DOWNSTREAM IMPACT PROPAGATION INDEX (DIPI)

GMI-IT-041IT OperationsInfrastructure-Dependent

Definition

Downstream Impact Propagation Index (DIPI) measures the share of consumer-reported data incidents that trace back to upstream pipelines whose quality score was above the halt threshold at the time the bad data was produced.

The Measurement Gap

A B2B SaaS company. The data platform team feeds 12 product services and 8 analytics workspaces. The data quality team set the halt threshold at 0.85: pipelines below that score get paused, pipelines above get pushed forward. The platform dashboard shows 96% of pipelines running above the threshold any given week. By the standard the team chose, the platform is healthy.

Then the customer success leader sends an escalation summary to the head of data: 14 customer escalations this month tied to incorrect usage analytics, totaling roughly 80 person-hours of customer success time and two churn events. "The numbers we show customers don't match what they see in their own logs." The data team digs in. Eleven of the 14 escalations trace back to upstream pipelines that scored between 0.85 and 0.92 the day the bad data was produced. Above threshold. Pushed forward. Degraded but not enough to trigger a halt.

The team had tuned the threshold to keep things flowing, because pausing pipelines created too many false-stop incidents. The result is a system where source-level quality monitoring runs green and consumer-level impact runs red. The decisions about what counts as "good enough" data live at the source. The cost lives at the consumer. The two are disconnected.

The legacy metrics do not span the gap.

Pipeline success rate

measures execution at the source: every pipeline ran, every pipeline completed.

Data quality score average

smooths out the consequential tail; a pipeline running 0.87 average can produce a single 0.85 day that drives a customer escalation, and the average never moves.

Incident count by source pipeline

tracks problems at the producer, not impact at the consumer.

SLA compliance

tracks uptime against availability targets, not correctness against business expectation.

Consumer ticket volume

tracks how many people complained, not whether those complaints trace back to source quality decisions.

MTTR for data incidents

tracks recovery time, not whether the right pipelines got the threshold adjustment they needed.

These metrics were not designed for AI-driven quality monitoring. They were designed for an era when the threshold for pushing data forward was either binary (pass / fail) or set conservatively by humans who would be paged when things broke. The thresholds were tunable, but the consequences of tuning them were visible: somebody got paged. With AI quality monitoring tuning thresholds dynamically and pushing data through automatically, the consequences moved from the on-call rotation to the customer's experience, and nobody is paged for that.

There is no metric that tells you what happened to downstream consumers when a pipeline pushed degraded-but-above-threshold data forward, which means data quality decisions at the source are disconnected from the impact those decisions create.

Formula

Downstream Impact Propagation Index (DIPI) measures the share of consumer-reported data incidents that trace back to upstream pipelines whose quality score was above the halt threshold at the time the bad data was produced. In plain language: of all the data problems your downstream consumers report, how many came from upstream pipelines that you decided were "good enough" to push forward? A high number means the halt threshold is too permissive. A low number means most consumer incidents trace to issues you correctly flagged at the source.

The formula:
DIPI = Consumer_incidents_from_passed_data / Total_consumer_data_incidents

Where:

  • Consumer_incidents_from_passed_data = data-related incidents reported by downstream consumers whose root cause attribution points to a source pipeline that was above the halt threshold at the time the bad data was produced
  • Total_consumer_data_incidents = all data-related incidents reported by downstream consumers during the measurement periodUnits are a percentage from 0 to 100. Zero means no consumer incidents trace to passed-through data: everything that escaped detection was either un-monitorable or below threshold. One hundred percent means the halt threshold is providing no protection at all. A worked example. The SaaS company from the opening scenario runs a quarterly DIPI measurement. Q3 numbers:

  • 47 consumer-reported data incidents during the quarter

  • Root cause attribution by the data engineering team: - 22 traced to upstream pipelines that were above the 0.85 halt threshold at the relevant timestamp - 18 traced to upstream pipelines that were paused below threshold (the issue was caught at source, but consumers had built downstream processes that broke when the expected data stopped arriving) - 7 traced to consumer-side bugs (downstream transformation errors, not source data quality)
    DIPI = 22 / 47 = 0.468 = 46.8%
    Per the proposed starting threshold (above 40% means the halt threshold is too permissive) this number triggers action. The team should lower the halt threshold on the worst-offending source pipelines, accepting some increase in halt-related downstream disruption in exchange for fewer consumer-facing quality incidents. Two edge cases require explicit handling. First, root cause attribution requires lineage. If the data engineer working an incident does not have tooling to trace from the consumer-side problem back to a specific upstream pipeline and that pipeline's quality score at the relevant timestamp, DIPI is not computable. Second, when bad data comes from multiple upstream sources, attribute proportionally based on contribution to the bad output, not all-or-nothing to one source. The minimum sample for a meaningful DIPI calculation is 20 consumer-reported data incidents per quarter. Lower volumes produce a noisy percentage. Cross-references. Upstream of DIPI is Semantic Data Quality Gap (SDQG). If SDQG is high, DIPI is almost certainly elevated, because the same semantic blind spots that pass through the quality monitor reach consumers. Downstream is

Cost Per Anomaly (CPA, GMI-IT-009): consumer incidents that DIPI counts drive remediation cost that CPA quantifies. The misinterpretation scenario is the opposite of the SDQG case. A pipeline reports SDQG = 0.50 (high: semantic blindness is bad) but DIPI is only 15% (low). Read in isolation this looks like SDQG is overstating the problem. The reality is usually that consumers are not yet using the semantically-broken data, or are filtering for it downstream, or have not noticed yet. SDQG is a leading indicator. DIPI is a lagging indicator. SDQG can be high before DIPI catches up. Watch both, and watch the gap between them, a widening gap suggests consumer impact is building up and will surface later.

Proposed formula, pending validation pilots.

At a Glance

Category
AI-Era IT Operations
← DEPRECATION IMPACT UNDERESTIMATION RATE (DIUR)DRIFT CLASSIFICATION ACCURACY (DCA) →

Assess Your Organization

The AI-GPS Assessment evaluates your AI governance maturity and recommends KPIs like this one for your implementation roadmap.

Take the Free AI-GPS Assessment