Troubleshooting OpenBao

For recovery key tasks and break-glass root tokens, see recovery key management. For Geo failover, see Geo disaster recovery.

Where OpenBao runs

OpenBao always runs in Kubernetes, even when GitLab uses the Linux package. The namespace and deployment name depend on the installation method:

These examples use the Cloud Native namespace gitlab. For a Linux package installation, replace gitlab with openbao in the kubectl commands.

OpenBao pods carry the label app.kubernetes.io/name=openbao. The active node also carries openbao-active=true.

Find OpenBao logs

Read the OpenBao logs with kubectl logs. The related GitLab Rails and Sidekiq logs are stored separately, depending on the installation method:

OpenBao posts audit events to GitLab and also writes them to the OpenBao pod logs.

Find the OpenBao pods

To list the OpenBao pods and see which node is active:

kubectl get pods -n gitlab -l app.kubernetes.io/name=openbao \
  --label-columns openbao-active,openbao-sealed

The pod with OPENBAO-ACTIVE set to true is the active node. The others are standby nodes.

Check OpenBao status

OpenBao must be unsealed to serve requests. To check, run bao status in a pod:

OPENBAO_POD=$(kubectl get pods -n gitlab -l app.kubernetes.io/name=openbao -o name | head -1)
kubectl exec -n gitlab "$OPENBAO_POD" -c openbao-server -- \
  sh -c "BAO_ADDR=http://127.0.0.1:8200 bao status"

In the output, Sealed must be false. The active node shows HA Mode active and a standby node shows HA Mode standby:

Seal Type       static
Initialized     true
Sealed          false
Storage Type    postgresql
HA Enabled      true
HA Mode         active

The sys/seal-status endpoint reports the same state as "sealed":false:

kubectl exec -n gitlab "$OPENBAO_POD" -c openbao-server -- \
  sh -c "BAO_ADDR=http://127.0.0.1:8200 bao read sys/seal-status"

In the logs, a node that unsealed successfully logs vault is unsealed. The active node logs acquired lock, enabling active operation and a standby node logs entering standby mode:

OPENBAO_POD=$(kubectl get pods -n gitlab -l app.kubernetes.io/name=openbao -o name | head -1)
kubectl logs -n gitlab "$OPENBAO_POD" -c openbao-server \
  | grep -E "acquired lock, enabling active operation|entering standby mode"

Find errors in a time window

To read OpenBao logs from a time window, use --since:

OPENBAO_POD=$(kubectl get pods -n gitlab -l app.kubernetes.io/name=openbao -o name | head -1)
kubectl logs -n gitlab "$OPENBAO_POD" -c openbao-server --since=30m \
  | grep -iE "error|warn|failed"

For a Linux package installation, search the Rails and Sidekiq log files by time. The logs are JSON, one event per line.

GitLab Rails logs

The Rails logs cover secret operations from the UI and GraphQL API, plus the audit callback from OpenBao.

For a Cloud Native installation:

kubectl logs -n gitlab -l app=webservice -c webservice \
  | grep -E "Projects::SecretsController|Groups::SecretsController|secrets_manager/audit_logs"

For a Linux package installation:

grep -E "Projects::SecretsController|Groups::SecretsController|secrets_manager/audit_logs" \
  /var/log/gitlab/gitlab-rails/production_json.log

GraphQL operations appear with a caller_id such as graphql:createProjectSecret or graphql:getGroupSecrets. The audit callback appears as the path /api/v4/internal/secrets_manager/audit_logs.

Sidekiq logs

The workers that provision, deprovision, and maintain Secrets Manager records run under the SecretsManagement:: namespace.

For a Cloud Native installation:

kubectl logs -n gitlab -l app=sidekiq -c sidekiq | grep "SecretsManagement::"

For a Linux package installation:

grep "SecretsManagement::" /var/log/gitlab/sidekiq/current

For provisioning problems, filter for ProvisionProjectSecretsManagerWorker or ProvisionGroupSecretsManagerWorker.

