Troubleshooting =============== .. raw:: html In this page, you can find some basic information on how to troubleshoot CloudNativePG in your Kubernetes cluster deployment. .. tip:: As a Kubernetes administrator, you should have the `kubectl `_ page bookmarked! 　 Before you start ---------------- Kubernetes environment ^^^^^^^^^^^^^^^^^^^^^^ What can make a difference in a troubleshooting activity is to provide clear information about the underlying Kubernetes system. Make sure you know: - the Kubernetes distribution and version you are using - the specifications of the nodes where PostgreSQL is running - as much as you can about the actual :ref:`Other S3-compatible Object Storages providers ` , including storage class and benchmarks you have done before going into production. - which relevant Kubernetes applications you are using in your cluster (i.e. Prometheus, Grafana, Istio, Certmanager, …) - the situation of continuous backup, in particular if it’s in place and working correctly: in case it is not, make sure you take an :ref:`Emergency backup ` before performing any potential disrupting operation Useful utilities ^^^^^^^^^^^^^^^^ On top of the mandatory ``kubectl`` utility, for troubleshooting, we recommend the following plugins/utilities to be available in your system: - :ref:`postgresql.cnpg.io/v1 ` for ``kubectl`` - `jq `_ , a lightweight and flexible command-line JSON processor - `grep `_ , searches one or more input files for lines containing a match to a specified pattern. It is already available in most \*nix distros. If you are on Windows OS, you can use `findstr `_ as an alternative to ``grep`` or directly use `wsl `_ and install your preferred \*nix distro and use the tools mentioned above. First steps ----------- To quickly get an overview of the cluster or installation, the ``kubectl`` plugin is the primary tool to use: 1. the :ref:`status subcommand ` provides an overview of a cluster 2. the :ref:`report subcommand ` provides the manifests for clusters and the operator deployment. It can also include logs using the ``--logs`` option. The report generated via the plugin will include the full cluster manifest. The plugin can be installed on air-gapped systems via packages. Please refer to the :ref:`plugin document ` for complete instructions. Are there backups? ------------------ After getting the cluster manifest with the plugin, you should verify if backups are set up and working. Before proceeding with troubleshooting operations, it may be advisable to perform an emergency backup depending on your findings regarding backups. Refer to the following section for instructions. It is **extremely risky** to operate a production database without keeping regular backups. Emergency backup ---------------- In some emergency situations, you might need to take an emergency logical backup of the main ``app`` database. .. Note:: The instructions you find below must be executed only in emergency situations and the temporary backup files kept under the data protection policies that are effective in your organization. The dump file is indeed stored in the client machine that runs the `kubectl` command, so make sure that all protections are in place and you have enough space to store the backup file. 　 The following example shows how to take a logical backup of the ``app`` database in the ``cluster-example`` Postgres cluster, from the ``cluster-example-1`` pod: .. code:: sh kubectl exec cluster-example-1 -c postgres \ -- pg_dump -Fc -d app > app.dump .. Note:: You can easily adapt the above command to backup your cluster, by providing the names of the objects you have used in your environment. 　 The above command issues a ``pg_dump`` command in custom format, which is the most versatile way to take `logical backups in PostgreSQL `_ . The next step is to restore the database. We assume that you are operating on a new PostgreSQL cluster that’s been just initialized (so the ``app`` database is empty). The following example shows how to restore the above logical backup in the ``app`` database of the ``new-cluster-example`` Postgres cluster, by connecting to the primary (``new-cluster-example-1`` pod): .. code:: sh kubectl exec -i new-cluster-example-1 -c postgres \ -- pg_restore --no-owner --role=app -d app --verbose < app.dump .. Note:: The example in this section assumes that you have no other global objects (databases and roles) to dump and restore, as per our recommendation. In case you have multiple roles, make sure you have taken a backup using `pg_dumpall -g` and you manually restore them in the new cluster. In case you have multiple databases, you need to repeat the above operation one database at a time, making sure you assign the right ownership. If you are not familiar with PostgreSQL, we advise that you do these critical operations under the guidance of a professional support company. 　 