GCP Per-Bucket Credentials

By default, all GCS operations use a single set of credentials - the service account referenced by the GOOGLE_APPLICATION_CREDENTIALS environment variable (or the ambient credentials on GCE/GKE). This is sufficient when all GCS buckets belong to the same GCP project and service account.

extra.gcp.application_creds lifts that restriction: it lets each bucket in AIS carry its own GCP service-account JSON path, completely overriding the global default for that bucket’s operations.

Table of Contents

• The extra.gcp section
• Setting per-bucket credentials
  • Example: validate per-bucket credentials on first access
• Credentials file format
• Registering a bucket that the default credentials cannot reach
• Multiple buckets, multiple GCP projects
• Startup behavior
• Clearing per-bucket credentials
• See also

The extra.gcp section

To inspect the GCP-specific knobs for a bucket:

1
$ ais bucket props show gs://abc extra.gcp
2
PROPERTY                         VALUE
3
extra.gcp.application_creds

or in JSON to see all supported fields regardless of whether they are set:

1
$ ais bucket props show gs://abc extra.gcp --json
2
{
3
    "extra": {
4
        "gcp": {
5
            "application_creds": "-"
6
        }
7
    }
8
}

Note on -: a dash in --json output means the field is supported but currently unset (no value configured). It is not a valid value to assign.

Setting per-bucket credentials

1
$ ais bucket props set gs://abc extra.gcp.application_creds /mnt/vault/sa-team-b.json
2
"extra.gcp.application_creds" set to: "/mnt/vault/sa-team-b.json" (was: "")

From this point on, every operation on gs://abc - GET, PUT, HEAD, LIST - will authenticate using the service account in /mnt/vault/sa-team-b.json instead of the cluster-wide default.

The path must be absolute and clean (no .. components). AIS validates this when the property is set and returns an error if the path is relative or malformed.

Example: validate per-bucket credentials on first access

Create bucket with non-default creds bypassing lookup:

1
$ ais create gs://abc --skip-lookup --props='extra.gcp.application_creds=/tmp/GOOD_gcp_creds.json'
2
"gs://abc" created
3
4
# first access resolves creds and establishes a session
5
$ ais ls gs://abc
6
7
NAME                     SIZE     CACHED
8
README.md                ...      no
9
...
10
Listed 778503 names (in-cluster: none)

On the other hand, if the credentials file is invalid or missing:

1
$ ais create gs://abc --skip-lookup --props='extra.gcp.application_creds=/tmp/BAD_gcp_creds.json'
2
"gs://abc" created
3
4
$ ais ls gs://abc
5
Error: gcp: failed to load application creds from ... for bucket ...

Credentials file format

The value is a file path to a standard GCP service-account JSON key:

1
{
2
  "type": "service_account",
3
  "project_id": "my-gcp-project",
4
  "private_key_id": "...",
5
  "private_key": "-----BEGIN RSA PRIVATE KEY-----\n...",
6
  "client_email": "sa-team-b@my-gcp-project.iam.gserviceaccount.com",
7
  ...
8
}

AIS reads the project_id from the file at the time the session is first established. The file path - not the credentials themselves - is what is stored in the bucket’s BMD entry, so secrets never leave the node’s filesystem.

The JSON key file itself is never replicated across nodes or transmitted over the network. Each AIS target must have access to the specified path locally.

In Kubernetes, service-account keys are typically mounted as files via a Secret volume:

1
volumes:
2
  - name: gcp-sa
3
    secret:
4
      secretName: team-b-gcp-sa
5
volumeMounts:
6
  - name: gcp-sa
7
    mountPath: /mnt/vault/sa
8
    readOnly: true

which makes /mnt/vault/sa/key.json a normal readable file from AIS’s perspective.

Registering a bucket that the default credentials cannot reach

If the cluster’s default service account has no access to the bucket, use --skip-lookup to bypass the initial HEAD check:

1
$ ais create gs://team-b-data --skip-lookup \
2
  --props="extra.gcp.application_creds=/mnt/vault/sa-team-b.json"
3
"gs://team-b-data" created

Then verify access:

1
$ ais ls gs://team-b-data
2
NAME             SIZE
3
...

--skip-lookup registers the bucket in BMD without verifying reachability. Use it only when you know the bucket exists and the per-bucket credentials are correct.

Multiple buckets, multiple GCP projects

extra.gcp.application_creds is per-bucket, so you can mix projects freely:

1
$ ais create gs://#proj-a/dataset --skip-lookup \
2
  --props="extra.gcp.application_creds=/mnt/vault/sa-proj-a.json"
3
4
$ ais create gs://#proj-b/dataset --skip-lookup \
5
  --props="extra.gcp.application_creds=/mnt/vault/sa-proj-b.json"

gs://#proj-a/dataset and gs://#proj-b/dataset are distinct BMD entries with separate on-disk paths, separate session caches, and independent credentials. The namespace component (e.g., #proj-a, #proj-b) disambiguates same-name buckets across projects.

AIS caches the GCP client session keyed by credentials file path, so the overhead of creating a new *storage.Client is paid once per unique credentials file across the lifetime of the target process.

Startup behavior

AIS initializes the default GCP session lazily - on first use, not at target startup. A missing or invalid GOOGLE_APPLICATION_CREDENTIALS will not prevent the target from starting

An error will occur only when accessing a bucket that requires the default (missing or invalid) credentials.

Per-bucket sessions (extra.gcp.application_creds) are resolved the same way: on demand, then cached.

Clearing per-bucket credentials

To revert a bucket to the cluster-wide default:

1
$ ais bucket props set gs://abc extra.gcp.application_creds ""
2
3
"extra.gcp.application_creds" set to: "" (was: "/mnt/vault/sa-team-b.json")

See also

• AWS Profiles and S3 Endpoints
• Bucket Properties
• Namespaced Buckets