resolver: auto-tune nx-heavy batch share by run health

selective-vpn-api/app/resolver.go

@@ -218,10 +218,13 @@ type resolverLiveBatchStats struct {
 	P1           int
 	P2           int
 	P3           int
+	NXHeavyPct   int
 	NXHeavyTotal int
 	NXHeavySkip  int
 	NextTarget   int
 	NextReason   string
+	NextNXPct    int
+	NextNXReason string
 	DNSAttempts  int
 	DNSTimeout   int
 	DNSCoolSkips int
@@ -413,13 +416,28 @@ func runResolverJob(opts ResolverOpts, logf func(string, ...any)) (resolverResul
 		liveBatchDefault = liveBatchMax
 	}
 	liveBatchTarget := loadResolverLiveBatchTarget(precheckStatePath, liveBatchDefault, liveBatchMin, liveBatchMax)
-	liveBatchNXHeavyPct := envInt("RESOLVE_LIVE_BATCH_NX_HEAVY_PCT", 10)
+	liveBatchNXHeavyMin := envInt("RESOLVE_LIVE_BATCH_NX_HEAVY_MIN_PCT", 5)
-	if liveBatchNXHeavyPct < 0 {
+	liveBatchNXHeavyMax := envInt("RESOLVE_LIVE_BATCH_NX_HEAVY_MAX_PCT", 35)
-		liveBatchNXHeavyPct = 0
+	liveBatchNXHeavyDefault := envInt("RESOLVE_LIVE_BATCH_NX_HEAVY_PCT", 10)
+	if liveBatchNXHeavyMin < 0 {
+		liveBatchNXHeavyMin = 0
 	}
-	if liveBatchNXHeavyPct > 100 {
+	if liveBatchNXHeavyMin > 100 {
-		liveBatchNXHeavyPct = 100
+		liveBatchNXHeavyMin = 100
 	}
+	if liveBatchNXHeavyMax < liveBatchNXHeavyMin {
+		liveBatchNXHeavyMax = liveBatchNXHeavyMin
+	}
+	if liveBatchNXHeavyMax > 100 {
+		liveBatchNXHeavyMax = 100
+	}
+	if liveBatchNXHeavyDefault < liveBatchNXHeavyMin {
+		liveBatchNXHeavyDefault = liveBatchNXHeavyMin
+	}
+	if liveBatchNXHeavyDefault > liveBatchNXHeavyMax {
+		liveBatchNXHeavyDefault = liveBatchNXHeavyMax
+	}
+	liveBatchNXHeavyPct := loadResolverLiveBatchNXHeavyPct(precheckStatePath, liveBatchNXHeavyDefault, liveBatchNXHeavyMin, liveBatchNXHeavyMax)
 	precheckEnvForced := resolvePrecheckForceEnvEnabled()
 	precheckFileForced := resolvePrecheckForceFileEnabled(precheckForcePath)
 	precheckDue := precheckEnvForced || precheckFileForced || (precheckEverySec > 0 && (precheckLastRun <= 0 || now-precheckLastRun >= precheckEverySec))
@@ -484,7 +502,7 @@ func runResolverJob(opts ResolverOpts, logf func(string, ...any)) (resolverResul
 		wildcardPolicy := wildcardDNSAttemptPolicy(1)
 		cEnabled, cMin, cRate, cStreak, cBan, cMaxBan := cooldown.configSnapshot()
 		logf(