The above steps might be integrated into the ``cnpg`` plugin at some stage in the future. Logs ---- All resources created and managed by CloudNativePG log to standard output in accordance with Kubernetes conventions, using :ref:`Standard output logging of PostgreSQL error messages in JSON format ` . While logs are typically processed at the infrastructure level and include those from CloudNativePG, accessing logs directly from the command line interface is critical during troubleshooting. You have three primary options for doing so: - Use the ``kubectl logs`` command to retrieve logs from a specific resource, and apply ``jq`` for better readability. - Use the :ref:`kubectl cnpg logs ` for CloudNativePG-specific logging. - Leverage specialized open-source tools like `stern `_ , which can aggregate logs from multiple resources (e.g., all pods in a PostgreSQL cluster by selecting the ``cnpg.io/clusterName`` label), filter log entries, customize output formats, and more. .. Note:: The following sections provide examples of how to retrieve logs for various resources when troubleshooting CloudNativePG. 　 Operator information -------------------- By default, the CloudNativePG operator is installed in the ``cnpg-system`` namespace in Kubernetes as a ``Deployment`` (see the :ref:`Details about the deployment
` for details). You can get a list of the operator pods by running: .. code:: shell kubectl get pods -n cnpg-system .. Note:: Under normal circumstances, you should have one pod where the operator is running, identified by a name starting with `cnpg-controller-manager-` . In case you have set up your operator for high availability, you should have more entries. Those pods are managed by a deployment named `cnpg-controller-manager` . 　 Collect the relevant information about the operator that is running in pod ```` with: .. code:: shell kubectl describe pod -n cnpg-system Then get the logs from the same pod by running: .. code:: shell kubectl logs -n cnpg-system Gather more information about the operator ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ Get logs from all pods in CloudNativePG operator Deployment (in case you have a multi operator deployment) by running: .. code:: shell kubectl logs -n cnpg-system \ deployment/cnpg-controller-manager --all-containers=true .. tip:: You can add `-f` flag to above command to follow logs in real time. 　 Save logs to a JSON file by running: .. code:: shell kubectl logs -n cnpg-system \ deployment/cnpg-controller-manager --all-containers=true | \ jq -r . > cnpg_logs.json Get CloudNativePG operator version by using ``kubectl-cnpg`` plugin: .. code:: shell kubectl-cnpg status Output: .. code:: shell Cluster in healthy state Name: cluster-example Namespace: default System ID: 7044925089871458324 PostgreSQL Image: ghcr.io/cloudnative-pg/postgresql:18.1-system-trixie Primary instance: cluster-example-1 Instances: 3 Ready instances: 3 Current Write LSN: 0/5000000 (Timeline: 1 - WAL File: 000000010000000000000004) Continuous Backup status Not configured Streaming Replication status Name Sent LSN Write LSN Flush LSN Replay LSN Write Lag Flush Lag Replay Lag State Sync State Sync Priority - --- -------- --------- --------- ---------- --------- --------- ---------- ----- ---------- ------------- cluster-example-2 0/5000000 0/5000000 0/5000000 0/5000000 00:00:00 00:00:00 00:00:00 streaming async 0 cluster-example-3 0/5000000 0/5000000 0/5000000 0/5000000 00:00:00.10033 00:00:00.10033 00:00:00.10033 streaming async 0 Instances status Name Database Size Current LSN Replication role Status QoS Manager Version - --- ------------- ----------- ---------------- ------ --- --------------- cluster-example-1 33 MB 0/5000000 Primary OK BestEffort 1.12.0 cluster-example-2 33 MB 0/5000000 Standby (async) OK BestEffort 1.12.0 cluster-example-3 33 MB 0/5000060 Standby (async) OK BestEffort 1.12.0 Cluster information ------------------- You can check the status of the ```` cluster in the ``NAMESPACE`` namespace with: .. code:: shell kubectl get cluster -n Output: .. code:: shell NAME AGE INSTANCES READY STATUS PRIMARY 10d4h3m 3 3 Cluster in healthy state -1 The above example reports a healthy PostgreSQL cluster of 3 instances, all in *ready* state, and with ``-1`` being the primary. In case of unhealthy conditions, you can discover more by getting the manifest of the ``Cluster`` resource: .. code:: shell kubectl get cluster -o yaml -n Another important command to gather is the ``status`` one, as provided by the ``cnpg`` plugin: .. code:: shell kubectl cnpg status -n .. tip:: You can print more information by adding the `--verbose` option. 　 