Problem

SlowNotificationHeuristics.hasRepeatedFailedServiceStarts() decides whether a user is failing to start the FCM service too often by comparing the failure ratio against serviceStartFailurePercentage (a Float, default 0.5). However, the ratio is computed with integer division:

if (failures.size / (failures.size + successes.size) >= failurePercentage) {

Both failures.size and successes.size are Int, so failures.size / (failures.size + successes.size) is integer division. Because failures.size is always <= the total, this evaluates to:

• 0 whenever there is at least one success (i.e. any mixed ratio), and
• 1 only when there are zero successes (100% failures).

Comparing that integer result against 0.5f means the branch only ever passes at a 100% failure rate, never at the configured 50% (or any other fractional threshold). In practice the failed-service-start signal of the delayed-notifications heuristic is dead for the realistic case of a user who fails, say, 50–99% of service starts.

The sibling heuristics (isFailingToDrainQueue, hasLongMessageLatency) compare raw counts/latencies, so they are unaffected — this is the only place a fractional rate is computed.

Fix

Promote the numerator to Float so the comparison uses the intended fractional failure rate:

-    if (failures.size / (failures.size + successes.size) >= failurePercentage) {
+    if (failures.size.toFloat() / (failures.size + successes.size) >= failurePercentage) {

Division by zero is not a concern here: an earlier guard returns when (successes.size + failures.size) < minimumEventCount (default 10).

Notes

One-line change, no behavior change for the 100%-failure case; it now also correctly triggers for the intended fractional thresholds.