Add DR drill test results and analysis charts

Charts and report from live DR drill validation of the background
refresh fix on account bgrefresh-mw-test-440 (East US + West US).

3 drills over 51 minutes, 103 background refreshes, 305 reads, 0 errors:
- Drill 1: MW→SW→MW transition (both detected)
- Drill 2: Region removal + re-add (detected, DIRECT mode connections persist)
- Drill 3: Failover priority change (fastest propagation ~1 min)

Config: Global endpoint, preferred=West US, DIRECT mode, 30s refresh
(overridden via COSMOS.UNAVAILABLE_LOCATIONS_EXPIRATION_TIME_IN_SECONDS).
Unpatched SDK: 0 refreshes (timer stops after init).

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: jeet1995

sdk/cosmos/azure-cosmos/test-results/background-refresh-dr-drill/dr-drill-report.md

@@ -0,0 +1,155 @@
+# DR Drill Report — Background Refresh Fix Validation
+
+**Account**: `bgrefresh-mw-test-440`
+**Resource Group**: `rg-bgrefresh-test`
+**Tenant**: `022026052026cdbprod.onmicrosoft.com` (ephemeral)
+**Regions**: East US (hub) + West US
+**SDK Version**: 4.80.0-beta.1 (patched with background refresh fix)
+**Client Config**: Global endpoint, preferred region = West US, DIRECT mode, 30s background refresh
+**Drill Duration**: 13:05 – 13:56 UTC (51 minutes)
+**Date**: 2026-04-10
+
+---
+
+## Executive Summary
+
+| Metric | Value |
+|--------|-------|
+| Background refreshes | **103** (every ~30s, continuous) |
+| Total point reads | **305** |
+| Errors | **0** |
+| Metadata transitions detected | **9** |
+| Read availability | **100%** |
+
+The patched SDK's background refresh timer ran continuously through all 3 drills, detecting every topology change. Zero read failures occurred. This validates the fix in [PR #48758](https://github.com/Azure/azure-sdk-for-java/pull/48758).
+
+**Contrast with unpatched SDK**: The unpatched 4.80.0-beta.1 stopped the background refresh timer after init (0 refreshes), and topology changes were never detected.
+
+---
+
+## Latency Summary
+
+| Scope | Count | Avg | P50 | P90 | P99 | Max |
+|-------|-------|-----|-----|-----|-----|-----|
+| **All reads** | 305 | 165ms | 145ms | 207ms | 687ms | 2007ms |
+| Drill 1 (MW↔SW) | 77 | 174ms | 145ms | — | 725ms | 725ms |
+| Drill 2 (Region removal) | 141 | 166ms | 142ms | — | 819ms | 2007ms |
+| Drill 3 (Failover change) | 87 | 156ms | 157ms | — | 270ms | 270ms |
+
+**Notable latency spikes**:
+- 13:05 (init): 725ms — cold start, first TCP connections to West US replicas
+- 13:24 (West US removed): 819ms spike, then bimodal 40-100ms (gateway rerouting) / 140ms (direct)
+- 13:36 (West US re-added): **2007ms** — highest spike as SDK re-establishes connections to newly available region
+
+---
+
+## Drill 1: MW → SW → MW Transition
+
+**Purpose**: Validate that the background refresh timer continues running in multi-write accounts and detects `enableMultipleWriteLocations` changes.
+
+| Time (UTC) | Control Plane Action | SDK Detection (via background refresh) | Lag |
+|---|---|---|---|
+| 13:07:11 | MW DISABLED | — | — |
+| 13:10:02 | — | ✅ `MW=false W=1 R=2` | ~3 min |
+| 13:13:15 | MW RE-ENABLED | — | — |
+| 13:14:56 | — | ✅ `MW=true W=2 R=2` | ~2 min |
+
+**Verdict**: ✅ **PASS** — Both transitions detected. `writableLocations` changed from `[EUS, WUS]` → `[EUS]` → `[EUS, WUS]`. This is the exact scenario that FAILS on the unpatched SDK (timer stops, transitions never detected).
+
+```
+13:05-13:10  MW=true  W=2 R=2   reads: West US ~145ms
+                                 ┌─ MW DISABLED at 13:07
+13:10-13:15  MW=false W=1 R=2   reads: West US ~145ms  ← SDK detected!
+                                 ┌─ MW RE-ENABLED at 13:13
+13:15-13:18  MW=true  W=2 R=2   reads: West US ~145ms  ← SDK detected!
+```
+
+---
+
+## Drill 2: Region Removal + Re-add
+
+**Purpose**: Validate SDK detection of region topology changes (readableLocations/writableLocations shrink and expand).
+
+| Time (UTC) | Control Plane Action | SDK Detection | Lag |
+|---|---|---|---|
+| 13:18:15 | Remove West US (started) | — | — |
