Let's try the following:

• Create a new C++ unit test in a new file, with similar dependencies as sbe_abt_test.cpp.
• Write a single example:
  • query in MQL
  • input documents
  • index specs
• Translate the query to ABT
• Get the query results as naively as possible. For example:
  • Disable constant folding
  • Disable exploration rewrites
  • Use a ScanDef with no indexes
• Generate all query plans:
  • Leave constant folding, exploration, everything else enabled.
  • Disable branch and bound.
  • Keep rejected plans.
  • Run OptPhaseManager once.
  • Write a new function that extracts all plans from the final memo.
• Lower to SBE, and run, each plan.
• Compare each result to the naive result.
  • Use sorting to ignore both field order, and document order.

In other words, this is a property test that holds the query and input data fixed, and asserts: all plans for this query are correct. This means optimization bugs can't hide behind costing decisions. But we still have control over which examples we test.

This will also be a good opportunity to test cost predictions.