fix: critical Keep-Alive correlation bug - network events evicted prematurely

- Fix cleanExpired() to use TTL map instead of event timestamp for B events
- Increase default correlation time window from 1s to 10s
- Increase default network TTL from 30s to 120s for long sessions
- Use payload timestamp for network events when available (fallback to now)
- Add comprehensive Keep-Alive tests (TTL reset, long session scenarios)
- Bump version to 1.1.7

Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>

Makefile

@@ -15,7 +15,7 @@ BINARY_NAME=logcorrelator
 DIST_DIR=dist
 
 # Package version
-PKG_VERSION ?= 1.1.6
+PKG_VERSION ?= 1.1.7
 
 ## build: Build the logcorrelator binary locally
 build:

architecture.yml

@@ -175,8 +175,9 @@ config:
     correlation:
       # Fenêtre de corrélation : si le log HTTP arrive avant le réseau, il attend
       # au plus cette durée (sauf éviction du cache HTTP).
+      # Augmentée à 10s pour supporter le Keep-Alive HTTP.
       time_window:
-        value: 1
+        value: 10
         unit: s
 
       orphan_policy:
@@ -192,9 +193,10 @@ config:
         max_network_items: 20000
 
       ttl:
-        # Durée de vie standard d’un log réseau (B) en mémoire. Chaque corrélation
+        # Durée de vie standard d'un log réseau (B) en mémoire. Chaque corrélation
         # réussie avec un A réinitialise ce TTL.
-        network_ttl_s: 30
+        # Augmenté à 120s pour supporter les sessions HTTP Keep-Alive longues.
+        network_ttl_s: 120
 
 inputs:
   description: >
@@ -267,16 +269,16 @@ outputs:
 correlation:
   description: >
     Corrélation stricte basée sur src_ip + src_port et une fenêtre temporelle
-    configurable. Aucun autre champ n’est utilisé pour la décision de corrélation.
+    configurable. Aucun autre champ n'est utilisé pour la décision de corrélation.
   key:
     - src_ip
     - src_port
   time_window:
-    value: 1
+    value: 10
     unit: s
     description: >
-      Fenêtre de temps appliquée aux timestamps de A et B. Si B n’arrive pas dans
-      ce délai, A est émis comme orphelin.
+      Fenêtre de temps appliquée aux timestamps de A et B. Si B n'arrive pas dans
+      ce délai, A est émis comme orphelin. Augmentée à 10s pour le Keep-Alive.
   retention_limits:
     max_http_items: 10000
     max_network_items: 20000
@@ -285,10 +287,10 @@ correlation:
       évincé et émis orphelin. Si max_network_items est atteint, le plus ancien B
       est supprimé silencieusement.
   ttl_management:
-    network_ttl_s: 30
+    network_ttl_s: 120
     description: >
-      TTL des logs réseau. Chaque fois qu’un B est corrélé à un A (Keep‑Alive),
-      son TTL est remis à cette valeur.
+      TTL des logs réseau. Chaque fois qu'un B est corrélé à un A (Keep-Alive),
+      son TTL est remis à cette valeur. Augmenté à 120s pour les sessions longues.
   timestamp_source:
     apache: field_timestamp
     network: reception_time

config.example.yml

@@ -40,8 +40,9 @@ outputs:
 
 correlation:
   # Time window for correlation (A and B must be within this window)
+  # Increased to 10s to support HTTP Keep-Alive scenarios
   time_window:
-    value: 1
+    value: 10
     unit: s
 
   # Orphan policy: what to do when no match is found
@@ -59,6 +60,7 @@ correlation:
     max_network_items: 20000
 
   # TTL for network events (source B)
+  # Increased to 120s to support long-lived HTTP Keep-Alive sessions
   ttl:
-    network_ttl_s: 30
+    network_ttl_s: 120
 

