-
Type:
Task
-
Resolution: Unresolved
-
Priority:
Major - P3
-
None
-
Affects Version/s: None
-
Component/s: None
-
Product Performance
-
None
-
None
-
None
-
None
-
None
-
None
-
None
SERVER-137897 found one workload shape (FindFilterSort in top_ten_queries_locust) where monguard costs more than noise with socket handoff. It measured these latencies:
| Arm | p95 | p50 | avg |
|---|---|---|---|
| No monguard | 3.01ms | 1.81ms | 2.33ms |
| Monguard proxy | 5.91ms | 2.71ms | 3.70ms |
| Monguard handoff | 4.89ms | 2.24ms | 3.08ms |
Handoff recovers part of the proxy regression. Handoff p95 latency is 1.9ms (~60%) higher than no-monguard (n=1). The original SERVER-137897 report calculated the regression as +37% p95 (using the stable region). Neither baseline is solid: the no-monguard numbers come from a single waterfall run, and the stable regions had drifted. The handoff vs proxy comparison (6v6) is solid (handoff p95 is 14.5% lower than proxy).
We need to investigate this further for broad-rollout performance sign-off. This can start now on the SERVER-137897 setup.
Out of scope
Fixing the regression (a follow-up ticket if needed).
Wider performance characterization will happen in SERVER-138610.