...

1. The following settings are required in hive-site.xml to enable ACID support for streaming:
   1. hive.txn.manager = org.apache.hadoop.hive.ql.lockmgr.DbTxnManager
   2. hive.compactor.initiator.on = true (See more important details here)
   3. hive.compactor.worker.threads > 0 
2. “stored as orc” must be specified during table creation. Only ORC storage format is supported currently.
3. tblproperties("transactional"="true") must be set on the table during creation.

4. User of the client streaming process must have the necessary permissions to write to the table or partition and create partitions in the table.

5. (Temporary requirements) When issuing queries on streaming tables, the client needs to set

   1. hive.vectorized.execution.enabled  to  false (for Hive version < 0.14.0)

   2. hive.input.format  to  org.apache.hadoop.hive.ql.io.HiveInputFormat

Limitations

Out of the box, currently, the streaming API only provides support for streaming delimited input data (such as CSV, tab separated, etc.), JSON and Regex formatted data. The records writers that are supported in Hive 3.0.0 release are

...

All these record writers expects strict schema match, meaning the schema of the records should exactly match with the table schema (Note: The writers does not perform schema check, it is upto up to the clients to make sure the record schema matches the table schema). 

...

Transactions are implemented slightly differently than traditional database systems. Each transaction has an id and multiple transactions are grouped into a “Transaction Batch”. This helps grouping records from multiple transactions into fewer files (rather than 1 file per transaction). During hive streaming connection creation, transaction batch size can be specified via builder API. Transaction management is completely hidden behind the API, in most cases users do not have to worry about tuning the transaction batch size (which is an expert level setting and might not be honored in future release). Also the API automatically rolls over to next transaction batch on beginTransaction() invocation if the current transaction batch is exhausted. The recommendation is to leave the transaction batch size at default value of 1 and group several thousands records together under a each transaction. Since each transaction corresponds to a delta directory in the filesystem, committing transaction too often can end up creating too many small directories. 

...