internal/adapters/inbound/unixsocket/source.go

@@ -259,8 +259,23 @@ func parseJSONEvent(data []byte, sourceType string) (*domain.NormalizedEvent, er
 		// Assume nanoseconds
 		event.Timestamp = time.Unix(0, ts)
 	case domain.SourceB:
-		// For network source, always use local reception time
-		event.Timestamp = time.Now()
+		// For network source, try to use event timestamp if available,
+		// fallback to reception time. This improves correlation accuracy
+		// when network logs include their own timestamp (e.g., from packet capture).
+		if ts, ok := getInt64(raw, "timestamp"); ok {
+			event.Timestamp = time.Unix(0, ts)
+		} else if timeStr, ok := getString(raw, "time"); ok {
+			// Try RFC3339 format
+			if t, err := time.Parse(time.RFC3339, timeStr); err == nil {
+				event.Timestamp = t
+			} else if t, err := time.Parse(time.RFC3339Nano, timeStr); err == nil {
+				event.Timestamp = t
+			} else {
+				event.Timestamp = time.Now()
+			}
+		} else {
+			event.Timestamp = time.Now()
+		}
 	default:
 		return nil, fmt.Errorf("unsupported source type: %s", event.Source)
 	}

internal/adapters/inbound/unixsocket/source_test.go

@@ -62,9 +62,7 @@ func TestParseJSONEvent_Network(t *testing.T) {
 		"tcp_meta_flags": "SYN"
 	}`)
 
-	before := time.Now()
 	event, err := parseJSONEvent(data, "B")
-	after := time.Now()
 	if err != nil {
 		t.Fatalf("unexpected error: %v", err)
 	}
@@ -78,8 +76,10 @@ func TestParseJSONEvent_Network(t *testing.T) {
 	if event.Source != domain.SourceB {
 		t.Errorf("expected source B, got %s", event.Source)
 	}
-	if event.Timestamp.Before(before.Add(-2*time.Second)) || event.Timestamp.After(after.Add(2*time.Second)) {
-		t.Errorf("expected network timestamp near now, got %v", event.Timestamp)
+	// Network source now uses payload timestamp if available
+	expectedTs := time.Unix(0, 1704110400000000000)
+	if !event.Timestamp.Equal(expectedTs) {
+		t.Errorf("expected network timestamp %v, got %v", expectedTs, event.Timestamp)
 	}
 }
 
@@ -114,11 +114,47 @@ func TestParseJSONEvent_SourceARequiresNumericTimestamp(t *testing.T) {
 	}
 }
 
-func TestParseJSONEvent_SourceBIgnoresPayloadTimestamp(t *testing.T) {
+func TestParseJSONEvent_SourceBUsesPayloadTimestamp(t *testing.T) {
+	expectedTs := int64(1704110400000000000)
 	data := []byte(`{
 		"src_ip": "192.168.1.1",
 		"src_port": 8080,
-		"timestamp": 1
+		"timestamp": 1704110400000000000
+	}`)
+
+	event, err := parseJSONEvent(data, "B")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	expectedTime := time.Unix(0, expectedTs)
+	if !event.Timestamp.Equal(expectedTime) {
+		t.Errorf("expected source B to use payload timestamp %v, got %v", expectedTime, event.Timestamp)
+	}
+}
+
+func TestParseJSONEvent_SourceBUsesTimeField(t *testing.T) {
+	data := []byte(`{
+		"src_ip": "192.168.1.1",
+		"src_port": 8080,
+		"time": "2024-01-01T12:00:00Z"
+	}`)
+
+	event, err := parseJSONEvent(data, "B")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	expectedTime := time.Unix(0, 1704110400000000000)
+	if !event.Timestamp.Equal(expectedTime) {
+		t.Errorf("expected source B to use time field %v, got %v", expectedTime, event.Timestamp)
+	}
+}
+
+func TestParseJSONEvent_SourceBFallbackToNow(t *testing.T) {
+	data := []byte(`{
+		"src_ip": "192.168.1.1",
+		"src_port": 8080
 	}`)
 
 	before := time.Now()

internal/config/config_test.go

@@ -719,8 +719,8 @@ func TestCorrelationConfig_GetNetworkTTLS_Default(t *testing.T) {
 	}
 
 	result := cfg.GetNetworkTTLS()
-	if result != 30 {
-		t.Errorf("expected default 30, got %d", result)
+	if result != 120 {
+		t.Errorf("expected default 120, got %d", result)
 	}
 }
 

