chroot, cgroup check, linger, host verify).

Key Files

An overview of the relevant files of the bisos.dockerProc package (starting-point template; to be updated to reflect bisos.airflow’s actual files):

  • py3/bin/containerProc-seed.cs — the engine-agnostic seed both .spcs delegate to.

  • py3/bin/dockerProc.spcs — Spread Planted CS for docker leaves.

  • py3/bin/podmanProc.spcs — Spread Planted CS for rootless-sysd leaves.

  • py3/bin/podmanHostVerify.cs — host readiness check for rootless-sysd.

  • py3/bin/dockerProc-sbom.pcs / py3/bin/podman-sbom.pcs — engine install via bisos.sbom.

  • py3/bin/dockerCmnds.cs / py3/bin/podmanCmnds.cs — cheat-sheet CS (symlinked).

  • py3/bisos/dockerProc/containerProc_seedInfo.py — Engine=/=Profile=/=CgroupVer enums, ContainerParams dataclass, paramsFromPlantPath() pure function.

  • py3/bisos/dockerProc/containerProc_seed.py — atexit registration.

  • py3/bisos/dockerProc/containerProc_csu.py — CS command implementations. Cmnds are organized by noun: containerProc_image* (build/delete), containerProc_instance* (up/down/delete/restart/ps/logs/exec/verify/status), and containerProc_fullClean (combined).

  • py3/images/containerProc-graphviz.pcs — source for the “Full Design at a Glance” figure.

  • py3/setup.py, py3/pypiProc.sh — PyPI packaging (setup.py is dblock-driven; do not hand-edit).

Documentation and Blee-Panels

bisos.airflow is part of the ByStar Digital Ecosystem http://www.by-star.net.

This module’s primary documentation is in the form of Blee-Panels. Blee-Panels are in the ./panels directory. From within Blee and BISOS these panels are accessible under the Blee “Panels” menu.

See file:./panels/bisos.airflow/_nodeBase_/fullUsagePanel-en.org for a starting point.

The companion repo bro_dockerfiles has its own README with details on each image variant, host cgroup v1/v2 compatibility, and the per-leaf build / verify recipes.

bisos.dockerProc is best developed with Blee, the By* BISOS Libre-Halaal Emacs Environment — a layer on top of Emacs and BISOS which creates a comprehensive integrated usage and development environment.

Support

For support, criticism, comments and questions; please contact the author/maintainer

Metadata

Release files for bisos.airflow 0.1

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for bisos.airflow 0.1
File Size Uploaded
bisos_airflow-0.1.tar.gz 70.8 kB Details

Release files / bisos_airflow-0.1.tar.gz

Download URL bisos_airflow-0.1.tar.gz
Size 70.8 kB
Tags Source
SHA-256 checksum
How to use checksums
7b759a4aa0cac1dff711a1efe9a2bc47a5ccc807a3775e5e6be24138340a01cf
BLAKE2b-256 checksum
How to use checksums
94261f0c2053af579ac5367d1af162a2dea0df2cbab0a196a5fd139dfccd7366
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/5.1.1 CPython/3.11.2

Release history Release notifications | RSS feed

0.13

1 release file

0.12

1 release file

0.11

1 release file

This release

0.1 This release

1 release file

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page