[HBASE-18626] Handle the incompatible change about the replication TableCFs' config

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.0.0-alpha-1, 1.4.0, 1.5.0, 2.0.0-alpha-3
• Fix Version/s: 1.4.0
• Assignee: Guanghao Zhang
• Reporter: Guanghao Zhang
• Created: 18/Aug/17 05:27
• Updated: 01/Aug/18 06:21
• Resolved: 22/Mar/18 20:53

Description

About compatibility, there is one incompatible change about the replication TableCFs' config. The old config is a string and it concatenate the list of tables and column families in format "table1:cf1,cf2;table2:cfA,cfB" in zookeeper for table-cf to replication peer mapping. When parse the config, it use ":" to split the string. If table name includes namespace, it will be wrong (See HBASE-11386). It is a problem since we support namespace (0.98). So HBASE-11393 (and HBASE-16653) changed it to a PB object. When rolling update cluster, you need rolling master first. And the master will try to translate the string config to a PB object. But there are two problems.
1. Permission problem. The replication client can write the zookeeper directly. So the znode may have different owner. And master may don't have the write permission for the znode. It maybe failed to translate old table-cfs string to new PB Object. See HBASE-16938
2. We usually keep compatibility between old client and new server. But the old replication client may write a string config to znode directly. Then the new server can't parse them.

Issue Links

• is related to: HBASE-11393 Replication TableCfs should be a PB object rather than a string
• is related to: HBASE-16653 Backport HBASE-11393 to all branches which support namespace
• is related to: HBASE-16938 TableCFsUpdater maybe failed due to no write permission on peerNode
• is related to: HBASE-11386 Replication#table,CF config will be wrong if the table name includes namespace