Definition
Stale Dependency Inflation Rate (SDIR) measures the share of auto-discovered dependencies that are no longer active.
In plain language: take a sample of dependencies from the map.
The Measurement Gap
A platform engineering team relies on the auto-discovered dependency map for change risk assessment. The map shows the change risk scorer thousands of edges. The scorer uses the map to estimate the blast radius of proposed changes, and changes with high estimated blast radius get more scrutiny.
Then a routine change to a customer onboarding service gets flagged as high risk because the dependency map shows the change affecting 27 downstream services. The team's senior engineer pushes back. "There's no way this change affects 27 services. This service was substantially refactored 18 months ago. Half of those dependencies don't exist anymore." Investigation reveals that the engineer is right. The dependency map shows dependencies that were valid before the refactor and have not been pruned since. Of the 27 dependencies, only 11 are actually active. The 16 phantom dependencies inflated the change risk score and triggered review attention the change did not actually warrant.
The team broadens the investigation. They sample 100 dependencies from the map and check whether they are still active. 31 of the 100 have had no traffic in the past 90 days. Some are explicable (DR-only paths, annual compliance integrations) but 22 of the 31 represent genuinely dead dependencies: connections to services that have been retired, refactored, or restructured. The map is inflated by 22%, and the inflation is affecting every change risk assessment that uses the map.
The legacy dependency mapping metrics report a healthy state.
Dependency count: high (inflated by stale entries).
Dependency map size: stable or growing.
Dependency confidence: tracked, mostly high.
Map updates: continuous.
None of these report on whether the existing dependencies are still real. The auto-discovery typically adds dependencies when it observes traffic and updates rarely or never when traffic stops. The pruning side of map maintenance has not been built or has not been enforced.
These metrics were designed when dependency maps were curated and architects removed obsolete dependencies as part of map maintenance. With auto-discovery, the addition is automated and the pruning is not, so the map accumulates stale entries over time.
There is no metric that tells you how many discovered dependencies are stale: no longer active but still inflating change risk assessments, which means the map is "conservative" in ways that waste engineering time.
Formula
Stale Dependency Inflation Rate (SDIR) measures the share of auto-discovered dependencies that are no longer active. In plain language: take a sample of dependencies from the map. For each, check whether there has been any traffic on that dependency in a verification window (recommend 90 days). The percentage with no traffic is SDIR. Exclude known-infrequent dependencies (DR paths, annual compliance integrations): those are dormant but not stale.
The formula:
SDIR = Dependencies_stale / Dependencies_sampled × 100
Where:
- Dependencies_stale = sampled dependencies with no traffic in the verification window
- Dependencies_sampled = total dependencies reviewed (excluding known-infrequent dependencies)
The recommended verification window is 90 days. Units are a percentage from 0 to 100. 0% means all sampled dependencies are active. Higher values mean more phantom connections. A worked example. The team runs SDIR on a sample of 50 dependencies from the customer onboarding service domain:
- Dependencies sampled: 50
- Known-infrequent dependencies (excluded from calculation): 7
- Remaining dependencies analyzed: 43
- Dependencies with no traffic in 90-day verification window: 14
SDIR = 14 / 43 = 32.6%
Per the catalog guidance (below 15% mostly clean, above 30% significant phantom dependency waste) this SDIR of 32.6% is above the action threshold. The team needs to implement pruning logic: any dependency with no traffic in 180 days (excluding documented exceptions) should be flagged for review and removed if confirmed obsolete. Three edge cases need explicit treatment. First, exclude known-infrequent dependencies (DR paths, annual compliance, emergency-only paths): they are dormant by design, not stale. Without exclusion, the metric conflates two different problems. Second, high SDIR combined with low DDCR is particularly bad: the map is missing real dependencies while including phantom ones, so both kinds of decisions based on the map are wrong. Track DDCR and SDIR together; the combination is more informative than either alone. Third, the minimum sample is 50 dependencies per assessment. Cross-references. Upstream of SDIR is
Dependency Discovery Coverage Rate (DDCR): high SDIR and low DDCR together indicate a fundamentally unreliable dependency map. Downstream is Change Risk Scoring Accuracy (CRSA, GMI-IT-012) and
Blast Radius Prediction Accuracy (BRPA, GMI-IT-065): phantom dependencies inflate the predictions both metrics depend on. The misinterpretation to watch for: a team reports SDIR = 10% (low) and concludes their dependency map is clean. They miss that the sample excluded a large portion of dependencies as "known-infrequent" without rigorous criteria for the exclusion, hiding stale dependencies in the exclusion category. Document the known-infrequent exclusion criteria explicitly and audit the excluded set periodically to ensure dormant-by-design and dormant-because-stale are distinguished correctly.
Proposed formula, pending validation pilots.