debian-mirror-gitlab/doc/ci/environments/deployment_safety.md
2020-06-23 00:09:42 +05:30

4.9 KiB

Deployment safety

Deployment jobs can be more sensitive than other jobs in a pipeline, and might need to be treated with extra care. GitLab has several features that help maintain deployment security and stability.

You can:

If you are using a continuous deployment workflow and want to ensure that concurrent deployments to the same environment do not happen, you should enable the following options:

Restrict write access to a critical environment

By default, environments can be modified by any team member that has Developer permission or higher. If you want to restrict write access to a critical environment (for example a production environment), you can set up protected environments.

Ensure only one deployment job runs at a time

Pipeline jobs in GitLab CI/CD run in parallel, so it's possible that two deployment jobs in two different pipelines attempt to deploy to the same environment at the same time. This is not desired behavior as deployments should happen sequentially.

You can ensure only one deployment job runs at a time with the resource_group keyword keyword in your .gitlab-ci.yml.

For example:

deploy:
  script: deploy-to-prod
  resource_group: prod

Example of a problematic pipeline flow before using the resource group:

  1. deploy job in Pipeline-A starts running.
  2. deploy job in Pipeline-B starts running. This is a concurrent deployment that could cause an unexpected result.
  3. deploy job in Pipeline-A finished.
  4. deploy job in Pipeline-B finished.

The improved pipeline flow after using the resource group:

  1. deploy job in Pipeline-A starts running.
  2. deploy job in Pipeline-B attempts to start, but waits for the first deploy job to finish.
  3. deploy job in Pipeline-A finishes.
  4. deploy job in Pipeline-B starts running.

For more information, see resource_group keyword in .gitlab-ci.yml.

Skip outdated deployment jobs

The execution order of pipeline jobs can vary from run to run, which could cause undesired behavior. For example, a deployment job in a newer pipeline could finish before a deployment job in an older pipeline. This creates a race condition where the older deployment finished later, overwriting the "newer" deployment.

You can ensure that older deployment jobs are cancelled automatically when a newer deployment runs by enabling the Skip outdated deployment jobs feature.

Example of a problematic pipeline flow before enabling Skip outdated deployment jobs:

  1. Pipeline-A is created on the master branch.
  2. Later, Pipeline-B is created on the master branch (with a newer commit SHA).
  3. The deploy job in Pipeline-B finishes first, and deploys the newer code.
  4. The deploy job in Pipeline-A finished later, and deploys the older code, overwriting the newer (latest) deployment.

The improved pipeline flow after enabling Skip outdated deployment jobs:

  1. Pipeline-A is created on the master branch.
  2. Later, Pipeline-B is created on the master branch (with a newer SHA).
  3. The deploy job in Pipeline-B finishes first, and deploys the newer code.
  4. The deploy job in Pipeline-A is automatically cancelled, so that it doesn't overwrite the deployment from the newer pipeline.

Restrict deployments for a particular period

If you want to prevent deployments for a particular period, for example during a planned vacation period when most employees are out, you can set up a Deploy Freeze. During a deploy freeze period, no deployment can be executed. This is helpful to ensure that deployments do not happen unexpectedly.

Troubleshooting

Pipelines jobs fail with The deployment job is older than the previously succeeded deployment job...

This is caused by the Skip outdated deployment jobs feature. If you have multiple jobs for the same environment (including non-deployment jobs), you might encounter this problem, for example:

build:service-a:
  environment:
    name: production

build:service-b:
  environment:
    name: production

The Skip outdated deployment jobs might not work well with this configuration, and will need to be disabled.

There is a plan to introduce a new annotation for environments to address this issue.