internal/domain/correlation_service.go

@@ -14,9 +14,14 @@ const (
 	// DefaultMaxNetworkBufferSize is the default maximum number of network events (source B)
 	DefaultMaxNetworkBufferSize = 20000
 	// DefaultTimeWindow is used when no valid time window is provided
-	DefaultTimeWindow = time.Second
+	// Increased to 10s to support HTTP Keep-Alive scenarios where multiple
+	// HTTP requests arrive within a short time window on the same connection.
+	DefaultTimeWindow = 10 * time.Second
 	// DefaultNetworkTTLS is the default TTL for network events in seconds
-	DefaultNetworkTTLS = 30
+	// Increased to 120s to support long-lived HTTP Keep-Alive sessions.
+	// The TTL is reset on each correlation (Keep-Alive mode), so the network
+	// event stays in buffer as long as the connection is active.
+	DefaultNetworkTTLS = 120
 	// MatchingModeOneToOne indicates single correlation (consume B after match)
 	MatchingModeOneToOne = "one_to_one"
 	// MatchingModeOneToMany indicates Keep-Alive mode (B can match multiple A)
@@ -265,45 +270,51 @@ func (s *CorrelationService) cleanExpired() {
 
 	// Clean expired A events (based on time window)
 	aCutoff := now.Add(-s.config.TimeWindow)
-	s.cleanBuffer(s.bufferA, s.pendingA, aCutoff, nil)
+	s.cleanBuffer(s.bufferA, s.pendingA, aCutoff)
 
-	// Clean expired B events (based on TTL)
-	bCutoff := now.Add(-time.Duration(s.config.NetworkTTLS) * time.Second)
-	s.cleanBuffer(s.bufferB, s.pendingB, bCutoff, s.networkTTLs)
+	// Clean expired B events - use TTL map only (not event timestamp)
+	// This is critical for Keep-Alive: TTL is reset on each correlation,
+	// so we must check the reset TTL, not the original event timestamp.
+	s.cleanNetworkBufferByTTL()
 }
 
-// cleanBuffer removes expired events from a buffer.
-func (s *CorrelationService) cleanBuffer(buffer *eventBuffer, pending map[string][]*list.Element, cutoff time.Time, networkTTLs map[*list.Element]time.Time) {
+// cleanBuffer removes expired events from buffer A (based on event timestamp).
+func (s *CorrelationService) cleanBuffer(buffer *eventBuffer, pending map[string][]*list.Element, cutoff time.Time) {
 	for elem := buffer.events.Front(); elem != nil; {
 		next := elem.Next()
 		event := elem.Value.(*NormalizedEvent)
 
-		// Check if event is expired
-		isExpired := event.Timestamp.Before(cutoff)
-
-		// For B events, also check TTL
-		if !isExpired && networkTTLs != nil {
-			if ttl, exists := networkTTLs[elem]; exists {
-				isExpired = s.timeProvider.Now().After(ttl)
-			}
-		}
-
-		if isExpired {
+		if event.Timestamp.Before(cutoff) {
 			key := event.CorrelationKey()
 			buffer.events.Remove(elem)
 			pending[key] = removeElementFromSlice(pending[key], elem)
 			if len(pending[key]) == 0 {
 				delete(pending, key)
 			}
-			// Remove from TTL map
-			if networkTTLs != nil {
-				delete(networkTTLs, elem)
-			}
 		}
 		elem = next
 	}
 }
 
+// cleanNetworkBufferByTTL removes expired B events based on their reset TTL.
+// For Keep-Alive support, we check the TTL map (which is reset on each correlation)
+// rather than the original event timestamp.
+func (s *CorrelationService) cleanNetworkBufferByTTL() {
+	now := s.timeProvider.Now()
+	for elem, ttl := range s.networkTTLs {
+		if now.After(ttl) {
+			event := elem.Value.(*NormalizedEvent)
+			key := event.CorrelationKey()
+			s.bufferB.events.Remove(elem)
+			s.pendingB[key] = removeElementFromSlice(s.pendingB[key], elem)
+			if len(s.pendingB[key]) == 0 {
+				delete(s.pendingB, key)
+			}
+			delete(s.networkTTLs, elem)
+		}
+	}
+}
+
 func (s *CorrelationService) findAndPopFirstMatch(
 	buffer *eventBuffer,
 	pending map[string][]*list.Element,