Get PostgreSQL container image version: .. code:: shell kubectl describe cluster -n | grep "Image Name" Output: .. code:: shell Image Name: ghcr.io/cloudnative-pg/postgresql:18.1-system-trixie .. Note:: Also you can use `kubectl-cnpg status -n ` to get the same information. 　 Pod information --------------- You can retrieve the list of instances that belong to a given PostgreSQL cluster with: .. code:: shell kubectl get pod -l cnpg.io/cluster= -L role -n Output: .. code:: shell NAME READY STATUS RESTARTS AGE ROLE -1 1/1 Running 0 10d4h5m primary -2 1/1 Running 0 10d4h4m replica -3 1/1 Running 0 10d4h4m replica You can check if/how a pod is failing by running: .. code:: shell kubectl get pod -n -o yaml - You can get all the logs for a given PostgreSQL instance with: .. code:: shell kubectl logs -n - If you want to limit the search to the PostgreSQL process only, you can run: .. code:: shell kubectl logs -n - | \ jq select(.logger=="postgres") | .record.message The following example also adds the timestamp: .. code:: shell kubectl logs -n - | \ jq -r select(.logger=="postgres") | [.ts, .record.message] | @csv If the timestamp is displayed in Unix Epoch time, you can convert it to a user-friendly format: .. code:: shell kubectl logs -n - | \ jq -r select(.logger=="postgres") | [(.ts|strflocaltime("%Y-%m-%dT%H:%M:%S %Z")), .record.message] | @csv Gather and filter extra information about PostgreSQL pods ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ Check logs from a specific pod that has crashed: .. code:: shell kubectl logs -n --previous - Get FATAL errors from a specific PostgreSQL pod: .. code:: shell kubectl logs -n - | \ jq -r .record | select(.error_severity == "FATAL") Output: .. code:: json { "log_time": "2021-11-08 14:07:44.520 UTC", "user_name": "streaming_replica", "process_id": "68", "connection_from": "10.244.0.10:60616", "session_id": "61892f30.44", "session_line_num": "1", "command_tag": "startup", "session_start_time": "2021-11-08 14:07:44 UTC", "virtual_transaction_id": "3/75", "transaction_id": "0", "error_severity": "FATAL", "sql_state_code": "28000", "message": "role \"streaming_replica\" does not exist", "backend_type": "walsender" } Filter PostgreSQL DB error messages in logs for a specific pod: .. code:: shell kubectl logs -n - | jq -r .err | select(. != null) Output: .. code:: shell dial unix /controller/run/.s.PGSQL.5432: connect: no such file or directory Get messages matching ``err`` word from a specific pod: .. code:: shell kubectl logs -n - | jq -r .msg | grep "err" Output: .. code:: shell 2021-11-08 14:07:39.610 UTC [15] LOG: ending log output to stderr Get all logs from PostgreSQL process from a specific pod: .. code:: shell kubectl logs -n - | \ jq -r . | select(.logger == "postgres") | select(.msg != "record") | .msg Output: .. code:: shell 2021-11-08 14:07:52.591 UTC [16] LOG: redirecting log output to logging collector process 2021-11-08 14:07:52.591 UTC [16] HINT: Future log output will appear in directory "/controller/log". 2021-11-08 14:07:52.591 UTC [16] LOG: ending log output to stderr 2021-11-08 14:07:52.591 UTC [16] HINT: Future log output will go to log destination "csvlog". Get pod logs filtered by fields with values and join them separated by ``|`` running: .. code:: shell kubectl logs -n - | \ jq -r [.level, .ts, .logger, .msg] | join(" | ") Output: .. code:: shell info | 1636380469.5728037 | wal-archive | Backup not configured, skip WAL archiving info | 1636383566.0664876 | postgres | record Backup information ------------------ You can list the backups that have been created for a named cluster with: .. code:: shell kubectl get backup -l cnpg.io/cluster= Storage information ------------------- Sometimes is useful to double-check the StorageClass used by the cluster to have some more context during investigations or troubleshooting, like this: .. code:: shell STORAGECLASS=$(kubectl get pvc -o jsonpath={.spec.storageClassName}) kubectl get storageclasses $STORAGECLASS -o yaml We are taking the StorageClass from one of the cluster pod here since often clusters are created using the default StorageClass. Node information ---------------- Kubernetes nodes is where ultimately PostgreSQL pods will be running. It’s strategically important to know as much as we can about them. You can get the list of nodes in your Kubernetes cluster with: .. code:: shell ## look at the worker nodes and their status kubectl get nodes -o wide Additionally, you can gather the list of nodes where the pods of a given cluster are running with: .. code:: shell kubectl get pod -l cnpg.io/cluster= \ -L role -n -o wide The latter is important to understand where your pods are distributed - very useful if you are using :ref:`affinity/anti-affinity rules and/or tolerations ` . Conditions ---------- Like many native kubernetes objects `like here `_ , Cluster exposes ``status.conditions`` as well. This allows