apache-airflow-providers-cncf-kubernetes

Content

Guides

References

Resources

Commits

Package apache-airflow-providers-cncf-kubernetes

Kubernetes

Release: 3.0.2: THIS VERSION IS YANKED - DO NOT INSTALL IT (SEE WARNING BELOW)

Provider package

This is a provider package for cncf.kubernetes provider. All classes for this provider package are in airflow.providers.cncf.kubernetes python package.

Warning

This version of the provider has been yanked as it had compatibility issues with core Airflow. Do not use this version. If you are on Airflow 2.1.* or 2.2.*, please downgrade the provider version to 3.0.0.

Both core Airflow and the cncf.kubernetes provider use the kubernetes library, however, the kubernetes version used for the provider clashes with core Airflow dependencies. In Airflow 2.1.* and 2.2.* the latest working version of the provider is 3.0.0. The other 3.* versions of provider are "yanked" which means that they are skipped when installed unless you explicitly choose to install this version.

Unless we see some serious bug to fix - all the future releases of the cncf.kubernetes provider (starting from 4.0.0) will only be installable for Airflow 2.3.* and above.

Installation

You can install this package on top of an existing Airflow 2.1+ installation via pip install apache-airflow-providers-cncf-kubernetes

PIP requirements

PIP package

Version required

apache-airflow

>=2.1.0

cryptography

>=2.0.0

kubernetes

>=3.0.0

Changelog

3.0.2 (YANKED)

Bug Fixes

  • Add missed deprecations for cncf (#20031)

3.0.1 (YANKED)

Misc

  • Update Kubernetes library version (#18797)

3.0.0

Breaking changes

  • Parameter is_delete_operator_pod default is changed to True (#20575)

  • Simplify KubernetesPodOperator (#19572)

  • Move pod_mutation_hook call from PodManager to KubernetesPodOperator (#20596)

  • Rename ''PodLauncher'' to ''PodManager'' (#20576)

Parameter is_delete_operator_pod has new default

Previously, the default for param is_delete_operator_pod was False, which means that after a task runs, its pod is not deleted by the operator and remains on the cluster indefinitely. With this release, we change the default to True.

Notes on changes KubernetesPodOperator and PodLauncher

Warning

Many methods in KubernetesPodOperator and PodLauncher have been renamed. If you have subclassed KubernetesPodOperator you will need to update your subclass to reflect the new structure. Additionally PodStatus enum has been renamed to PodPhase.

Overview

Generally speaking if you did not subclass KubernetesPodOperator and you didn't use the PodLauncher class directly, then you don't need to worry about this change. If however you have subclassed KubernetesPodOperator, what follows are some notes on the changes in this release.

One of the principal goals of the refactor is to clearly separate the "get or create pod" and "wait for pod completion" phases. Previously the "wait for pod completion" logic would be invoked differently depending on whether the operator were to "attach to an existing pod" (e.g. after a worker failure) or "create a new pod" and this resulted in some code duplication and a bit more nesting of logic. With this refactor we encapsulate the "get or create" step into method KubernetesPodOperator.get_or_create_pod, and pull the monitoring and XCom logic up into the top level of execute because it can be the same for "attached" pods and "new" pods.

The KubernetesPodOperator.get_or_create_pod tries first to find an existing pod using labels specific to the task instance (see KubernetesPodOperator.find_pod). If one does not exist it creates a pod <~.PodManager.create_pod>.

The "waiting" part of execution has three components. The first step is to wait for the pod to leave the Pending phase (~.KubernetesPodOperator.await_pod_start). Next, if configured to do so, the operator will follow the base container logs and forward these logs to the task logger until the base container is done. If not configured to harvest the logs, the operator will instead KubernetesPodOperator.await_container_completion either way, we must await container completion before harvesting xcom. After (optionally) extracting the xcom value from the base container, we await pod completion <~.PodManager.await_pod_completion>.

Previously, depending on whether the pod was "reattached to" (e.g. after a worker failure) or created anew, the waiting logic may have occurred in either handle_pod_overlap or create_new_pod_for_operator.

After the pod terminates, we execute different cleanup tasks depending on whether the pod terminated successfully.

If the pod terminates unsuccessfully, we attempt to log the pod events PodLauncher.read_pod_events>. If additionally the task is configured not to delete the pod after termination, we apply a label KubernetesPodOperator.patch_already_checked> indicating that the pod failed and should not be "reattached to" in a retry. If the task is configured to delete its pod, we delete it KubernetesPodOperator.process_pod_deletion>. Finally, we raise an AirflowException to fail the task instance.

If the pod terminates successfully, we delete the pod KubernetesPodOperator.process_pod_deletion> (if configured to delete the pod) and push XCom (if configured to push XCom).

Details on method renames, refactors, and deletions

In KubernetesPodOperator:

  • Method create_pod_launcher is converted to cached property pod_manager

  • Construction of k8s CoreV1Api client is now encapsulated within cached property client

  • Logic to search for an existing pod (e.g. after an airflow worker failure) is moved out of execute and into method find_pod.

  • Method handle_pod_overlap is removed. Previously it monitored a "found" pod until completion. With this change the pod monitoring (and log following) is orchestrated directly from execute and it is the same whether it's a "found" pod or a "new" pod. See methods await_pod_start, follow_container_logs, await_container_completion and await_pod_completion.

