Check Your Knowledge

Test your understanding of this section’s AI-assisted migration concepts before moving on.

1. In the hybrid model this section introduces, what does the “20% human accuracy” half actually consist of?
   • A. Rewriting every AI-drafted translation from scratch by hand
   • B. A structured review discipline – prompts, named failure checks, and a formal audit gate – applied to AI output
   • C. Waiting for a newer model version to fix the errors automatically
   • D. Disabling AI assistance entirely for procedural code
2. Which of the four required blocks in a deterministic transpilation prompt asks the model to flag its own uncertainty?
   • A. Context
   • B. Contract
   • C. Constraints
   • D. Self-Audit
3. Why does “pinning the model” matter for a transpilation prompt?
   • A. It makes the prompt shorter
   • B. It ensures a rerun of the same prompt later produces a comparable result instead of silent drift from a model update
   • C. It is required by Databricks licensing
   • D. It prevents the model from reading the schema
4. In the window-function hallucination, what specifically goes wrong when an LLM substitutes DENSE_RANK() for a source procedure’s RANK()?
   • A. The query fails to compile
   • B. DENSE_RANK()’s gap-free numbering can let a different set of rows pass a rank <= N filter than RANK() would have
   • C. DENSE_RANK() runs but always returns zero rows
   • D. The two functions are fully interchangeable with no behavioral difference
5. What invariant helps detect that DENSE_RANK() produced a migrated result, whether or not that was intended?
   • A. The row count always doubles
   • B. The maximum rank value in a partition equals the count of distinct ranked values in that partition
   • C. The minimum rank value is always zero
   • D. Every partition has exactly one row
6. In the “Syntactically Valid, Semantically Wrong” gallery, why does a naive Oracle DATE → Spark DATE mapping lose information?
   • A. Spark’s DATE type stores more precision than Oracle’s
   • B. Oracle’s DATE always carries a time-of-day component that Spark’s timeless DATE type has no way to hold
   • C. Oracle does not support a DATE type
   • D. The mapping causes a compile error, so it’s caught immediately
7. Why does an Oracle CHAR(n) to Spark STRING port risk silently breaking a join?
   • A. STRING columns cannot be used in joins
   • B. Oracle blank-pads and ignores trailing spaces in comparisons; Spark’s STRING does neither, so a value that matched before may not match after
   • C. Oracle CHAR columns cannot hold text data
   • D. Spark automatically trims all string columns before joining
8. In the transaction-semantics gap example, what is the correct redesign for an Oracle cursor loop that commits every hundred rows?
   • A. Reproduce a per-hundred-row commit checkpoint literally in PySpark
   • B. Replace the row-by-row commit loop with a single atomic MERGE INTO keyed on the record’s natural key
   • C. Remove all commit logic and let the table stay in whatever state the job leaves it
   • D. Run the procedure twice and compare results
9. Why must a redesigned transaction boundary (like the one in question 8) be explicitly documented rather than shipped silently?
   • A. Documentation is a compliance requirement unrelated to correctness
   • B. It changes real failure semantics – from partial-batch commits to all-or-nothing – which stakeholders need to know about
   • C. Delta Lake requires a comment above every MERGE statement
   • D. It has no actual effect on behavior, so documentation is just good practice
10. What does the six-point gate audit consider a “partial pass” – for example, five of six checks succeeding?
    • A. Acceptable, since most checks passed
    • B. A full fail – the golden-diff and other checks are binary pass/fail, not partial credit
    • C. Grounds for shipping with a follow-up ticket
    • D. Irrelevant, since only check 1 (row-count parity) actually matters

Answer Key

1. B – The human-accuracy half is the review discipline itself: structured prompts, named failure-mode checks, and a formal audit gate, not manual rewriting or waiting for a better model.
2. D – The Self-Audit block asks the model to list its own assumptions and flag constructs it’s uncertain about.
3. B – Pinning the model version makes a later rerun of the same prompt comparable rather than silently different due to an underlying model update.
4. B – DENSE_RANK()’s gap-free numbering can admit a different set of rows through a rank <= N filter than RANK()’s gapped numbering would, with no error raised.
5. B – DENSE_RANK()’s maximum value in a partition always equals the count of distinct ranked values; this equality breaking is the sign RANK() (with a tie) was actually intended.
6. B – Oracle’s DATE type always carries a time-of-day component; Spark’s timeless DATE type silently discards it on a naive mapping.
7. B – Oracle’s CHAR(n) blank-pads and ignores trailing spaces in comparisons; Spark’s STRING does neither, so previously-matching values can fail to match.
8. B – The correct move is a deliberate redesign to a single atomic MERGE INTO on the natural key, not a literal translation of the per-row commit loop.
9. B – The redesign changes real failure semantics (partial-batch vs. all-or-nothing), which is a decision stakeholders need to be told about, not discover later.
10. B – The six-point audit treats each check as binary; a golden-diff mismatch or any other failed check is a full fail regardless of how many other checks passed.

---

| ← Previous: The LLM Prompt Library and Audit Checklist | Next: The Ingestion Decision Tree → | |:—|—:|