Skip to content

Systems and records

Multi-system dataset card: what 6,589 retail campaign episodes expose

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

A multi-system dataset card should define the episode first, then measure every join, period, gap and right that episode depends on, instead of describing tables one at a time. Filled for the public Complete Journey retail data, the card shows 6,589 household-campaign episodes, a redemption-to-basket link that holds for 1,129 of 2,102 redemptions, campaign windows that run past the data and a CC0 package line whose upstream terms could not be located.

License terms

Multi system dataset card

This page covers multi-system dataset card: what 6,589 retail campaign episodes expose.Illustrative

Key takeaways

  • Define the episode first: in the Complete Journey package, one household receiving one campaign gives 6,589 episodes across six source tables.
  • Measure join cardinality, not only coverage: the package's documented coupon join turns 2,102 redemptions into 2,265,375 rows.
  • Where no shared key exists, publish the matching rule and its result: 1,129 of 2,102 redemptions tie to exactly one basket.
  • Give each system its own period and mark derived values: 932 episodes run past the purchase year, and every timestamp's seconds come from basket ID digits.
  • Record rights per source: the package declares CC0, but the upstream study's own terms were not located.

What should a multi-system dataset card describe first?

The episode: the unit of work that spans systems, with its key, its period and the joins it depends on. A table-by-table card can list every column correctly and still hide the fact a buyer needs most, which is whether one business episode survives the trip from five source systems into one dataset. Describe the episode, then measure each link it relies on.

Data Cards, as Pushkarna, Zaldivar and Kjartansson of Google Research describe them in their 2022 FAccT paper, summarize upstream sources, collection and annotation methods, training and evaluation methods, intended use and decisions that affect model performance. That structure fits a single source well. When records come from a point-of-sale system, a marketing system and a coupon clearing process, the facts that decide usability sit between tables: shared keys, overlapping periods and context that exists for some episodes and not others.

The decision this page supports: can a joined, multi-system dataset carry a stated task, and which numbers on its card prove it? We answer it by filling a card for a public retail dataset and measuring every field rather than describing it.

The test case: one retail episode spread across six tables

The test case is completejourney, an R data package released under CC0 that redistributes a cleaned version of a retail study: one year of grocery purchases by 2,469 households, with product data, direct-marketing campaigns, coupons and, for some households, demographics. We read its data files at commit 5b5d061 of 25 November 2025 without running any package code.

Its tables map to distinct business functions: 1,469,307 purchase lines in 155,848 baskets, 6,589 rows recording which household received which of 27 campaigns, a campaign calendar, 116,204 coupon-to-product rows, 2,102 coupon redemptions, 92,331 products and 801 household profiles, plus a sample of in-store display and mailer placements. We define the episode as one household receiving one campaign, so the dataset holds 6,589 episodes across 1,559 households.

An episode then reads as a chain: the campaign reaches the household, the household shops during the campaign window, some purchases use a campaign coupon, and the redemption is recorded. Each step lives in a different table, and each hand-off is a join the card has to measure.

The episode card, filled from the data

Each field of the card states what to measure, what the data showed and what the result limits. Percentages are computed from the integer counts shown, rounded half up to one decimal place.

The card below is filled from the package data as published; a supplier's own card would use the same fields on its own systems:

The episode card, filled from the data
Card fieldWhat to measureComplete Journey resultWhat it limits
Episode unitThe key that spans systems, and how many episodes existHousehold-campaign pairs: 6,589 episodes, 1,559 households, 27 campaignsEvery rate uses episodes or lines as its denominator
Join coverageShare of rows on each side that find a partner4,836 purchase lines name 17 products missing from the product table; all 2,102 redemptions match a campaign and a couponProduct-level tasks lose a small, named slice
Join cardinalityRows the documented join producesRedemptions joined to coupons on coupon_upc: 2,102 rows become 2,265,375A redemption cannot be labeled with a product from the coupon list
Missing link keyEvents that must be matched by rule because no shared ID existsNo basket ID on redemptions; 1,129 of 2,102 (53.7 percent) tie to exactly one basketCoupon-effect labels rest on a rule-based link
Period alignmentEach system's start and end, and windows that cross themPurchases cover 2017 except 25 December; 932 episodes (14.1 percent) have windows running past that yearWindow outcomes are truncated for those episodes
Derived valuesFields the publisher computed rather than recordedDates rebased onto 2017; timestamp seconds built from basket ID digits; campaign IDs renumbered 1 to 27No sub-minute timing; no rejoin to source campaign IDs
Missing contextContext present per episode, not per tableDemographics for 801 of 2,469 households and 3,835 of 6,589 episodes (58.2 percent)Segment claims cover a heavier-shopping subset
Sample fidelityWhether a published sample keeps episodes wholeThe 75,000-line sample keeps 47 of 45,585 multi-line baskets completeUse full tables for any basket or episode task
Rights by sourceLicense text located for each upstream sourcePackage declares CC0; the upstream source's own terms were not locatedReuse needs the upstream terms, not only the package line