-			"resolver policy: direct_try=%d direct_budget_ms=%d wildcard_try=%d wildcard_budget_ms=%d nx_early_stop=%t nx_hard_quarantine=%t cooldown_enabled=%t cooldown_min_attempts=%d cooldown_timeout_rate=%d cooldown_fail_streak=%d cooldown_ban_sec=%d cooldown_max_ban_sec=%d live_batch_target=%d live_batch_min=%d live_batch_max=%d live_batch_nx_heavy_pct=%d stale_keep_sec=%d precheck_every_sec=%d precheck_max=%d precheck_forced_env=%t precheck_forced_file=%t",
+			"resolver policy: direct_try=%d direct_budget_ms=%d wildcard_try=%d wildcard_budget_ms=%d nx_early_stop=%t nx_hard_quarantine=%t cooldown_enabled=%t cooldown_min_attempts=%d cooldown_timeout_rate=%d cooldown_fail_streak=%d cooldown_ban_sec=%d cooldown_max_ban_sec=%d live_batch_target=%d live_batch_min=%d live_batch_max=%d live_batch_nx_heavy_pct=%d live_batch_nx_heavy_min=%d live_batch_nx_heavy_max=%d stale_keep_sec=%d precheck_every_sec=%d precheck_max=%d precheck_forced_env=%t precheck_forced_file=%t",
 			directPolicy.TryLimit,
 			directPolicy.DomainBudget.Milliseconds(),
 			wildcardPolicy.TryLimit,
@@ -501,6 +519,8 @@ func runResolverJob(opts ResolverOpts, logf func(string, ...any)) (resolverResul
 			liveBatchMin,
 			liveBatchMax,
 			liveBatchNXHeavyPct,
+			liveBatchNXHeavyMin,
+			liveBatchNXHeavyMax,
 			staleKeepSec,
 			precheckEverySec,
 			precheckMaxDomains,
@@ -783,7 +803,7 @@ func runResolverJob(opts ResolverOpts, logf func(string, ...any)) (resolverResul
 		dnsErrors := dnsStats.totalErrors()
 		unresolvedSuppressed := cacheNegativeHits + quarantineHits + liveDeferred
 		logf(
-			"resolve summary: domains=%d cache_hits=%d cache_neg_hits=%d quarantine_hits=%d stale_hits=%d resolved_now=%d unresolved=%d unresolved_live=%d unresolved_suppressed=%d live_batch_target=%d live_batch_deferred=%d live_batch_p1=%d live_batch_p2=%d live_batch_p3=%d live_batch_nxheavy_total=%d live_batch_nxheavy_skip=%d static_entries=%d static_skipped=%d unique_ips=%d direct_ips=%d wildcard_ips=%d ptr_lookups=%d ptr_errors=%d dns_attempts=%d dns_ok=%d dns_nxdomain=%d dns_timeout=%d dns_temporary=%d dns_other=%d dns_cooldown_skips=%d dns_errors=%d timeout_recheck_checked=%d timeout_recheck_recovered=%d timeout_recheck_recovered_ips=%d timeout_recheck_still_timeout=%d timeout_recheck_now_nxdomain=%d timeout_recheck_now_temporary=%d timeout_recheck_now_other=%d timeout_recheck_no_signal=%d duration_ms=%d",
