[FLINK-15775] SourceFunctions are instantiated twice when pulled on from 2 Sinks

• Type: Bug
• Status: Closed
• Priority: Minor
• Resolution: Invalid
• Affects Version/s: 1.9.1, 1.10.0
• Fix Version/s: 1.11.0
• Component/s: Table SQL / Planner
• Labels: auto-deprioritized-major
• Release Note: Has fixed after flink 1.11
• Assignee: Unassigned
• Reporter: Benoît Paris
• Created: 27/Jan/20 11:05
• Updated: 03/Dec/21 14:13
• Resolved: 28/May/21 08:39

Description

When pulled on by two sinks, the SourceFunctions of a TableSource will get instantiated twice; (and subsequently opened by the parallelism number, which is expected behavior):

The following will instantiate the FooTableSource's SourceFunction once (OK behavior, but not the processing we want):

 

tEnv.registerTableSource("foo_table", new FooTableSource());
Table out0 = tEnv.sqlQuery("SELECT * FROM foo_table WHERE field_1 = 0");
tEnv.registerTableSink("syso_sink_0", new SysoSink());
out0.insertInto("syso_sink_0");

 

This will instantiate the FooTableSource's SourceFunction twice (Not OK, as we're missing half the inputs in each SysoSink):

 

tEnv.registerTableSource("foo_table", new FooTableSource());
Table out0 = tEnv.sqlQuery("SELECT * FROM foo_table WHERE field_1 = 0");
Table out1 = tEnv.sqlQuery("SELECT * FROM foo_table WHERE field_1 = 1");
tEnv.registerTableSink("syso_sink_0", new SysoSink());
tEnv.registerTableSink("syso_sink_1", new SysoSink());
out0.insertInto("syso_sink_0");
out1.insertInto("syso_sink_1"); 

 

This might not be a problem for Kafka's SourceFunctions, as we can always reread from a log; but it is a data loss problem when the source data can't be reproduced.

Actually, this might be me not understanding the API. Is there a way to make the runtime read from the same opened SourceFunctions?

Attached is Java code that logs the faulty opening of the SourceFunctions, pom.xml, and logical execution plans for the duplicated case, and the workaround.

 

---

Workaround: make a conversion to an appendStream. Somehow this makes the planner think it has to put a materialization barrier after the Source and read from that:

 

tEnv.registerTableSource("foo_table_source", new FooTableSource());
Table sourceTable = tEnv.sqlQuery("SELECT * FROM foo_table_source");
Table appendingSourceTable = tEnv.fromDataStream(
 tEnv.toAppendStream(sourceTable, Types.ROW(new String[]{"field_1"}, new TypeInformation[]{Types.LONG()}))
);
tEnv.registerTable("foo_table", appendingSourceTable);

 

 

Best Regards,

Ben