Automated canary analysis observes several time windows rather than trusting one small sample. Promotion requires sustained health, while one conclusive regression stops the rollout.
Design a CanaryWindowAnalyzer:
- The constructor receives
minRequests, an allowed error-rate increase in percentage points, an allowed p95 latency increase in percent, and the required number of consecutive healthy windows. evaluate(...) returns "promote", "rollback", or "continue".decisionWindow(...) returns the zero-based window that produced promotion or rollback, or -1 for continue.
A window is eligible only when both versions have at least minRequests. An ineligible window resets the healthy streak.
An eligible window is unhealthy when either:
or:
Equality is healthy. Roll back immediately on the first unhealthy eligible window. Otherwise promote at the first required consecutive healthy window and ignore later windows.
Example 1:
Input:
Output:
Explanation: Two consecutive eligible windows stay within both thresholds.
Example 2:
Input:
Output:
Explanation: The second window exceeds the error threshold, so the rollout stops before it can collect three healthy windows.
Constraints
- All six input arrays have equal length, at most
10^4. - Request totals and p95 values are positive integers.
- Error counts are between zero and their request totals.
- Thresholds are non-negative integers.
1 <= requiredHealthyWindows <= 100- Cross-products fit signed 64-bit integers.
- At most
50 total method calls are made.
Analyze Canary Health Windowsmedium