[IGNITE-6024] SQL: execute DML statements on the server when possible

• Type: Task
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.1
• Fix Version/s: 2.3
• Component/s: sql
• Labels: important performance
• Assignee: Sergey Kalashnikov
• Reporter: Vladimir Ozerov
• Created: 10/Aug/17 09:05
• Updated: 13/Oct/17 09:37
• Resolved: 13/Oct/17 09:34

Description

Currently we execute DML statements as follows:
1) Get query result set to the client
2) Construct entry processors and send them to servers in batches

This approach is inefficient as it causes a lot of unnecessary network communication Instead, we should execute DML statements directly on server nodes when it is possible.

Implementation considerations:
1) Determine set of queries which could be processed in this way. E.g., LIMIT/OFFSET, GROUP BY, ORDER BY, DISTINCT, etc. are out of question - they must go through the client anyway. Probably skipMergeTable flag is a good starting point (good, not precise!)
2) Send request to every server and execute local DML right there
3) No failover support at the moment - throw "partial update" exception if topology is unstable
4) Handle partition reservation carefully
5) Transactions: we still have single coordinator - this is a client. When MVCC and TX SQL is ready, client will assign proper counters to server requests.

Issue Links

• blocks: IGNITE-6319 SQL: implement server-side DML for lazy mode
• relates to: IGNITE-6620 Document JDBC/ODBC "skipReducerOnUpdate" flag
• links to: GitHub Pull Request #2488