Object Storage

GitLab supports using an object storage service for holding numerous types of data. It’s recommended over NFS and in general it’s better in larger setups as object storage is typically much more performant, reliable, and scalable.

Options

Object storage options that GitLab has tested, or is aware of customers using include:

Configuration guides

For configuring GitLab to use Object Storage refer to the following guides:

  1. Configure object storage for backups.
  2. Configure object storage for job artifacts including incremental logging.
  3. Configure object storage for LFS objects.
  4. Configure object storage for uploads.
  5. Configure object storage for merge request diffs.
  6. Configure object storage for Container Registry (optional feature).
  7. Configure object storage for Mattermost (optional feature).
  8. Configure object storage for packages (optional feature).
  9. Configure object storage for Dependency Proxy (optional feature).
  10. Configure object storage for Pseudonymizer (optional feature).
  11. Configure object storage for autoscale Runner caching (optional - for improved performance).
  12. Configure object storage for Terraform state files

Other alternatives to filesystem storage

If you’re working to scale out your GitLab implementation, or add fault tolerance and redundancy, you may be looking at removing dependencies on block or network filesystems. See the following guides and note that Pages requires disk storage:

  1. Make sure the git user home directory is on local disk.
  2. Configure database lookup of SSH keys to eliminate the need for a shared authorized_keys file.

Warnings, limitations, and known issues

Use separate buckets

Using separate buckets for each data type is the recommended approach for GitLab.

A limitation of our configuration is that each use of object storage is separately configured. We have an issue for improving this and easily using one bucket with separate folders is one improvement that this might bring.

There is at least one specific issue with using the same bucket: when GitLab is deployed with the Helm chart restore from backup will not properly function unless separate buckets are used.

One risk of using a single bucket would be that if your organisation decided to migrate GitLab to the Helm deployment in the future. GitLab would run, but the situation with backups might not be realised until the organisation had a critical requirement for the backups to work.

S3 API compatibility issues

Not all S3 providers are fully compatible with the Fog library that GitLab uses. Symptoms include: 

411 Length Required

GitLab Pages requires NFS

If you’re working to add more GitLab servers for scaling or fault tolerance and one of your requirements is GitLab Pages this currently requires NFS. There is work in progress to remove this dependency. In the future, GitLab Pages may use object storage.

The dependency on disk storage also prevents Pages being deployed using the GitLab Helm chart.

Incremental logging is required for CI to use object storage

If you configure GitLab to use object storage for CI logs and artifacts, you must also enable incremental logging.

Proxy Download

A number of the use cases for object storage allow client traffic to be redirected to the object storage back end, like when Git clients request large files via LFS or when downloading CI artifacts and logs.

When the files are stored on local block storage or NFS, GitLab has to act as a proxy. This is not the default behavior with object storage.

The proxy_download setting controls this behavior: the default is generally false. Verify this in the documentation for each use case. Set it to true so that GitLab proxies the files.

When not proxying files, GitLab returns an HTTP 302 redirect with a pre-signed, time-limited object storage URL. This can result in some of the following problems:

  • If GitLab is using non-secure HTTP to access the object storage, clients may generate https->http downgrade errors and refuse to process the redirect. The solution to this is for GitLab to use HTTPS. LFS, for example, will generate this error: 

     LFS: lfsapi/client: refusing insecure redirect, https->http

  • Clients will need to trust the certificate authority that issued the object storage certificate, or may return common TLS errors such as: 

     x509: certificate signed by unknown authority

  • Clients will need network access to the object storage. Errors that might result if this access is not in place include: 

     Received status code 403 from server: Forbidden

Getting a 403 Forbidden response is specifically called out on the package repository documentation as a side effect of how some build tools work.

ETag mismatch

Using the default GitLab settings, some object storage back-ends such as MinIO and Alibaba might generate ETag mismatch errors.

If you are seeing this ETag mismatch error with Amazon Web Services S3, it’s likely this is due to encryption settings on your bucket. See the section on using Amazon instance profiles on how to fix this issue.

When using GitLab direct upload, the workaround for MinIO is to use the --compat parameter on the server.

We are working on a fix to the GitLab Workhorse component.

Using Amazon instance profiles

Instead of supplying AWS access and secret keys in object storage configuration, GitLab can be configured to use IAM roles to set up an Amazon instance profile. When this is used, GitLab will fetch temporary credentials each time an S3 bucket is accessed, so no hard-coded values are needed in the configuration.

Encrypted S3 buckets

Introduced in GitLab 13.1 for instance profiles only.

When configured to use an instance profile, GitLab Workhorse will properly upload files to S3 buckets that have SSE-S3 or SSE-KMS encryption enabled by default. Note that customer master keys (CMKs) and SSE-C encryption are not yet supported since this requires supplying keys to the GitLab configuration.

Without instance profiles enabled (or prior to GitLab 13.1), GitLab Workhorse will upload files to S3 using pre-signed URLs that do not have a Content-MD5 HTTP header computed for them. To ensure data is not corrupted, Workhorse checks that the MD5 hash of the data sent equals the ETag header returned from the S3 server. When encryption is enabled, this is not the case, which causes Workhorse to report an ETag mismatch error during an upload.

With instance profiles enabled, GitLab Workhorse uses an AWS S3 client that properly computes and sends the Content-MD5 header to the server, which eliminates the need for comparing ETag headers. If the data is corrupted in transit, the S3 server will reject the file.

Disabling the feature

The Workhorse S3 client is enabled by default when the use_iam_profile configuration option is set to true.

The feature can be disabled using the :use_workhorse_s3_client feature flag. To disable the feature, ask a GitLab administrator with Rails console access to run the following command: 

Feature.disable(:use_workhorse_s3_client)

IAM Permissions

To set up an instance profile:

  1. Create an Amazon Identity Access and Management (IAM) role with the necessary permissions. The following example is a role for an S3 bucket named test-bucket: 

    {
    "Version": "2012-10-17",
    "Statement": [
        {
            "Sid": "VisualEditor0",
            "Effect": "Allow",
            "Action": [
                "s3:PutObject",
                "s3:GetObject",
                "s3:AbortMultipartUpload",
                "s3:DeleteObject"
            ],
            "Resource": "arn:aws:s3:::test-bucket/*"
        }
    ]
}
  2. Attach this role to the EC2 instance hosting your GitLab instance.
  3. Configure GitLab to use it via the use_iam_profile configuration option. For example, when configuring uploads to use object storage, see the AWS IAM profiles section in S3-compatible connection settings.