Package org.apache.flink.table.runtime.operators.join.stream.utils

Class JoinHelper<STATE_VIEW,​OUTER_STATE_VIEW extends STATE_VIEW>

  • public abstract class JoinHelper<STATE_VIEW,​OUTER_STATE_VIEW extends STATE_VIEW>
extends Object
    A helper to do the logic of streaming join.

    • Method Detail

      • processJoin

        public void processJoin​(RowData input,
                        STATE_VIEW inputSideAsyncStateView,
                        STATE_VIEW otherSideAsyncStateView,
                        boolean inputIsLeft,
                        AssociatedRecords otherSideAssociatedRecords,
                        boolean isSuppress)
                 throws Exception
        Process an input element and output incremental joined records, retraction messages will be sent in some scenarios.

        Following is the pseudo code to describe the core logic of this method. The logic of this method is too complex, so we provide the pseudo code to help understand the logic. We should keep sync the following pseudo code with the real logic of the method.

        Note: "+I" represents "INSERT", "-D" represents "DELETE", "+U" represents "UPDATE_AFTER", "-U" represents "UPDATE_BEFORE". We forward input RowKind if it is inner join, otherwise, we always send insert and delete for simplification. We can optimize this to send -U & +U instead of D & I in the future (see FLINK-17337). They are equivalent in this join case. It may need some refactoring if we want to send -U & +U, so we still keep -D & +I for now for simplification. See FlinkChangelogModeInferenceProgram.SatisfyModifyKindSetTraitVisitor. 

         if input record is accumulate
 |  if input side is outer
 |  |  if there is no matched rows on the other side, send +I[record+null], state.add(record, 0)
 |  |  if there are matched rows on the other side
 |  |  | if other side is outer
 |  |  | |  if the matched num in the matched rows == 0, send -D[null+other]
 |  |  | |  if the matched num in the matched rows > 0, skip
 |  |  | |  otherState.update(other, old + 1)
 |  |  | endif
 |  |  | send +I[record+other]s, state.add(record, other.size)
 |  |  endif
 |  endif
 |  if input side not outer
 |  |  state.add(record)
 |  |  if there is no matched rows on the other side, skip
 |  |  if there are matched rows on the other side
 |  |  |  if other side is outer
 |  |  |  |  if the matched num in the matched rows == 0, send -D[null+other]
 |  |  |  |  if the matched num in the matched rows > 0, skip
 |  |  |  |  otherState.update(other, old + 1)
 |  |  |  |  send +I[record+other]s
 |  |  |  else
 |  |  |  |  send +I/+U[record+other]s (using input RowKind)
 |  |  |  endif
 |  |  endif
 |  endif
 endif

 if input record is retract
 |  state.retract(record)
 |  if there is no matched rows on the other side
 |  | if input side is outer, send -D[record+null]
 |  endif
 |  if there are matched rows on the other side, send -D[record+other]s if outer, send -D/-U[record+other]s if inner.
 |  |  if other side is outer
 |  |  |  if the matched num in the matched rows == 0, this should never happen!
 |  |  |  if the matched num in the matched rows == 1, send +I[null+other]
 |  |  |  if the matched num in the matched rows > 1, skip
 |  |  |  otherState.update(other, old - 1)
 |  |  endif
 |  endif
 endif
        Parameters:
        input - the input element
        inputSideAsyncStateView - state of input side
        otherSideAsyncStateView - state of other side
        inputIsLeft - whether input side is left side
        otherSideAssociatedRecords - associated records in the state of the other side
        Throws:
        Exception