Most links hold; one only looks like a key, and one key is missing altogether. Every redemption's household is listed as receiving that campaign, every redemption's coupon code exists for that campaign, and 6,404 of the 6,589 episodes (97.2 percent) contain at least one purchase inside the campaign window. Purchases join to products almost completely: 4,836 lines (0.3 percent) carry 17 product IDs the product table lacks. The package's prep script drops three corporate-use product groups from the product table without a matching filter on purchases, a plausible cause we could not confirm without the source files.

The coupon code is where documentation and data part ways. The package documents coupon_upc as uniquely identifying each coupon, unique to household and campaign, yet 151 codes appear in more than one campaign and 293 of the 491 redeemed codes were redeemed by more than one household. The coupon table also repeats 4,872 rows exactly. The package's own example joins redemptions to coupons on coupon_upc alone, which turns 2,102 redemptions into 2,265,375 rows; joining on code and campaign still yields 2,169,956, because each coupon row lists one eligible product.

The missing key is the basket. A redemption records household, coupon, campaign and date but no basket, so it must be matched by rule: same household, same New York calendar date, and a coupon-discounted line for a product that coupon covers. Under that rule 1,129 redemptions tie to exactly one basket, 1 ties to two or more and 972 find a same-day basket with no qualifying line. Only 220 of the 792 episodes with a redemption have every redemption tied to one basket. And of 21,889 purchase lines with a coupon discount, which the package documents as a manufacturer coupon, only 1,173 qualify as the matching line for any redemption, so campaign redemptions explain a small share of coupon activity.

A join map of one retail episode: purchases at the center linked to the campaign list, product table, household survey and redemptions, with the coupon list linked to redemptions; each link labeled by its measured coverage or row growth.
Measured links around one household-campaign episode. Source: completejourney package, 2025.

Periods: where each system's clock starts and stops

Each table keeps its own clock, and the card should state all of them. Purchases run from 1 January to 31 December 2017 in New York time, with no line on 25 December; the prep script removes one Christmas Day transaction on the assumption that stores were closed. The year itself is assigned: the script maps day 285 of the source study to 1 January 2017, so calendar effects such as holidays and weekdays describe the package's mapping, not recorded dates.

Campaign windows do not respect that year. Three of the 27 campaigns start before 1 January 2017 and five end after 31 December 2017, so 932 episodes have windows that run past the purchase data, and the prep script keeps only 2017 redemptions, so their out-of-year redemptions are gone as well. Within the year, all 2,102 redemptions fall inside their campaign windows.

Time of day is partly constructed too. For every one of the 1,469,307 purchase lines, the seconds of the timestamp equal the last two digits of the basket ID times 0.6, rounded down, as the prep script's comment about adding a seconds component describes. A card that lists transaction_timestamp as a recorded time would invite sequence analysis at a precision the data does not have.

Missing context and samples that split episodes

Context that exists for some episodes and not others is the gap joins cannot repair. Demographic profiles cover 801 of the 2,469 households (32.4 percent), but those households make 828,850 of the purchase lines (56.4 percent) and 3,835 of the episodes (58.2 percent), so profiled households account for more purchase lines than their share of households. Within the profiles, home ownership is missing for 233 and marital status for 137 after the prep script turns Unknown into a missing value and fills several household composition fields by rule. Stores carry only an ID.