GitLab Runner logs

When a CI/CD job fails to fetch a secret, the cause appears in the job log in the GitLab UI. Search the job log for these strings:

Healthy startup logs

After a restart, the active node logs this sequence. A standby node stops at vault is unsealed and then logs entering standby mode. The line format varies by configuration, so match the message text rather than a prefix.

Error messages

OpenBao messages come from the openbao-server container. GitLab messages come from the Rails or Sidekiq logs.

Secrets Manager is stuck in provisioning

When you enable the Secrets Manager, the toggle can stay in a loading state with the status at provisioning. The Secrets Manager has no failed state, so any step that fails before activation leaves the record stuck. The usual cause is that Sidekiq cannot reach OpenBao.

To diagnose:

1. Check the Sidekiq logs for the provisioning worker:

   kubectl logs -n gitlab -l app=sidekiq -c sidekiq \
     | grep -E "ProvisionProjectSecretsManagerWorker|ProvisionGroupSecretsManagerWorker"

2. Test whether Sidekiq can reach OpenBao, from a Sidekiq pod or node:

   curl "https://openbao.example.com/v1/sys/health"

A maintenance worker retries a stale task up to three times, then stops. After that, the record stays in provisioning with no automated recovery, and the retries log Retrying failed secrets_manager maintenance task.

After you fix the connectivity, disable and re-enable the Secrets Manager to provision it again.

Authentication mount missing after self-initialization

On a fresh installation with multiple OpenBao pods, a self-initialization race can leave OpenBao unsealed but without the gitlab_rails_jwt/ authentication mount. The pods look healthy, but secret operations fail with permission denied. Run bao auth list with a root token to confirm the mount exists. To prevent the race, start a fresh installation with a single replica, confirm initialization completes, then scale up.

GitLab cannot connect to OpenBao

GitLab Rails and Sidekiq connect to OpenBao over HTTP. Rails uses internal_url, and falls back to url when internal_url is not set. To inspect the configuration, run this in the Rails console:

Gitlab.config.openbao.to_h

Common causes:

• An https:// URL against an OpenBao listener that serves http fails with wrong version number. global.openbao.https sets the scheme GitLab connects with, not the OpenBao listener TLS. The listener serves plain HTTP by default. Either leave global.openbao.https unset to match it, or enable listener TLS with openbao.config.tlsDisable: false and set global.openbao.https to true.
• OIDC discovery and audit logging fail over an untrusted TLS certificate. Use a certificate that GitLab trusts.
• A request that produces no OpenBao audit entry never reached the authentication backend. Check the Ingress or reverse proxy.

For a Cloud Native installation, a working configuration looks like this:

global:
  openbao:
    enabled: true
    url: http://gitlab-openbao-active:8200
    internal_url: http://gitlab-openbao-active:8200

For a Linux package installation, GitLab uses the gitlab_rails['openbao']['url'] setting in /etc/gitlab/gitlab.rb to connect to OpenBao. The bundled NGINX reverse proxy routes to OpenBao with the oak['components']['openbao'] settings. For more information, see Install OpenBao for a Linux package deployment.

JWT authentication fails

GitLab authenticates to OpenBao with a JWT. The aud (audience) claim in the JWT must exactly match the bound_audiences value on the OpenBao authentication role. Any difference fails authentication, including a trailing slash, http compared to https, or a port.

OpenBao stores bound_audiences at initialization time, derived from the OpenBao URL. The stored value does not change when you later change the URL. Changing the URL therefore breaks authentication, because the stored bound_audiences no longer matches the aud GitLab sends. To set the audience independently of the connect URL, use global.openbao.jwt_audience.

To find the audience GitLab sends, run this in the Rails console:

SecretsManagement::ProjectSecretsManager.jwt_audience

The method returns the configured jwt_audience, or the OpenBao url when jwt_audience is not set. To inspect the stored value, read the authentication role with a root token and compare bound_audiences to that audience.