+| 13:29:22 | Remove West US (completed) | — | — |
+| 13:24:15 | — | ✅ `MW=true W=1 R=1` | ~6 min from start |
+| 13:36:26 | West US re-added + MW enabled | — | — |
+| 13:36:16 | — | ✅ `MW=true W=2 R=2` | immediate |
+
+**Verdict**: ✅ **PASS** — Region removal and re-addition both detected.
+
+**DIRECT mode observation**: Even after West US was removed from the account metadata (`R=1`), reads continued on West US via existing RNTBD connections. This is expected — DIRECT mode maintains active TCP sockets to replicas, which keep serving until decommissioned. Latency shifted to bimodal pattern (~40-100ms vs ~140ms), suggesting some gateway-level rerouting.
+
+```
+13:18-13:24  MW=true  W=2 R=2   reads: West US ~145ms
+                                 ┌─ West US REMOVED
+13:24-13:36  MW=true  W=1 R=1   reads: West US ~95ms (RNTBD connections alive)
+                                 ┌─ West US RE-ADDED  (2007ms spike!)
+13:36-13:42  MW=true  W=2 R=2   reads: West US ~160ms (connections re-established)
+```
+
+---
+
+## Drill 3: Failover Priority Change
+
+**Purpose**: Validate SDK detection of write region swap via `failover-priority-change`.
+
+| Time (UTC) | Control Plane Action | SDK Detection | Lag |
+|---|---|---|---|
+| 13:43:06 | MW disabled | — | — |
+| 13:44:09 | Write swapped EUS→WUS | — | — |
+| 13:45:06 | — | ✅ `MW=false W=1 R=2` | ~2 min |
+| 13:49:09 | Write swapped WUS→EUS | — | — |
+| 13:49:31 | — | ✅ `MW=false W=2 R=2` (transitional) | <1 min |
+| 13:50:01 | — | ✅ `MW=false W=1 R=2` (settled) | ~1 min |
+| 13:51:16 | MW re-enabled | — | — |
+| 13:55:06 | — | ✅ `MW=true W=2 R=2` | ~4 min |
+
+**Verdict**: ✅ **PASS** — All transitions detected. Fastest detection was <1 min (failover-priority-change propagates quickly via global endpoint). Transient `W=2` state at 13:49 reflects the failover mid-transition.
+
+```
+13:42-13:45  MW=false W=1 R=2   (MW just disabled)
+                                 ┌─ Write swapped EUS→WUS at 13:44
+13:45-13:49  MW=false W=1 R=2   (WUS is now write region)
+                                 ┌─ Write swapped WUS→EUS at 13:49
+13:49-13:50  MW=false W=2 R=2   (transitional)
+13:50-13:55  MW=false W=1 R=2   (EUS restored as write)
+                                 ┌─ MW RE-ENABLED at 13:51
+13:55+       MW=true  W=2 R=2   ← Full cycle complete
+```
+
+---
+
+## Background Refresh Continuity
+
+The core validation of the fix — the background refresh timer must never stop:
+
+| Period | Refreshes | Status |
+|--------|-----------|--------|
+| Startup (13:05) | 1 (init) | ✅ |
+| Drill 1: MW=true steady state | ~10 | ✅ Timer kept running |
+| Drill 1: MW=false (SW mode) | ~10 | ✅ Timer kept running |
+| Drill 1: MW=true restored | ~6 | ✅ Timer kept running |
+| Drill 2: Region removal | ~24 | ✅ Timer kept running |
+| Drill 2: Region re-add | ~12 | ✅ Timer kept running |
+| Drill 3: Failover changes | ~28 | ✅ Timer kept running |
+| Drill 3: MW restored | ~12 | ✅ Timer kept running |
+| **Total** | **103** | ✅ Continuous |
+
+**Unpatched comparison**: 0 refreshes after init. Timer stopped. None of these transitions would have been detected.
+
+---
+
+## Key Findings
+
+1. **Fix validated**: Background refresh timer runs continuously in MW accounts — the fix in `GlobalEndpointManager.refreshLocationPrivateAsync()` works as designed.
+
+2. **Metadata propagation latency**: Global endpoint reflects changes in 1-6 minutes depending on the operation type. `failover-priority-change` is fastest (~1 min), region removal is slowest (~6 min).
+
+3. **DIRECT mode resilience**: Region removal doesn't immediately cause read failures — RNTBD connections to replicas persist after metadata removal. This is by design for DIRECT mode.
+
+4. **Zero errors**: All 305 reads succeeded (HTTP 200) across all 3 drills. The SDK handled every topology change gracefully.
+
+5. **6-year-old bug**: This bug existed since [PR #6139](https://github.com/Azure/azure-sdk-for-java/pull/6139) (Nov 2019). The PR author explicitly acknowledged it in point #4 but it was never fixed until now.