Other gaps sit inside fields. Product category is missing for 540 products, touching 2,209 purchase lines, and package size for 30,586 products. 8,869 lines show a quantity of zero, 49 of them with a positive sales value, and 11,226 lines show zero sales value; the package documents neither case. And 910 households that bought during the year received no campaign at all. The card must say whether that means not targeted or not recorded, and the package does not say.

Samples break episodes in a different way. The bundled transactions sample draws 75,000 lines at random, with seed 8451 in the prep script. It touches 47,243 baskets but keeps every line of only 1,705 of them, and 1,658 of those were single-line baskets to begin with: just 47 of the 45,585 multi-line baskets it touches arrive complete. The promotions sample keeps 2 stores, where 2,705 of 13,952 purchase lines (19.4 percent) have a placement row for their product, store and week. A sample can be fine for line-level statistics and useless for episodes; the card has to say which.

Rights and lineage: one license line is not enough

A multi-system card needs one rights row per source, because a dataset's license line describes its packager, not necessarily its origin. The package's DESCRIPTION file declares CC0 and says the data originates from the 84.51° Complete Journey 2.0 source files on that company's website. A search limited to 8451.com on 28 October 2026 returned no page for the dataset or its terms, so the card records the package license as declared, the upstream terms as not located, and does not infer the second from the first.

Lineage belongs on the same card. The prep script slices one year of source days, removes the Christmas Day transaction, sets positive retail discounts to zero before flipping all discount signs to positive, renumbers the 27 campaigns from 1 to 27 so source campaign IDs cannot be recovered except by their sort order, and recodes demographic categories. None of these steps is wrong. Each changes what a field means, and a buyer comparing this data with the source study, or with a supplier's own export, needs the list.

For licensed business records the same logic holds per system: the CRM, ticketing and billing exports in one dataset can carry different contract terms, retention periods and approvals, and each needs its own row. This page gives general information about documentation, not legal advice on any license.

Does a clean load prove the card is right?

No, and this dataset shows why. Every table loads, every row count matches the package's own dimension tests (6,589 campaign rows, 27 campaign descriptions, 2,102 redemptions, 116,204 coupon rows, 801 profiles, 92,331 products, 360,535 promotion sample rows and 75,000 sample lines), and the full purchase table holds the 1,469,307 lines its README states. Each basket ID maps to one household, one store and one calendar date, and no basket lists a product twice.

A card built from those checks alone would describe a healthy dataset. The episode-level measures tell a different story: a documented join that multiplies rows more than a thousandfold, a link that holds for 53.7 percent of redemptions, windows truncated for 14.1 percent of episodes and a sample that keeps 0.1 percent of multi-line baskets whole. A loadable file and a matching row count prove structure; only episode-level tests prove the relationships a task depends on.

Acceptance checklist for a multi-system card

When the records come from a company's own systems, the same card belongs in the data request and in acceptance. SourceX sources datasets on request; a buyer's request describes scope, volume, format, timing and licensing needs, and adding the episode definition, the joins it depends on and the acceptance test tells the supplier what must survive preparation. Scope and terms are approved and signed by the supplying company before any records move, so the card's rights rows can be settled first.

Accept a multi-system card when each item below is measured, not described:

Checklist · 9 items

  1. The episode is defined as a key spanning systems, and episodes are counted before tables.
  2. Every join reports coverage in both directions, its cardinality and the row count the documented join produces.
  3. Events without a shared key state the matching rule and how many match once, more than once and not at all.
  4. Each system's period is stated, with every episode window that runs past it.
  5. Derived values are marked: rebased dates, constructed times, recoded or imputed fields and renumbered IDs.
  6. Missing context is measured per episode, with the denominator named.
  7. Any bundled sample states whether it keeps episodes whole.
  8. Rights are recorded per source with the license text located, not inherited from the package.
  9. A script that regenerates every number on the card ships with it.

0 of 9 ticked. Tick each item as you go.

Methods note: package, script and measurement rules

Data: the completejourney package at github.com/bradleyboehmke/completejourney, commit 5b5d06192b9856edd04e4d405787af2f2e4a1fef (25 November 2025, package version 1.1.1), read on 28 October 2026. Files: data/*.rda and data/transactions.rds, parsed with the Python rdata package (version 1.1.0); transactions.rds has SHA-256 1fa0700033f1e5d9bb6b09e2be063d8d68474d346e95c50f2833e09d083e0007. Script and output: bin/xs0947_cj_data_card.py, bin/xs0947_output.txt and bin/xs0947_summary.json in this batch's scratch folder (counts only). No package code was run. This is the package's cleaned copy, not 84.51°'s source files, which we did not read.