To fix the mismatch without deleting stored secrets, reconfigure authentication with a recovery key. For the procedure, see Reconfigure authentication with a recovery key.

If you do not have a recovery key, reset OpenBao data. This deletes all stored secrets.

OpenBao pods are sealed

If bao status reports Sealed true at startup, auto-unseal failed:

• With the default static unseal, the cause is usually a missing or incorrect unseal secret. The secret is gitlab-openbao-unseal for a Cloud Native installation, and openbao-static-unseal for a Linux package installation.
• With KMS auto-unseal, currently AWS KMS (awskms), the cause is usually that OpenBao cannot reach the KMS.

To check the seal status, see Check OpenBao status.

Database problems

OpenBao requires its own PostgreSQL database. The GitLab chart fails the installation or upgrade if you enable OpenBao without a dedicated database.

Other database problems:

• Connection pool exhaustion or high latency causes intermittent timeouts.
• Incorrect md5_auth_cidr_addresses, sslMode, or password values in a Linux package PostgreSQL configuration send the OpenBao pods into CrashLoopBackOff. For the correct settings, see Install OpenBao for a Linux package deployment.

Audit events are missing

OpenBao posts audit events to GitLab at /api/v4/internal/secrets_manager/audit_logs. The GitLab chart enables audit logging by default. If audit events do not arrive:

• Setting config.audit.http.enabled to false stops OpenBao from posting events. Confirm that audit logging is enabled.
• A shared audit token mismatch returns 401 on the audit endpoint. Confirm that GitLab and OpenBao use the same audit token.

Troubleshoot Geo deployments

OpenBao runs as an active node on the primary Geo site and as a standby node on each secondary site. A secondary node connects to a read-only PostgreSQL replica, so it logs failed to acquire lock and cannot execute INSERT in a read-only transaction. These messages are expected.

If a secondary node logs cipher: message authentication failed or unknown key ID, its seal key does not match the primary. The fix depends on the seal mechanism:

• With a static unseal, copy the gitlab-openbao-unseal secret from the primary cluster to the secondary cluster, then restart the OpenBao pods:

  kubectl -n gitlab get secret gitlab-openbao-unseal -o yaml

• With a KMS seal, configure both sites to use the same KMS key.

If JWT authentication fails after a failover, the audience no longer matches the stored bound_audiences. The fix depends on the domain:

• If both sites use the primary OpenBao URL, set jwt_audience to the primary OpenBao URL on both sites. See Install OpenBao on a secondary site.
• If the secondary site uses a different domain, this configuration is not supported. Reconfiguring the audience does not restore authentication, because every project and group namespace also needs re-provisioning. Update DNS so the primary domain points to the promoted secondary. For more information, see Geo deployment.

Diagnose slow secret operations

When CI/CD jobs are slow to fetch secrets, or secret operations time out, use the following queries to find the cause. Run these queries in the Prometheus or Grafana instance that scrapes the OpenBao metrics. To expose those metrics, see OpenBao metrics.

Confirm latency is elevated

Use the following query to measure average request latency in milliseconds. The query works at any traffic level, including low-traffic deployments:

rate(openbao_core_handle_request_sum[5m])
/
rate(openbao_core_handle_request_count[5m])

Under normal load, average latency across all request types is typically 3 to 7 ms. Investigate if average latency consistently exceeds 20 ms.

When OpenBao is actively processing requests, use the following query for P99 latency:

openbao_core_handle_request{quantile="0.99"}

Normal P99 is below 10 ms. This query returns NaN when OpenBao is idle because the summary window has no recent observations. Use the rate-based query in that case.

Identify potential issues

If PostgreSQL latency is elevated, check whether the connection pool is saturated. If all connections are busy, additional requests queue and cause latency. For connection pool configuration, see Database resources.