Skip to main content
This page covers common CoreWeave Database issues. For each, it lists the symptom, the likely cause, and how to verify or fix it.

Database never becomes Ready

Symptom

The following wait command times out, and kubectl describe shows the database stuck in Provisioning or Pending:

What to check

  1. Inspect the database status:
    Look at the Status.Conditions section. The condition messages typically point at the underlying problem (scheduling, storage, image pull, certificate generation).
  2. Check the pods:
    A pod in Pending state usually means the scheduler can’t place it. kubectl describe pod on that pod surfaces the reason, most often “no CPU nodes available” or “Insufficient memory”.
  3. Check the PersistentVolumeClaims (PVCs):
    A PVC stuck in Pending means Distributed File Storage (DFS) couldn’t provision the volume.

Common causes

  • The CKS cluster has no CPU nodes with sufficient capacity. Provision CPU capacity through your usual CKS workflow.
  • spec.postgres.synchronousReplication is set and the CKS cluster doesn’t have a separate Node for each instance. Synchronous replication requires each instance to run on a different Node, so the extra instances stay Pending until you add Nodes.
  • CoreWeave Database isn’t enabled on this CKS cluster. Confirm kubectl get crd | grep coreweave.com shows the cwdbclusters and cwdbbackups CRDs. If not, request enablement through your customer Slack channel.
  • The requested storage size exceeds available DFS capacity in the zone.

Resource is created in the wrong namespace

Symptom

You apply a CWDBCluster manifest and the database appears in the default namespace, not the one you expected. Applications in your target namespace can’t resolve the pooler Service.

Cause

The manifest didn’t specify a namespace and you didn’t pass --namespace on the command line. Kubernetes defaults to the default namespace in that case.

Fix

Delete the resource from default and recreate it in the correct namespace:
To prevent this, either set metadata.namespace in the manifest or always pass --namespace [NAMESPACE] on kubectl apply.

Database is rejected for insufficient memory

Symptom

kubectl apply fails with an error similar to the following:
If the manifest sets spec.postgres.connections without spec.resources.requests.memory, the error message is spec.resources.requests.memory is required when spec.postgres.connections is set.

Cause

When spec.postgres.connections is set, the API server requires 10 MiB of spec.resources.requests.memory for each PostgreSQL server connection. The number of server connections is 350 in transaction mode, and spec.postgres.connections.max plus 50 in session and none modes.

Fix

Raise spec.resources.requests.memory to at least 3500Mi in transaction mode, or to (max + 50) × 10 MiB in session or none mode. In session or none mode, you can instead lower max, or switch to transaction mode if your application doesn’t rely on session state. For memory calculations and pooler behavior, see Connection limits and memory requirements and Choose a pooler mode.

Application can’t connect to the database

Symptom

Your application reports a connection error such as connection refused, name resolution failed, or password authentication failed.

What to check

  1. Confirm the database is Ready:
  2. Verify the Service exists:
  3. Check that your application connects to the right Service. Application traffic should go to [CLUSTER-NAME]-pooler-rw or [CLUSTER-NAME]-pooler-ro, or to [CLUSTER-NAME]-rw or [CLUSTER-NAME]-ro in pooler mode none. The uri field in the [CLUSTER-NAME]-[DBNAME]-credentials Secret already points at the right Service. [DBNAME] is the initial database name set in spec.postgres.dbName, which defaults to app. For example, a CWDBCluster named vector-db with dbName: vectors has the Secret vector-db-vectors-credentials. If dbName isn’t a valid RFC 1035 label (at most 63 lowercase letters, digits, and hyphens, starting with a letter and ending with a letter or digit), the operator uses app in the Secret name instead, so dbName: orders_db produces [CLUSTER-NAME]-app-credentials.
  4. Verify cross-namespace name resolution. From a pod in a different namespace, use the fully qualified name: [CLUSTER-NAME]-pooler-rw.[DB-NAMESPACE].svc.cluster.local.
  5. Re-decode the credentials Secret. If the password has rotated, your application may be using a stale value:
See Connect to a database for the full list of connection options.

psql works from a debug pod but fails from the application

Symptom

You can connect from a debug pod using the URI from the Secret, but the application pod cannot.

Common causes

  • The application is reading the wrong Secret. Confirm it mounts [CLUSTER-NAME]-[DBNAME]-credentials for read-write traffic ([CLUSTER-NAME]-app-credentials for the default dbName: app) or [CLUSTER-NAME]-[DBNAME]-credentials-ro for read-only traffic to replicas.
  • The application has a NetworkPolicy that blocks egress to the database namespace. Check NetworkPolicies in the application’s namespace.
  • The application is configured with a hard-coded hostname or port that doesn’t match the pooler Service.

On-demand backup stays in Pending

Symptom

The backup stays in Pending for longer than a few minutes when you check its status. Replace [BACKUP-NAME] with the name of the CWDBBackup resource:

What to check

  1. Confirm the source database is Ready: an on-demand backup can’t start if the database is unhealthy.
  2. Check the backup status detail:
    Look for events such as BackupFailed or quota errors.
  3. Check whether you’re over the on-demand backup quota for the database. Delete unneeded backups, or contact your CoreWeave account team to raise the quota.

Restore takes too long

Symptom

A new CWDBCluster created with spec.postgres.restore set stays in Provisioning for much longer than expected.

Notes

  • Restore time scales with database size and the write-ahead log (WAL) replay distance from the base backup. A small database restored from a recent base backup completes quickly. Restoring a multi-TiB database from older base backups can take hours.
  • The new database only becomes Ready after the bootstrap completes. Don’t cut application traffic over until the new CWDBCluster reports Ready.
  • If a restore appears stuck (no progress for more than 30 minutes for a small database), escalate through your customer Slack channel with the new database name, the spec.postgres.restore.sourceClusterName, and namespace.

Pods aren’t spread across distinct nodes

Symptom

The pod listing shows two or more instances on the same node:

Common causes

  • You added a custom spec.affinity block that turned off the default anti-affinity rules.
  • The CKS cluster doesn’t have enough CPU nodes for all instances. Unless spec.postgres.synchronousReplication is set, the scheduler falls back to co-locating pods when no spread is possible.

Fix

Either remove the custom affinity override and let the operator manage scheduling, or add CPU node capacity so the scheduler can place each instance on a distinct node. The default rules use “preferred” anti-affinity (not “required”) to avoid blocking provisioning, but on a CKS cluster with enough capacity you should see one instance per node. When spec.postgres.synchronousReplication is set, anti-affinity is required, so instances stay Pending instead of sharing a Node.

Escalate an issue

If the steps on this page don’t resolve your issue, contact CoreWeave through your customer Slack channel with:
  • The database name and namespace.
  • The full output of kubectl describe cwdbcluster [CLUSTER-NAME] --namespace [NAMESPACE].
  • Relevant pod or backup status: kubectl get pods,cwdbbackup --namespace [NAMESPACE].
  • A short description of what you expected to happen and what happened instead.
Last modified on November 13, 2026