[HIVE-7341] Support for Table replication across HCatalog instances

• Type: New Feature
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 0.13.1
• Fix Version/s: 0.14.0
• Component/s: HCatalog
• Labels: backward-incompatible data-replication
• Assignee: Mithun Radhakrishnan
• Reporter: Mithun Radhakrishnan
• Created: 02/Jul/14 22:52
• Updated: 06/Nov/15 01:16
• Resolved: 18/Aug/14 11:02

Description

The HCatClient currently doesn't provide very much support for replicating HCatTable definitions between 2 HCatalog Server (i.e. Hive metastore) instances.

Systems similar to Apache Falcon might find the need to replicate partition data between 2 clusters, and keep the HCatalog metadata in sync between the two. This poses a couple of problems:

1. The definition of the source table might change (in column schema, I/O formats, record-formats, serde-parameters, etc.) The system will need a way to diff 2 tables and update the target-metastore with the changes. E.g.

   targetTable.resolve( sourceTable, targetTable.diff(sourceTable) );
   hcatClient.updateTableSchema(dbName, tableName, targetTable);
   

2. The current HCatClient.addPartitions() API requires that the partition's schema be derived from the table's schema, thereby requiring that the table-schema be resolved before partitions with the new schema are added to the table. This is problematic, because it introduces race conditions when 2 partitions with differing column-schemas (e.g. right after a schema change) are copied in parallel. This can be avoided if each HCatAddPartitionDesc kept track of the partition's schema, in flight.
3. The source and target metastores might be running different/incompatible versions of Hive.

The impending patch attempts to address these concerns (with some caveats).

1. HCatTable now has
   1. a diff() method, to compare against another HCatTable instance
   2. a resolve(diff) method to copy over specified table-attributes from another HCatTable
   3. a serialize/deserialize mechanism (via HCatClient.serializeTable() and HCatClient.deserializeTable()), so that HCatTable instances constructed in other class-loaders may be used for comparison
2. HCatPartition now provides finer-grained control over a Partition's column-schema, StorageDescriptor settings, etc. This allows partitions to be copied completely from source, with the ability to override specific properties if required (e.g. location).
3. HCatClient.updateTableSchema() can now update the entire table-definition, not just the column schema.
4. I've cleaned up and removed most of the redundancy between the HCatTable, HCatCreateTableDesc and HCatCreateTableDesc.Builder. The prior API failed to separate the table-attributes from the add-table-operation's attributes. By providing fluent-interfaces in HCatTable, and composing an HCatTable instance in HCatCreateTableDesc, the interfaces are cleaner(ish). The old setters are deprecated, in favour of those in HCatTable. Likewise, HCatPartition and HCatAddPartitionDesc.

I'll post a patch for trunk shortly.

Issue Links

• is depended upon by: HIVE-7576 Add PartitionSpec support in HCatClient API
• is duplicated by: HIVE-4274 Table created using HCatalog java client doesn't set the owner
• is related to: HIVE-7770 Undo backward-incompatible behaviour change introduced by HIVE-7341
• is related to: HIVE-6378 HCatClient::createTable() doesn't allow SerDe class to be specified