Monthly Data Latency & Ingestion Health Dashboard
[OVERVIEW]
A saved CX Enterprise Coworker skill that produces a monthly health check on your Adobe Experience Platform data, how fresh it is and how reliably it's arriving, for any set of datasets you name. One request generates interactive dashboards in chat plus a downloadable, Adobe-branded PDF. It knows the difference between event/streaming datasets (like Web SDK and Mobile SDK) and profile datasets, and reports the right metrics for each.
[TRYABLE]
1. Ask Coworker to build a monthly data-health/latency dashboard for your datasets, and refine it until the dashboards and PDF look right.
2. Say "save this as a skill." Confirm how you want to trigger it, what it outputs by default, and its default sandbox and month.
3. Review the draft and publish, it's then available in every future session.
4. Run the quick test to confirm it works, then just ask for the dashboard each month (naming the month or datasets when you want to override the defaults).
[SETUP]
- CX Enterprise Coworker with skill creation enabled.
- Adobe Experience Platform sandbox access, with the datasets you want to monitor actively ingesting data (e.g., Web SDK and Mobile SDK event datasets, plus profile datasets).
- Access to ingestion/batch details and Query Service (for event-data freshness sampling).
- The exact names of the datasets you want covered.
[IMPACT]
The goal is simple: ask one question and get a clear, trustworthy picture of whether your data is showing up on time and without problems, with no manual digging across multiple tools. Each run gives you a dashboard for your event data, a comparison table for your profile data, a polished PDF you can share with stakeholders, and a plain-language verdict on what’s healthy and what needs a look. Because you can run it month after month, it’s easy to spot when something changes in one real run, comparing two months is exactly what revealed a data feed that had quietly started lagging.
[LESSONS]
- "Typical" and "worst case" beat "average." The report leads with the typical experience (p50) and the near-worst-case (p95) instead of an average, because a few slow outliers can make an average misleading. When those two numbers are close, things are steady; when they're far apart, something is occasionally slow.
- Trust the data's arrival time, not its claimed event time. Device clocks are often wrong, so the skill measures freshness from batches it knows landed cleanly, and reports how much of the data had believable timestamps.
- Not every "failure" is a problem. A batch that "fails" but loses no records is usually just an empty or skipped send, not lost data. The skill calls that out so you don't chase a false alarm.
- A steady, identical delay is a clue. If every batch is late by about the same amount, that usually points to a backfill or delayed feed rather than normal variation and is worth investigating.
- Freshness only applies to event data. Profile datasets don't carry an event time, so for those the skill focuses on reliability and data quality instead, and says why.
