Study Guide: Week 6#
Open-ended questions on the material from week 6. The exam asks multiple-choice versions of questions like these. No answers are given: each question can be answered in a few sentences from the pages listed under its heading. The exam’s format is described in Exam Preparation.
Tests and Data Validation#
Read: Unit Tests, Data Validation, and Agenda Item 2 of the week 6 chapter.
A unit test and a plausibility check can both stop a pipeline. What does each one check?
A daily job pulls a series, checks it, and publishes a page. Why should the checks run before the page is published, and what should a failed check leave on the site?
A pipeline’s unit tests and plausibility checks all pass, and the number it publishes is wrong. How can that happen, and what kind of check would have caught it?
You reconcile your series against a second source that was computed from the same inputs. What can that comparison catch, and what can it not?
Give one data validation failure that should stop a pipeline from publishing and one that should only be flagged. What makes the difference?
The FTSFR Benchmark#
Read: The Financial Time Series Forecasting Repository and Bejarano et al. (2026).
Two studies compare the same forecasting methods on what both call corporate bond returns, and they rank the methods differently. Other than the methods, what can explain the difference?
Much of the data in a financial forecasting benchmark comes from licensed databases such as CRSP and OptionMetrics. Why does such a benchmark publish the code that builds each dataset rather than the datasets?
A benchmark builds corporate bond portfolios twice from the same source data, changing only how the sorting signal is cleaned, and the two sets of portfolio returns are 98 to 99 percent correlated. The same forecasting models are run on both. What happens to the ranking of the models?
If a ranking of forecasting methods can change when only the cleaning changes, why must a published comparison record the exact version of the data it used?
Hashes, Snapshots, and Pins#
Read: ChartBook: Know Where Every Number Came From and Agenda Items 4 and 5 of the week 6 chapter.
What is a content hash of a file, and what happens to it when one byte of the file changes?
Git names each commit by a hash computed from the hashes of its files and the hash of its parent commit. Why does editing a file in an old commit change the hash of every commit after it?
A lockfile records, for each dependency, one hash computed from the hashes of every asset pinned from it. Why is that one hash enough to tell whether anything pinned from the dependency has changed?
What is the difference between reading another project’s working folder and reading a snapshot of it?
chartbook pinrecords which snapshot of a dependency a project reads, and never moves a pin that is already there;chartbook updatemoves it. Why keep the two commands separate?A pull request’s diff of
chartbook.lockincludes these lines, under the record forftsfr/fed_yield_curve. What does the change say happened, and which command made it?- "snapshot": "2026-11-02T05:22:19Z", - "pinned_hash": "sha256:aff4…" + "snapshot": "2026-11-09T05:20:41Z", + "pinned_hash": "sha256:51e0…" ... - "content_hash": "sha256:b63d…" + "content_hash": "sha256:bc10…"
You are editing an upstream pipeline and the report that reads it, on the same machine. The report reads the upstream’s snapshot in your store, not the upstream’s working folder. What could go wrong if it read the folder, and what does the store give you instead?
Lineage and Provenance#
Read: ChartBook: Know Where Every Number Came From.
What is the difference between the lineage of a chart and its provenance?
Task Runners, Orchestrators, and Catalogs#
Read: Agenda Item 4 of the week 6 chapter.
doit, Airflow, and ChartBook each answer a different question about a pipeline. What question does each one answer?A task that appends the day’s rows to a file every time it runs is not idempotent. What goes wrong when a scheduler retries it after a failure?
Parquet#
Read: Strategies for Medium-Sized Data.
Parquet stores a table column by column. Why does reading three columns of a hundred-column Parquet file cost little more than the three columns themselves?
A Parquet file keeps the minimum and maximum of each column for every row group. How does a reader use those statistics to skip data when a query filters on that column?