+			"resolve summary: domains=%d cache_hits=%d cache_neg_hits=%d quarantine_hits=%d stale_hits=%d resolved_now=%d unresolved=%d unresolved_live=%d unresolved_suppressed=%d live_batch_target=%d live_batch_deferred=%d live_batch_p1=%d live_batch_p2=%d live_batch_p3=%d live_batch_nxheavy_pct=%d live_batch_nxheavy_total=%d live_batch_nxheavy_skip=%d static_entries=%d static_skipped=%d unique_ips=%d direct_ips=%d wildcard_ips=%d ptr_lookups=%d ptr_errors=%d dns_attempts=%d dns_ok=%d dns_nxdomain=%d dns_timeout=%d dns_temporary=%d dns_other=%d dns_cooldown_skips=%d dns_errors=%d timeout_recheck_checked=%d timeout_recheck_recovered=%d timeout_recheck_recovered_ips=%d timeout_recheck_still_timeout=%d timeout_recheck_now_nxdomain=%d timeout_recheck_now_temporary=%d timeout_recheck_now_other=%d timeout_recheck_no_signal=%d duration_ms=%d",
 			len(domains),
 			len(fresh),
 			cacheNegativeHits,
@@ -798,6 +818,7 @@ func runResolverJob(opts ResolverOpts, logf func(string, ...any)) (resolverResul
 			liveP1,
 			liveP2,
 			liveP3,
+			liveBatchNXHeavyPct,
 			liveNXHeavyTotal,
 			liveNXHeavySkip,
 			len(staticEntries),
@@ -850,6 +871,27 @@ func runResolverJob(opts ResolverOpts, logf func(string, ...any)) (resolverResul
 	}
 	if precheckDue {
 		nextTarget, nextReason := computeNextLiveBatchTarget(liveBatchTarget, liveBatchMin, liveBatchMax, dnsStats, liveDeferred)
+		nextNXPct, nextNXReason := computeNextLiveBatchNXHeavyPct(
+			liveBatchNXHeavyPct,
+			liveBatchNXHeavyMin,
+			liveBatchNXHeavyMax,
+			dnsStats,
+			resolvedNowDNS,
+			liveP3,
+			liveNXHeavyTotal,
+			liveNXHeavySkip,
+		)
+		if logf != nil {
+			logf(
+				"resolve live-batch nxheavy: pct=%d next=%d reason=%s selected=%d total=%d skipped=%d",
+				liveBatchNXHeavyPct,
+				nextNXPct,
+				nextNXReason,
+				liveP3,
+				liveNXHeavyTotal,
+				liveNXHeavySkip,
+			)
+		}
 		saveResolverPrecheckState(
 			precheckStatePath,
 			now,
@@ -861,10 +903,13 @@ func runResolverJob(opts ResolverOpts, logf func(string, ...any)) (resolverResul
 				P1:           liveP1,
 				P2:           liveP2,
 				P3:           liveP3,
+				NXHeavyPct:   liveBatchNXHeavyPct,
 				NXHeavyTotal: liveNXHeavyTotal,
 				NXHeavySkip:  liveNXHeavySkip,
 				NextTarget:   nextTarget,
 				NextReason:   nextReason,
+				NextNXPct:    nextNXPct,
+				NextNXReason: nextNXReason,
 				DNSAttempts:  dnsStats.Attempts,
 				DNSTimeout:   dnsStats.Timeout,
 				DNSCoolSkips: dnsStats.Skipped,
@@ -2473,6 +2518,34 @@ func loadResolverLiveBatchTarget(path string, fallback, minV, maxV int) int {
 	return v
 }
 