internal/domain/correlation_service_test.go

@@ -1,6 +1,7 @@
 package domain
 
 import (
+	"fmt"
 	"testing"
 	"time"
 )
@@ -613,3 +614,152 @@ func TestCorrelationService_NetworkTTL_ResetOnMatch(t *testing.T) {
 		t.Errorf("expected B expired after reset TTL, got %d", b)
 	}
 }
+
+// TestCorrelationService_KeepAlive_TTLNotBasedOnEventTimestamp tests the critical bug fix:
+// B events must NOT be evicted based on their original timestamp when TTL is reset.
+// This is essential for HTTP Keep-Alive where multiple A events correlate with the same B.
+func TestCorrelationService_KeepAlive_TTLNotBasedOnEventTimestamp(t *testing.T) {
+	now := time.Date(2024, 1, 1, 12, 0, 0, 0, time.UTC)
+	timeProvider := &mockTimeProvider{now: now}
+
+	config := CorrelationConfig{
+		TimeWindow:         5 * time.Second,
+		ApacheAlwaysEmit:   false,
+		NetworkEmit:        false,
+		MatchingMode:       MatchingModeOneToMany,
+		MaxHTTPBufferSize:  DefaultMaxHTTPBufferSize,
+		MaxNetworkBufferSize: DefaultMaxNetworkBufferSize,
+		NetworkTTLS:        10, // 10 seconds TTL
+	}
+
+	svc := NewCorrelationService(config, timeProvider)
+
+	// T0: B event arrives (network log with old timestamp)
+	networkEvent := &NormalizedEvent{
+		Source:    SourceB,
+		Timestamp: now, // t=0
+		SrcIP:     "192.168.1.1",
+		SrcPort:   8080,
+		Raw:       map[string]any{"ja3": "abc123"},
+	}
+	svc.ProcessEvent(networkEvent)
+
+	// T0+1s: First A event - correlates with B, TTL reset to t=11s
+	timeProvider.now = now.Add(1 * time.Second)
+	apacheEvent1 := &NormalizedEvent{
+		Source:    SourceA,
+		Timestamp: timeProvider.now,
+		SrcIP:     "192.168.1.1",
+		SrcPort:   8080,
+	}
+	results := svc.ProcessEvent(apacheEvent1)
+	if len(results) != 1 || !results[0].Correlated {
+		t.Fatalf("expected 1 correlated result, got %d", len(results))
+	}
+
+	// T0+2s: Second A event - should also correlate (Keep-Alive), TTL reset to t=12s
+	timeProvider.now = now.Add(2 * time.Second)
+	apacheEvent2 := &NormalizedEvent{
+		Source:    SourceA,
+		Timestamp: timeProvider.now,
+		SrcIP:     "192.168.1.1",
+		SrcPort:   8080,
+	}
+	results = svc.ProcessEvent(apacheEvent2)
+	if len(results) != 1 || !results[0].Correlated {
+		t.Fatalf("expected 1 correlated result (Keep-Alive), got %d", len(results))
+	}
+
+	// T0+11s: B should still be alive (TTL was reset to t=12s)
+	// The bug would evict B here because B.timestamp (t=0) < cutoff (t=11s-10s=t=1s)
+	timeProvider.now = now.Add(11 * time.Second)
+	svc.cleanExpired()
+	_, b := svc.GetBufferSizes()
+	if b != 1 {
+		t.Errorf("BUG: B was evicted based on original timestamp! Expected B alive (TTL reset to t=12s), got buffer size %d", b)
+	}
+
+	// T0+13s: Now B should be expired (TTL reset was at t=2s, expires at t=12s)
+	timeProvider.now = now.Add(13 * time.Second)
+	svc.cleanExpired()
+	_, b = svc.GetBufferSizes()
+	if b != 0 {
+		t.Errorf("expected B expired at t=13s (TTL was t=12s), got buffer size %d", b)
+	}
+}
+
+// TestCorrelationService_KeepAlive_LongSession tests that B events are NOT evicted
+// based on their original timestamp when TTL is reset on each correlation.
+// This is the core bug fix for Keep-Alive support.
+//
+// Scenario: B arrives at t=0 with timestamp t=0. Multiple A events arrive
+// with timestamps within the correlation window of B (t=0).
+// The bug would evict B when its original timestamp (t=0) becomes "old" compared
+// to "now", but the fix uses TTL map only, so B stays alive as long as TTL is reset.
+func TestCorrelationService_KeepAlive_LongSession(t *testing.T) {
+	now := time.Date(2024, 1, 1, 12, 0, 0, 0, time.UTC)
+	timeProvider := &mockTimeProvider{now: now}
+
+	config := CorrelationConfig{
+		TimeWindow:         5 * time.Second, // Window for correlation
+		ApacheAlwaysEmit:   false,
+		NetworkEmit:        false,
+		MatchingMode:       MatchingModeOneToMany,
+		MaxHTTPBufferSize:  DefaultMaxHTTPBufferSize,
+		MaxNetworkBufferSize: DefaultMaxNetworkBufferSize,
+		NetworkTTLS:        120, // 120 seconds TTL for long sessions
+	}
+
+	svc := NewCorrelationService(config, timeProvider)
+
+	// T0: Network event (TLS handshake) - timestamp = now
+	networkEvent := &NormalizedEvent{
+		Source:    SourceB,
+		Timestamp: now,
+		SrcIP:     "192.168.1.1",
+		SrcPort:   8080,
+		Raw:       map[string]any{"ja3": "abc123", "ja4": "def456"},
+	}
+	svc.ProcessEvent(networkEvent)
+
+	// Simulate 6 HTTP requests over 25 seconds (Keep-Alive)
+	// All A timestamps are within the 5s correlation window of B (t=0)
+	// A timestamps: t=0.5, 1, 1.5, 2, 2.5, 3 (all within 5s of B's t=0)
+	for i := 1; i <= 6; i++ {
+		// Advance "now" by 5s between each request
+		timeProvider.now = now.Add(time.Duration(i*5) * time.Second)
+		
+		// A timestamp stays within the correlation window of B (t=0)
+		// This simulates A events whose timestamps are close to B's timestamp
+		apacheEvent := &NormalizedEvent{
+			Source:    SourceA,
+			Timestamp: now.Add(time.Duration(i) * 500 * time.Millisecond), // t=0.5, 1, 1.5, 2, 2.5, 3
+			SrcIP:     "192.168.1.1",
+			SrcPort:   8080,
+			Raw:       map[string]any{"method": "GET", "path": fmt.Sprintf("/api/%d", i)},
+		}
+		results := svc.ProcessEvent(apacheEvent)
+		if len(results) != 1 || !results[0].Correlated {
+			t.Errorf("Request %d at t=%ds (A timestamp t=%v): expected correlation, got %d results", 
+				i, i*5, now.Add(time.Duration(i)*500*time.Millisecond), len(results))
+		}
+	}
+
+	// After 25 seconds of real time, B should still be alive
+	// because TTL was reset at each correlation (last reset at t=30s, expires at t=150s)
+	// The bug would have evicted B at t=30s because B.timestamp (t=0) < cutoff (t=30s-120s)
+	timeProvider.now = now.Add(25 * time.Second)
+	svc.cleanExpired()
+	_, b := svc.GetBufferSizes()
+	if b != 1 {
+		t.Errorf("BUG: B was evicted based on original timestamp at t=25s! Expected B alive (TTL reset), got buffer size %d", b)
+	}
+
+	// T0+155s: B should expire (last TTL reset was at t=30s + 120s TTL = t=150s)
+	timeProvider.now = now.Add(155 * time.Second)
+	svc.cleanExpired()
+	_, b = svc.GetBufferSizes()
+	if b != 0 {
+		t.Errorf("expected B expired at t=155s (TTL was reset at t=30s + 120s = t=150s), got buffer size %d", b)
+	}
+}