  • Method create_pod_request_obj is renamed build_pod_request_obj. It now takes argument context in order to add TI-specific pod labels; previously they were added after return.

  • Method create_labels_for_pod is renamed _get_ti_pod_labels. This method doesn't return all labels, but only those specific to the TI. We also add parameter include_try_number to control the inclusion of this label instead of possibly filtering it out later.

  • Method _get_pod_identifying_label_string is renamed _build_find_pod_label_selector

  • Method _try_numbers_match is removed.

  • Method create_new_pod_for_operator is removed. Previously it would mutate the labels on self.pod, launch the pod, monitor the pod to completion etc. Now this logic is in part handled by get_or_create_pod, where a new pod will be created if necessary. The monitoring etc is now orchestrated directly from execute. Again, see the calls to methods await_pod_start, follow_container_logs, await_container_completion and await_pod_completion.

In class PodManager (formerly PodLauncher):

  • Method start_pod is removed and split into two methods: create_pod and await_pod_start.

  • Method monitor_pod is removed and split into methods follow_container_logs, await_container_completion, await_pod_completion

  • Methods pod_not_started, pod_is_running, process_status, and _task_status are removed. These were needed due to the way in which pod phase was mapped to task instance states; but we no longer do such a mapping and instead deal with pod phases directly and untransformed.

  • Method _extract_xcom is renamed extract_xcom.

  • Method read_pod_logs now takes kwarg container_name

Other changes in pod_manager.py (formerly pod_launcher.py):

  • Class pod_launcher.PodLauncher renamed to pod_manager.PodManager

  • Enum-like class PodStatus is renamed PodPhase, and the values are no longer lower-cased.

  • The airflow.settings.pod_mutation_hook is no longer called in cncf.kubernetes.utils.pod_manager.PodManager.run_pod_async. For KubernetesPodOperator, mutation now occurs in build_pod_request_obj.

  • Parameter is_delete_operator_pod default is changed to True so that pods are deleted after task completion and not left to accumulate. In practice it seems more common to disable pod deletion only on a temporary basis for debugging purposes and therefore pod deletion is the more sensible default.

Features

  • Add params config, in_cluster, and cluster_context to KubernetesHook (#19695)

  • Implement dry_run for KubernetesPodOperator (#20573)

  • Clarify docstring for ''build_pod_request_obj'' in K8s providers (#20574)

Bug Fixes

  • Fix Volume/VolumeMount KPO DeprecationWarning (#19726)

2.2.0

Features

  • Added namespace as a template field in the KPO. (#19718)

  • Decouple name randomization from name kwarg (#19398)

Bug Fixes

  • Checking event.status.container_statuses before filtering (#19713)

  • Coalesce 'extra' params to None in KubernetesHook (#19694)

  • Change to correct type in KubernetesPodOperator (#19459)

2.1.0

Features

  • Add more type hints to PodLauncher (#18928)

  • Add more information to PodLauncher timeout error (#17953)

2.0.3

Bug Fixes

  • Fix KubernetesPodOperator reattach when not deleting pods (#18070)

  • Make Kubernetes job description fit on one log line (#18377)

  • Do not fail KubernetesPodOperator tasks if log reading fails (#17649)

2.0.2

Bug Fixes

  • Fix using XCom with ''KubernetesPodOperator'' (#17760)

  • Import Hooks lazily individually in providers manager (#17682)

2.0.1

Features

  • Enable using custom pod launcher in Kubernetes Pod Operator (#16945)

Bug Fixes

  • BugFix: Using 'json' string in template_field causes issue with K8s Operators (#16930)

2.0.0

Breaking changes

  • Auto-apply apply_default decorator (#15667)

Warning

Due to apply_default decorator removal, this version of the provider requires Airflow 2.1.0+. If your Airflow version is < 2.1.0, and you want to install this provider version, first upgrade Airflow to at least version 2.1.0. Otherwise your Airflow package version will be upgraded automatically and you will have to manually run airflow upgrade db to complete the migration.

Features

  • Add 'KubernetesPodOperat' 'pod-template-file' jinja template support (#15942)

  • Save pod name to xcom for KubernetesPodOperator (#15755)

Bug Fixes

  • Bug Fix Pod-Template Affinity Ignored due to empty Affinity K8S Object (#15787)

  • Bug Pod Template File Values Ignored (#16095)

  • Fix issue with parsing error logs in the KPO (#15638)

  • Fix unsuccessful KubernetesPod final_state call when 'is_delete_operator_pod=True' (#15490)

1.2.0

Features

  • Require 'name' with KubernetesPodOperator (#15373)

  • Change KPO node_selectors warning to proper deprecationwarning (#15507)

Bug Fixes

  • Fix timeout when using XCom with KubernetesPodOperator (#15388)

  • Fix labels on the pod created by ''KubernetsPodOperator'' (#15492)

1.1.0

Features

  • Separate Kubernetes pod_launcher from core airflow (#15165)

  • Add ability to specify api group and version for Spark operators (#14898)

  • Use libyaml C library when available. (#14577)

1.0.2

Bug fixes

  • Allow pod name override in KubernetesPodOperator if pod_template is used. (#14186)

  • Allow users of the KPO to *actually* template environment variables (#14083)

1.0.1

Updated documentation and readme files.

Bug fixes

  • Pass image_pull_policy in KubernetesPodOperator correctly (#13289)

1.0.0

Initial version of the provider.

Was this entry helpful?