+func loadResolverLiveBatchNXHeavyPct(path string, fallback, minV, maxV int) int {
+	if fallback < minV {
+		fallback = minV
+	}
+	if fallback > maxV {
+		fallback = maxV
+	}
+	m := loadJSONMap(path)
+	if len(m) == 0 {
+		return fallback
+	}
+	raw := m["live_batch_nxheavy_next_pct"]
+	if raw == nil {
+		raw = m["live_batch_nxheavy_pct"]
+	}
+	v, ok := parseAnyInt(raw)
+	if !ok {
+		return fallback
+	}
+	if v < minV {
+		v = minV
+	}
+	if v > maxV {
+		v = maxV
+	}
+	return v
+}
+
 func computeNextLiveBatchTarget(current, minV, maxV int, dnsStats dnsMetrics, deferred int) (int, string) {
 	if current < minV {
 		current = minV
@@ -2517,6 +2590,73 @@ func computeNextLiveBatchTarget(current, minV, maxV int, dnsStats dnsMetrics, de
 	return next, reason
 }
 
+func computeNextLiveBatchNXHeavyPct(
+	current, minV, maxV int,
+	dnsStats dnsMetrics,
+	resolvedNowDNS int,
+	liveP3 int,
+	liveNXHeavyTotal int,
+	liveNXHeavySkip int,
+) (int, string) {
+	if current < minV {
+		current = minV
+	}
+	if current > maxV {
+		current = maxV
+	}
+	next := current
+	reason := "stable"
+	attempts := dnsStats.Attempts
+	timeoutRate := 0.0
+	nxRate := 0.0
+	okRate := 0.0
+	if attempts > 0 {
+		timeoutRate = float64(dnsStats.Timeout) / float64(attempts)
+		nxRate = float64(dnsStats.NXDomain) / float64(attempts)
+		okRate = float64(dnsStats.OK) / float64(attempts)
+	}
+	nxSelectedRatio := 0.0
+	if liveNXHeavyTotal > 0 {
+		nxSelectedRatio = float64(liveP3) / float64(liveNXHeavyTotal)
+	}
+
+	switch {
+	case attempts == 0:
+		reason = "no_dns_attempts"
+	case timeoutRate >= 0.20 || dnsStats.Skipped > 0:
+		next = current - 3
+		reason = "timeout_very_high_or_cooldown"
+	case timeoutRate >= 0.12:
+		next = current - 2
+		reason = "timeout_high"
+	case dnsStats.OK == 0 && dnsStats.NXDomain > 0:
+		next = current - 2
+		reason = "no_success_nx_only"
+	case nxRate >= 0.90 && resolvedNowDNS == 0:
+		next = current - 2
+		reason = "nx_dominant_no_resolve"
+	case nxSelectedRatio >= 0.95 && resolvedNowDNS == 0:
+		next = current - 1
+		reason = "nxheavy_selected_dominant"
+	case timeoutRate <= 0.02 && okRate >= 0.10 && liveNXHeavySkip > 0:
+		next = current + 2
+		reason = "healthy_fast_reintroduce_nxheavy"
+	case timeoutRate <= 0.04 && resolvedNowDNS > 0 && liveNXHeavySkip > 0:
+		next = current + 1
+		reason = "healthy_reintroduce_nxheavy"
+	}
+	if next < minV {
+		next = minV
+	}
+	if next > maxV {
+		next = maxV
+	}
+	if next == current && reason != "no_dns_attempts" {
+		reason = "stable"
+	}
+	return next, reason
+}
+
 func classifyLiveBatchHost(
 	host string,
 	cache domainCacheState,
@@ -2701,8 +2841,11 @@ func saveResolverPrecheckState(path string, ts int, timeoutStats resolverTimeout
 	state["live_batch_p1"] = live.P1
 	state["live_batch_p2"] = live.P2
 	state["live_batch_p3"] = live.P3
+	state["live_batch_nxheavy_pct"] = live.NXHeavyPct
 	state["live_batch_nxheavy_total"] = live.NXHeavyTotal
 	state["live_batch_nxheavy_skip"] = live.NXHeavySkip
+	state["live_batch_nxheavy_next_pct"] = live.NextNXPct
+	state["live_batch_nxheavy_next_reason"] = live.NextNXReason
 	state["live_batch_next_target"] = live.NextTarget
 	state["live_batch_next_reason"] = live.NextReason
 	state["live_batch_dns_attempts"] = live.DNSAttempts

selective-vpn-gui/dashboard_controller.py

@@ -672,6 +672,7 @@ class DashboardController:
         live_batch_p1 = int(pairs.get("live_batch_p1", 0))
         live_batch_p2 = int(pairs.get("live_batch_p2", 0))
         live_batch_p3 = int(pairs.get("live_batch_p3", 0))
+        live_batch_nxheavy_pct = int(pairs.get("live_batch_nxheavy_pct", 0))
         live_batch_nxheavy_skip = int(pairs.get("live_batch_nxheavy_skip", 0))
 
         r_checked = int(pairs.get("timeout_recheck_checked", 0))
@@ -688,7 +689,7 @@ class DashboardController:
             f"quarantine_hits={q_hits} | dns_timeout={dns_timeout} "
             f"| cooldown_skips={dns_cooldown_skips} | attempts={dns_attempts} "
             f"| live_batch={live_batch_target} deferred={live_batch_deferred} "
-            f"(p1={live_batch_p1}, p2={live_batch_p2}, p3={live_batch_p3}, nx_skip={live_batch_nxheavy_skip})"
+            f"(p1={live_batch_p1}, p2={live_batch_p2}, p3={live_batch_p3}, nx_pct={live_batch_nxheavy_pct}, nx_skip={live_batch_nxheavy_skip})"
         )
         recheck_text = (
             f"Timeout recheck: checked={r_checked} recovered={r_recovered} "