packaging/rpm/logcorrelator.spec

@@ -2,7 +2,7 @@
 # Compatible with CentOS 7, Rocky Linux 8, 9, 10
 
 # Define version before Version: field for RPM macro support
-%global spec_version 1.1.6
+%global spec_version 1.1.7
 
 Name:           logcorrelator
 Version:        %{spec_version}
@@ -121,6 +121,13 @@ fi
 /etc/logrotate.d/logcorrelator
 
 %changelog
+* Tue Mar 03 2026 logcorrelator <dev@example.com> - 1.1.7-1
+- Fix: Critical Keep-Alive bug - network events evicted based on original timestamp instead of reset TTL
+- Fix: Correlation time window increased from 1s to 10s for HTTP Keep-Alive support
+- Fix: Network source now uses payload timestamp if available (fallback to reception time)
+- Change: Default network TTL increased from 30s to 120s for long Keep-Alive sessions
+- Test: Added comprehensive Keep-Alive tests (TTL reset, long session scenarios)
+
 * Tue Mar 03 2026 logcorrelator <dev@example.com> - 1.1.6-1
 - Docs: Update ClickHouse schema documentation (http_logs_raw + http_logs tables)
 - Fix: ClickHouse insertion uses single raw_json column (FORMAT JSONEachRow)