Rules as applied: IDs are compared as exact strings after trimming whitespace. R dates are days since 1970-01-01; purchase timestamps are seconds since 1970-01-01 UTC converted to America/New_York, the zone the prep script uses, before taking calendar dates and seconds. An episode is one row of the campaigns table; its window is the campaign's start and end dates, inclusive. A purchase inside a window is a line of the same household dated within it. A redemption's candidate basket is a same-household basket on the redemption date holding a line with a coupon discount above zero for a product listed in coupons under the same coupon code and campaign. Repeated coupon rows are rows identical in all three columns, counted beyond the first; a code in more than one campaign has two or more distinct campaign IDs in the coupon table; the documented-join counts are left joins of redemptions to coupons on coupon_upc, then on coupon_upc and campaign_id. Demographic coverage counts households, lines and episodes whose household ID appears in the demographics table. Sample completeness compares sampled lines per basket with that basket's lines in the full table. Promotion coverage counts purchase lines in the 2 sampled stores whose product, store and week appear in the promotions sample.

Not measured: the full 20,940,529-row promotions table, store attributes (none are published), the accuracy of any value, and the upstream source's terms. Percentages are computed from integer counts and rounded half up to one decimal. Release check: rerun on 84.51°'s own files if they become available, and confirm the upstream terms.

Frequently asked questions

What is a multi-system dataset card?

It is a dataset card whose unit is an episode that spans several source systems, such as a campaign, a purchase and a redemption, rather than one table. It measures the joins, periods, missing context, derived fields, samples and rights the episode depends on, and ships with the script that produces each number.

How is it different from a standard dataset card?

A standard card describes sources, collection, intended use and known limits, usually one table or source at a time. A multi-system card adds what lives between tables: key coverage and cardinality, matching rules where no key exists, each system's period and per-episode context, because those decide whether a joined task is possible.

Which join metrics belong on the card?

Coverage in both directions, cardinality and the row count the documented join produces, plus, for events without a shared key, how many match once, more than once and not at all. In the Complete Journey package the documented coupon join turns 2,102 redemptions into 2,265,375 rows, which no coverage rate would reveal.

Why do periods need their own field?

Because each system starts and stops on its own schedule, and episodes that cross a boundary are silently truncated. In the test dataset, 932 of 6,589 household-campaign episodes have windows running past the purchase year, and the dates themselves were assigned by the packager rather than recorded.

Can a sample stand in for the full dataset on the card?

Only for the unit it preserves. A random line sample keeps line statistics but splits baskets and episodes: the Complete Journey sample of 75,000 lines keeps 47 of 45,585 multi-line baskets whole. The card should name the sampling unit and test whether episodes survive it.

Sources

  • The completejourney DESCRIPTION file declares version 1.1.1 and a CC0 license and says the data originates from the 84.51° Complete Journey 2.0 source files. Source
  • The completejourney README describes grocery transactions over one year for 2,469 households, demographics and direct-marketing history for certain households, and a full transactions table of 1,469,307 rows. Source
  • The package's prep script slices source days 285 to 649 onto 2017, removes a Christmas Day transaction, builds timestamp seconds from basket IDs, zeroes positive retail discounts, renumbers campaigns 1 to 27, recodes demographics and draws the samples with seed 8451. Source
  • The package's dimension tests expect 6,589 campaign rows, 27 campaign descriptions, 2,102 redemptions, 116,204 coupon rows, 801 demographic rows, 92,331 products, 360,535 promotion sample rows and 75,000 sample transactions. Source
  • The package documents coupon_upc as uniquely identifying each coupon, unique to household and campaign, and shows an example join of redemptions to coupons on coupon_upc. Source
  • The package documents coupon_disc as the discount applied due to a manufacturer coupon and basket_id as identifying a purchase occasion. Source
  • Pushkarna, Zaldivar and Kjartansson (Google Research, FAccT 2022) describe Data Cards as summaries covering upstream sources, collection and annotation methods, intended use and decisions affecting model performance. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies