Twenty-two tables, four layers
0:07
Twenty-two tables, four layers. From exactly as it arrived to shaped for a question.
0:12
The catalogThe catalog lists every governed table at Acme Home Appliances, a demo company, with who owns it, whether its checks pass and when it last refreshed.
0:24
Twenty-two tablesTwenty-two governed tables, eight of them certified gold.
0:28
ScaleTwo hundred and seventy-eight million rows, about sixty-one gigabytes, in open format in the company's own cloud.
0:36
Need attentionNine need attention. Three are failing checks, and six have gone stale.
0:43
Four layersEvery table sits in one of four layers, and the layer is a promise about how far it can be trusted.
0:53
RawRaw is exactly as it arrived. Nothing is cleaned and nothing is promised.
1:01
StagingStaging is typed and deduplicated: safe to build on, not to report from.
1:07
CoreCore holds the conformed facts and dimensions: one definition of a distributor, one of an invoice.
1:17
MartAnd a mart is shaped for a question somebody actually asks. Dashboards read marts.
1:25
ChecksChecks shows plain counts when everything passes, and a tag only when something fails.
1:32
FreshnessFreshness has no colour under a day old. Older than that, it warms up, and past a week it turns solid.
Reading a table
1:42
Reading a table. Failing, stale, healthy, and not checked at all.
1:47
FailingThe cost-per-lead mart is failing three of seven checks. The panel leads with what that means: anything reading it reports numbers nobody is standing behind.
1:58
DowngradedIt was gold until this morning. The analytics engineer downgraded it to silver, and that change is in the audit trail.
2:09
StaleThe channel stock fact passes all eleven checks and has not landed for nine days. Correct and old is the harder problem to notice.
2:19
WhenLast refreshed nine days ago, on 30 July.
2:27
HealthyThe billing fact is what healthy looks like: fourteen of fourteen checks, refreshed eight minutes ago, gold certified, owned by the data platform team.
2:38
SizeThirty-eight million invoice lines, nine gigabytes on disk.
2:45
StagingThe service CRM extract passes its checks but is a day and a half old, because the CRM feed behind it is rate-limited.
2:58
UncheckedThe call events landed last week and have no checks yet. The catalog shows that as a gap, not as a pass.
3:10
FadedAnd the paused price tracker's table fades: unchecked and weeks old, on purpose.
Gold is not current
3:17
Gold is not current. Certification and freshness are different questions.
3:23
By certificationSorted by certification, the eight gold tables come first.
3:28
What gold meansGold means reviewed, monitored and safe to run the business on. Silver was reviewed once. None means nobody has vouched for it.
3:38
Gold and staleBut three gold tables are stale. The service SLA mart is a day and a half behind the CRM.
3:46
Gold and staleSo is the service tickets fact it is built from.
3:52
Gold and nine days oldAnd the channel stock fact is gold and nine days old.
3:58
Two questionsA certificate says the definition was checked. Freshness says the data arrived. The catalog shows both on the same row, so neither can vouch for the other.
4:14
A dimensionThe distributor master is small, 470 rows, and gold. Every rollup by region depends on it.
Finding a table
4:22
Finding a table. Search by name, owner or state, and sort by anything.
4:30
By subjectSearch reads names, descriptions, layers, owners and states. Service finds five tables, and three of them are stale.
4:40
By ownerBy owner, Finance owns one table: receivables, gold and healthy.
4:48
By stateBy state, the three failing tables: the cost-per-lead mart and both customer feedback tables.
4:58
Oldest firstSorted by freshness, oldest first: the paused price tracker, then the three channel stock tables, all from 30 July.
5:09
Largest firstBy size, largest first, the web sessions export leads at fourteen gigabytes, unmodelled and read by one mart.
Back to the cause
5:19
Back to the cause. Every table names the connector that feeds it.
5:24
A failing factThe customer feedback fact is failing three of nine checks. Leadership reads it as customer satisfaction.
5:31
Fed byFed by names the connector behind it: the customer feedback drop, broken for four days. The table is waiting on it.
5:41
Query itQuery it gives the exact name to use from any SQL tool. The data sits in open format, so the company's own tools can read it directly.
5:55
The sourceOne click lands on the source: a file drop whose name pattern stopped matching when the agency renamed its files.
6:10
The whole chainLineage draws the same chain the other way: from the survey drop, through both feedback tables, to the models, dashboards and alerts that read them.
Where it shows up
6:21
Where it shows up. The catalog decides what a dashboard is allowed to claim.
6:27
In IntelligenceIn Antvia Intelligence, the Business pulse dashboard reads the same catalog. Two of its tiles are not showing current data, and the dashboard says so above them.
6:39
UnverifiedCustomer satisfaction is marked unverified, and the tile names the failing checks on the feedback table.
6:47
StaleChannel stock is marked nine days old, with the table and the date it last refreshed.
6:54
Still shownThe numbers stay on screen. Hiding them would be worse. What changes is that nobody can read them as current by accident.
7:06
Across every dashboardThe dashboard register counts it across the board: eight dashboards are fed by at least one table the catalog does not stand behind.