one to ‘wait’ for a particular event to occur instead of relying on the overall cluster health state. Available conditions as of now are: - LastBackupSucceeded - ContinuousArchiving - Ready ``LastBackupSucceeded`` is reporting the status of the latest backup. If set to ``True`` the last backup has been taken correctly, it is set to ``False`` otherwise. ``ContinuousArchiving`` is reporting the status of the WAL archiving. If set to ``True`` the last WAL archival process has been terminated correctly, it is set to ``False`` otherwise. ``Ready`` is ``True`` when the cluster has the number of instances specified by the user and the primary instance is ready. This condition can be used in scripts to wait for the cluster to be created. How to wait for a particular condition ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ - Backup: .. code:: bash $ kubectl wait --for=condition=LastBackupSucceeded cluster/ -n - ContinuousArchiving: .. code:: bash $ kubectl wait --for=condition=ContinuousArchiving cluster/ -n - Ready (Cluster is ready or not): .. code:: bash $ kubectl wait --for=condition=Ready cluster/ -n Below is a snippet of a ``cluster.status`` that contains a failing condition. .. code:: bash $ kubectl get cluster/ -o yaml . . . status: conditions: - message: unexpected failure invoking barman-cloud-wal-archive: exit status 2 reason: ContinuousArchivingFailing status: "False" type: ContinuousArchiving - message: exit status 2 reason: LastBackupFailed status: "False" type: LastBackupSucceeded - message: Cluster Is Not Ready reason: ClusterIsNotReady status: "False" type: Ready Networking ---------- CloudNativePG requires basic networking and connectivity in place. You can find more information in the :ref:`Networking ` section. If installing CloudNativePG in an existing environment, there might be network policies in place, or other network configuration made specifically for the cluster, which could have an impact on the required connectivity between the operator and the cluster pods and/or the between the pods. You can look for existing network policies with the following command: .. code:: sh kubectl get networkpolicies There might be several network policies set up by the Kubernetes network administrator. .. code:: sh $ kubectl get networkpolicies NAME POD-SELECTOR AGE allow-prometheus cnpg.io/cluster=cluster-example 47m default-deny-ingress 57m PostgreSQL core dumps --------------------- Although rare, PostgreSQL can sometimes crash and generate a core dump in the ``PGDATA`` folder. When that happens, normally it is a bug in PostgreSQL (and most likely it has already been solved - this is why it is important to always run the latest minor version of PostgreSQL). CloudNativePG allows you to control what to include in the core dump through the ``cnpg.io/coredumpFilter`` annotation. .. Note:: Please refer to :ref:`Labels and Annotations ` for more details on the standard annotations that CloudNativePG provides. 　 By default, the ``cnpg.io/coredumpFilter`` is set to ``0x31`` in order to exclude shared memory segments from the dump, as this is the safest approach in most cases. .. Note:: Please refer to `Core dump filtering settings `_ . for more details on how to set the bitmask that controls the core dump filter. 　 .. Note:: Beware that this setting only takes effect during Pod startup and that changing the annotation doesn't trigger an automated rollout of the instances. 　 Although you might not personally be involved in inspecting core dumps, you might be asked to provide them so that a Postgres expert can look into them. First, verify that you have a core dump in the ``PGDATA`` directory with the following command (please run it against the correct pod where the Postgres instance is running): .. code:: sh kubectl exec -ti POD -c postgres \ -- find /var/lib/postgresql/data/pgdata -name core.* Under normal circumstances, this should return an empty set. Suppose, for example, that we have a core dump file: :: /var/lib/postgresql/data/pgdata/core.14177 Once you have verified the space on disk is sufficient, you can collect the core dump on your machine through ``kubectl cp`` as follows: .. code:: sh kubectl cp POD:/var/lib/postgresql/data/pgdata/core.14177 core.14177 You now have the file. Make sure you free the space on the server by removing the core dumps. Visualizing and Analyzing Profiling Data ---------------------------------------- CloudNativePG integrates with `pprof `_ to collect and analyze profiling data at two levels: - **Operator level** – enable by adding the ``--pprof-server=true`` option to the operator deployment (see :ref:`Operator configuration ` ). - **Postgres cluster level** – enable by adding the ``alpha.cnpg.io/enableInstancePprof`` annotation to a ``Cluster`` resource (described below). When the ``alpha.cnpg.io/enableInstancePprof`` annotation is set to ``"true"`` , each instance pod exposes a Go pprof HTTP server provided by the instance manager. - The server listens on ``0.0.0.0:6060`` inside the pod. - A container port named ``pprof`` (``6060/TCP`` ) is automatically added to the pod spec. You can disable pprof at any time by either removing the annotation or setting it to ``"false"`` . The operator will roll out changes automatically to remove the pprof port and flag. .. Note:: The pprof server only serves plain HTTP on port `6060` . 　 Example ^^^^^^^ Enable pprof on a cluster by adding the annotation: .. code:: yaml apiVersion: postgresql.cnpg.io/v1 kind: Cluster metadata: name: cluster-example annotations: alpha.cnpg.io/enableInstancePprof: "true" spec: instances: 3 # ... Changing this annotation updates the instance pod spec (adds port ``6060`` and the corresponding flag) and triggers a rolling update. .. Warning:: The example below uses `kubectl port-forward` for local testing only. This is **not** the intended way to expose the feature in production. Treat pprof as a sensitive debugging interface and never expose it publicly. If you must access it remotely, secure it with proper network policies and access controls. 　 Use port-forwarding to access the pprof endpoints: .. code:: bash kubectl port-forward -n pod/ 6060 curl -sS http://localhost:6060/debug/pprof/ go tool pprof http://localhost:6060/debug/pprof/profile?seconds=30 You can also access pprof using the browser at `http://localhost:6060/debug/pprof/ `_ . .. _troubleshooting-1: Troubleshooting ^^^^^^^^^^^^^^^ First, verify that the cluster has the ``alpha.cnpg.io/enableInstancePprof: "true"`` annotation set. Next, check that the instance manager command includes the ``--pprof-server`` flag and that port ``6060/TCP`` is exposed. You can do this by running: .. code:: bash kubectl -n describe pod Then review the **Command** and **Ports** sections in the output. Finally, if you are not using port-forwarding, make sure that your NetworkPolicies allow access to port ``6060/TCP`` . Some known issues ----------------- Storage is full ^^^^^^^^^^^^^^^ In case the storage is full, the PostgreSQL pods will not be able to write new data, or, in case of the disk containing the WAL segments being full, PostgreSQL will shut down. If you see messages in the logs about the disk being full, you should increase the size of the affected PVC. You can do this by editing the PVC and changing the ``spec.resources.requests.storage`` field. After that, you should also update the Cluster resource with the new size to apply the same change to all the pods. Please look at the :ref:`Volume expansion ` in the documentation. If the space for WAL segments is exhausted, the pod will be crash-looping and the cluster status will report ``Not enough disk space`` . Increasing the size in the PVC and then in the Cluster resource will solve the issue. See also the :ref:`Disk Full Failure ` Pods are stuck in ``Pending`` state ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ In case a Cluster’s instance is stuck in the ``Pending`` phase, you should check the pod’s ``Events`` section to get an idea of the reasons behind this: .. code:: shell kubectl describe pod -n Some of the possible causes for this are: - No nodes are matching the ``nodeSelector`` - Tolerations are not correctly configured to match the nodes’ taints - No nodes are available at all: this could also be related to ``cluster-autoscaler`` hitting some limits, or having some temporary issues In this case, it could also be useful to check events in the namespace: .. code:: shell kubectl get events -n ## list events in chronological order kubectl get events -n --sort-by=.metadata.creationTimestamp Replicas out of sync when no backup is configured ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ Sometimes replicas might be switched off for a bit of time due to maintenance reasons (think of when a Kubernetes nodes is drained). In case your cluster does not have backup configured, when replicas come back up, they might require a WAL file that is not present anymore on the primary (having been already recycled according to the WAL management policies as mentioned in :ref:`The `postgresql` section ` ), and fall out of synchronization. Similarly, when ``pg_rewind`` might require a WAL file that is not present anymore in the former primary, reporting ``pg_rewind: error: could not open file`` . In these cases, pods cannot become ready anymore, and you are required to delete the PVC and let the operator rebuild the replica. If you rely on dynamically provisioned Persistent Volumes, and you are confident in deleting the PV itself, you can do so with: .. code:: shell PODNAME= VOLNAME=$(kubectl get pv -o json | \ jq -r .items[]|select(.spec.claimRef.name==\"$PODNAME\")|.metadata.name) kubectl delete pod/$PODNAME pvc/$PODNAME pvc/$PODNAME-wal pv/$VOLNAME Cluster stuck in ``Creating new replica`` ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ Cluster is stuck in “Creating a new replica”, while pod logs don’t show relevant problems. This has been found to be related to the next issue :ref:`Networking is impaired by installed Network Policies ` . Networking issues are reflected in the status column as follows: .. code:: text Instance Status Extraction Error: HTTP communication issue Networking is impaired by installed Network Policies ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ As pointed out in the :ref:`Networking ` , local network policies could prevent some of the required connectivity. A tell-tale sign that connectivity is impaired is the presence in the operator logs of messages like: .. code:: text "Cannot extract Pod status", […snipped…] "Get \"http://:8000/pg/status\": dial tcp :8000: i/o timeout" You should list the network policies, and look for any policies restricting connectivity. .. code:: sh $ kubectl get networkpolicies NAME POD-SELECTOR AGE allow-prometheus cnpg.io/cluster=cluster-example 47m default-deny-ingress 57m For example, in the listing above, ``default-deny-ingress`` seems a likely culprit. You can drill into it: .. code:: sh $ kubectl get networkpolicies default-deny-ingress -o yaml <…snipped…> spec: podSelector: {} policyTypes: - Ingress In the :ref:`networking page ` you can find a network policy file that you can customize to create a ``NetworkPolicy`` explicitly allowing the operator to connect cross-namespace to cluster pods. Error while bootstrapping the data directory ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ If your Cluster’s initialization job crashes with a “Bus error (core dumped) child process exited with exit code 135”, you likely need to fix the Cluster hugepages settings. The reason is the incomplete support of hugepages in the cgroup v1 that should be fixed in v2. For more information, check the PostgreSQL `BUG #17757: Nothonoring huge_pages setting during initdb causes DB crash inKubernetes `_ . To check whether hugepages are enabled, run ``grep HugePages /proc/meminfo`` on the Kubernetes node and check if hugepages are present, their size, and how many are free. If the hugepages are present, you need to configure how much hugepages memory every PostgreSQL pod should have available. For example: .. code:: yaml postgresql: parameters: shared_buffers: "128MB" resources: requests: memory: "512Mi" limits: hugepages-2Mi: "512Mi" Please remember that you must have enough hugepages memory available to schedule every Pod in the Cluster (in the example above, at least 512MiB per Pod must be free). Bootstrap job hangs in running status ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ If your Cluster’s initialization job hangs while in ``Running`` status with the message: “error while waiting for the API server to be reachable”, you probably have a network issue preventing communication with the Kubernetes API server. Initialization jobs (like most of jobs) need to access the Kubernetes API. Please check your networking. Another possible cause is when you have sidecar injection configured. Sidecars such as Istio may make the network temporarily unavailable during startup. If you have sidecar injection enabled, retry with injection disabled. Replicas take over two minutes to reconnect after a failover ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ When the primary instance fails, the operator promotes the most advanced standby to the primary role. Other standby instances then attempt to reconnect to the ``-rw`` service for replication. However, during this reconnection process, ``kube-proxy`` may not have updated its routing information yet. As a result, the initial ``SYN`` packet sent by the standby instances might fail to reach its intended destination. If the network is configured to silently drop packets instead of rejecting them, standby instances will not receive a response and will retry the connection after an exponential backoff period. On Linux systems, the default value for the ``tcp_syn_retries`` kernel parameter is 6, meaning the system will attempt to establish the connection for approximately 127 seconds before giving up. This prolonged retry period can significantly delay the reconnection process. For more details, consult the `tcp_syn_retries documentation `_ . You can work around this issue by setting ``STANDBY_TCP_USER_TIMEOUT`` in the :ref:`Operator configuration ` . This will cause the standby instances to close the TCP connection if the initial ``SYN`` packet is not acknowledged within the specified timeout, allowing them to retry the connection more quickly.