Cloudera Data Hub rolling upgrade limitations and issues
Cloudera Data Hub rolling upgrades have the following limitations:
Cloudera Operational Database clusters
-
HBase commands may fail during the rolling upgrade with the error "ServerNotRunningYetException: Server is not running yet." HBase retries DDL operations submitted while the master is initializing until the master is fully initialized to serve the request. However, a situation might arise where the default number of retries or intervals proves to be insufficient for an operation submitted by the client to complete.
Implementing the following configuration adjustments in your client application can support the master getting initialized up to 10 minutes:
<property> <name>hbase.client.pause</name> <value>300</value> </property> <property> <name>hbase.client.retries.number</name> <value>20</value> </property>
If you have seen a longer or shorter master initialization period, you can modify these values accordingly. These retry settings apply to all types of calls to HBase service, encompassing GET, SCAN, MUTATE, and DDLs.
-
In a rolling restart, if a COD cluster has less than 10 datanodes, existing writes can fail with an error indicating a new block cannot be allocated and all nodes are excluded. This is because the client has attempted to use all the datanodes in the cluster, and failed to write to each of them as they were restarted. This will only happen on small clusters of less than 10 datanodes, as larger clusters have more spare nodes to allow the write to continue.
- When performing a maintenance upgrade or other cluster upgrade, in some occasions there can
been an error when the upgrade is nearly complete, but trying to re-start services/roles.
The error is similar to: "Failed to start role
hue-HUE_SERVER-8cc9321b2213cc5c6846c64e1fc6b1cb of service hue in cluster
cod--xoaitnb0wnl1. This role requires the following additional parcels to be activated
before it can start: [cdh]."
This is due to an agent operation that sometimes is delayed and can interfere with the role start. When this happens, resume the failed upgrade from Cloudera Manager as a 'Full Admin' user.
- During the VM replacement as part of OS upgrade, every new node gets a new IP Address, and
if the old IP address is cached somewhere, HDFS requests fail with
UnknownHostException
. It recovers after some time (10 mins max). - If Knox is HA and one of the Knox servers is down, then accessing the service through a Cloudera Control Plane endpoint URL (i.e., through cloud load balancer) will take approximately 30 seconds to failover the request to the available Knox instance. This also means that the services that are reached through Knox will not be available behind Knox during this period time.
- During OS upgrades, attempts to access Knox on the host being upgraded may produce occasional 403 HTTP responses. Wait and retry the failed requests.
- When upgrading Cloudera Data Hub clusters to Cloudera Runtime 7.2.18.100, you might encounter staleness in
knox.jwt.client.gateway.address
configuration in case its value points to the address of the Data Lake node. If staleness occurs after the upgrade, you need to run Deploy Client Configuration in Cloudera Manager.
For more limitations of Cloudera Operational Database, see Rolling upgrade limitations (Cloudera Operational Database).
Cloudera Streams Messaging clusters
- Rolling upgrades are not supported for Cruise Control or Streams Messaging Manager (SMM). When upgrading a Streams Messaging cluster, expect that both of these services will be temporarily unavailable during the upgrade. This, however, does not impact Kafka's ability to perform a rolling upgrade.
- Rolling upgrades for Schema Registry are only supported from Cloudera Runtime 7.2.18 or higher to higher versions. When upgrading a Streams Messaging cluster from a lower version, expect that clients connecting to the Schema Registry service might experience downtime. This, however, does not impact Kafka's ability to perform a rolling upgrade.
- If Knox is HA and one of the Knox servers is down, then accessing the service through a Cloudera Control Plane endpoint URL (i.e., through cloud load balancer) will take approximately 30 seconds to failover the request to the available Knox instance. This also means that the services that are reached through Knox will not be available behind Knox during this period time.
- During OS upgrades, attempts to access Knox on the host being upgraded may produce occasional 403 HTTP responses. Wait and retry the failed requests.
- When upgrading Cloudera Data Hub clusters to Cloudera Runtime 7.2.18.100, you might encounter staleness in
knox.jwt.client.gateway.address
configuration in case its value points to the address of the Data Lake node. If staleness occurs after the upgrade, you need to run Deploy Client Configuration in Cloudera Manager.