Definition
Replica Consistency Divergence Score (RCDS) measures the actual data divergence between primary and replica at the most recent consistency checkpoint, normalized against the RPO target.
The Measurement Gap
A B2B SaaS company runs a regulated customer data platform with strict RPO targets. The replication strategy uses asynchronous replication to a secondary region with a documented RPO of 5 minutes. The replication monitoring dashboard reports lag in milliseconds during normal operation, well within the RPO budget. The team has not had to fail over in 18 months. The replication posture is reported as healthy.
Then a sustained primary-region degradation forces a failover at peak load. The failover succeeds technically: the secondary region takes over within the RTO. The compliance team checks RPO compliance and finds that the actual data lost at failover was 14 minutes of transactions, not the 5 minutes the dashboard had been reporting. Investigation reveals that during the high-load period leading up to the failover, replication had been falling progressively further behind. The lag metric the dashboard showed was a point-in-time measurement that did not capture the cumulative divergence under sustained load. The replication appeared healthy at every individual measurement and was unhealthy in aggregate.
The customer data platform is regulated under data residency and continuity requirements. The 14-minute data loss exceeds the SLA committed to customers. The post-incident review highlights that the AI-optimized replica placement, which had been tuned for cost efficiency by moving replicas to lower-cost zones, had reduced inter-zone bandwidth between primary and replica. The configuration met the RPO under normal load and degraded under the peak load that the optimizer had not modeled.
The legacy DR metrics on the team's dashboard report a healthy replication posture.
Replication lag (current): milliseconds.
Replication health: green.
RPO compliance at last test: 100%.
Replica availability: 100%.
Last failover test: passed within RPO.
Replication throughput: meeting required rate at normal load.
None of these report on the divergence under sustained operational conditions. The current-lag metric is a snapshot. The compliance check at the last test is point-in-time. The throughput at normal load does not reveal the optimizer's hidden tradeoffs that surface under peak load.
These metrics were designed for synchronous replication or for asynchronous replication where the team explicitly managed bandwidth and prioritization. With AI-optimized replica placement and AI-tuned replication parameters, the bandwidth and prioritization choices are made automatically based on cost and average-case performance. The divergence behavior under load is not in the optimizer's objective function and not in the monitoring.
There is no metric that tells you whether AI-optimized replica placement maintains data consistency under actual failover conditions, which means RPO targets are met on paper while real data divergence at failover is unknown.
Formula
Replica Consistency Divergence Score (RCDS) measures the actual data divergence between primary and replica at the most recent consistency checkpoint, normalized against the RPO target.
In plain language: at each consistency checkpoint, measure the time difference between the latest transaction committed on primary and the latest transaction confirmed on replica. Divide by the RPO target. A value near 0 means the replica is essentially in sync. A value approaching 1 means you are right at the RPO limit. A value above 1 means you have already violated RPO.
The formula:
RCDS = (T_primary_latest − T_replica_latest) / RPO_target
Where:
- T_primary_latest = timestamp of the most recent committed transaction on primary
- T_replica_latest = timestamp of the most recent confirmed transaction on replica
- RPO_target = the stated recovery point objective time duration
Units are a dimensionless ratio. 0 means perfect synchronization. Below 1.0 means within RPO. Above 1.0 means RPO already violated. The metric should be reported as a continuous time series, with both the current value and the trailing 24-hour distribution.
A worked example. The B2B SaaS company instruments RCDS on their customer data platform. The RPO target is 5 minutes (300 seconds). Measurements over a 24-hour period:
- Normal off-peak hours: T_primary − T_replica = 0.4 seconds, RCDS = 0.001 (well within RPO)
- Normal business hours: T_primary − T_replica = 2.1 seconds, RCDS = 0.007
- Mid-day load spike: T_primary − T_replica = 38 seconds, RCDS = 0.127
- Sustained load: T_primary − T_replica = 95 seconds, RCDS = 0.317
- Peak: T_primary − T_replica = 167 seconds, RCDS = 0.557
- Continued peak: T_primary − T_replica = 245 seconds, RCDS = 0.817 (approaching RPO threshold)
The RCDS time series shows the replica progressively falling behind as load increases. At 0.817, the replica is within RPO but uncomfortably close. Per the catalog guidance (below 0.5 comfortable margin, above 0.8 dangerously close, above 1.0 RPO violated) this is the territory that should trigger an alert to the DR team. If load continued, the next measurement would likely cross 1.0.
In the actual failover scenario described in Section 1, the divergence reached 14 minutes (840 seconds, RCDS = 2.8) during peak load before the team failed over. The dashboard had been showing point-in-time lag in milliseconds because the measurement was taking instantaneous snapshots rather than tracking the trailing divergence.
Three edge cases need explicit treatment. First, in multi-region or multi-replica deployments, compute RCDS per replica and use the worst score as the operationally relevant metric, a single replica with high RCDS is the recovery posture for any failover that lands on that replica. Second, measure during sustained load, not just idle: RCDS during idle is trivially low and uninformative. Third, for inherently asynchronous replication, establish a non-zero baseline RCDS during normal operation; spikes above baseline are what matters, not the absolute value.
Cross-references. Upstream of RCDS is
Multi-Cloud Portability Erosion Index (MCPEI): AI optimization decisions that lock workloads into cost-optimized configurations often have replication-related consequences that RCDS surfaces. Downstream is Failover Necessity Rate (FNR). If RCDS is high, the failover that does occur will involve more data loss, raising the operational cost of every failover event.
Proposed formula, pending validation pilots.