Monitoring and Operations

Overview

It is in production. Three things will wake you up at 3am if you do not monitor them.

01. The Three Things

After running graph databases in production across multiple organizations, three categories of failure account for the vast majority of 3am pages:

1. Query latency spikes. A query that ran in 200ms starts taking 15 seconds, and users see timeouts. The cause is usually an unindexed property lookup, a Cartesian product in a Cypher query, or a traversal that hit a supernode — a single node with 100,000 or more relationships.

2. Graph size growth. The graph was designed for 1 million nodes. It now has 50 million because someone added a data source without considering cardinality. Memory exhausts, queries slow to a crawl, and the page cache starts evicting.

3. Extraction pipeline failures. The CDC pipeline from Chapter 13 stops processing events. The graph falls hours behind the relational database. Applications serve stale data, and nobody notices until a business user reports wrong results.

02. Key Metrics Dashboard

Every graph database deployment needs a dashboard with these metrics. Whether you use Grafana, Datadog, or CloudWatch, track all of them from day one.

Query Metrics

Resource Metrics

Graph-Specific Metrics

Pipeline Metrics

Collecting Graph-Specific Metrics

import time
import logging
from dataclasses import dataclass, field
from neo4j import GraphDatabase

logger = logging.getLogger("graph_metrics")


@dataclass
class GraphMetrics:
    """Snapshot of graph database metrics."""
    timestamp: float = field(default_factory=time.time)
    node_counts: dict = field(default_factory=dict)
    rel_counts: dict = field(default_factory=dict)
    total_nodes: int = 0
    total_relationships: int = 0
    max_degree: int = 0
    max_degree_node: str = ""
    store_size_mb: float = 0.0


def collect_graph_metrics(session) -> GraphMetrics:
    """Collect key graph metrics for dashboard."""
    metrics = GraphMetrics()

    # Node counts by label
    result = session.run("""
        CALL db.labels() YIELD label
        CALL {
            WITH label
            MATCH (n) WHERE label IN labels(n)
            RETURN count(n) AS cnt
        }
        RETURN label, cnt ORDER BY cnt DESC
    """)
    for record in result:
        metrics.node_counts[record["label"]] = record["cnt"]
        metrics.total_nodes += record["cnt"]

    # Relationship counts by type
    result = session.run("""
        CALL db.relationshipTypes() YIELD relationshipType AS type
        CALL {
            WITH type
            MATCH ()-[r]->() WHERE type(r) = type
            RETURN count(r) AS cnt
        }
        RETURN type, cnt ORDER BY cnt DESC
    """)
    for record in result:
        metrics.rel_counts[record["type"]] = record["cnt"]
        metrics.total_relationships += record["cnt"]

    # Supernode detection: find the node with most relationships
    result = session.run("""
        MATCH (n)
        WITH n, size([(n)--() | 1]) AS degree
        ORDER BY degree DESC
        LIMIT 1
        RETURN coalesce(n.name, n.id, toString(id(n))) AS node_id,
               labels(n) AS labels, degree
    """)
    record = result.single()
    if record:
        metrics.max_degree = record["degree"]
        metrics.max_degree_node = (
            f"{record['labels'][0]}:{record['node_id']}"
        )

    return metrics


def format_metrics_for_prometheus(metrics: GraphMetrics) -> str:
    """Format metrics as Prometheus text exposition."""
    lines = []
    lines.append(
        f"neo4j_total_nodes {metrics.total_nodes}"
    )
    lines.append(
        f"neo4j_total_relationships {metrics.total_relationships}"
    )
    lines.append(
        f"neo4j_max_node_degree {metrics.max_degree}"
    )

    for label, count in metrics.node_counts.items():
        lines.append(
            f'neo4j_node_count{{label="{label}"}} {count}'
        )

    for rel_type, count in metrics.rel_counts.items():
        lines.append(
            f'neo4j_rel_count{{type="{rel_type}"}} {count}'
        )

    return "\n".join(lines)

03. Alerting Rules: What to Alert On vs What to Log

Not everything is an alert. Alert fatigue kills on-call teams faster than actual incidents.

Alert Hierarchy

Specific Thresholds

Alerting Implementation

from dataclasses import dataclass
from enum import Enum

class AlertLevel(Enum):
    INFO = "info"
    WARNING = "warning"
    URGENT = "urgent"
    PAGE = "page"


@dataclass
class AlertRule:
    name: str
    metric: str
    warning_threshold: float
    urgent_threshold: float
    page_threshold: float
    sustained_minutes: int = 0    # 0 = instant alert
    description: str = ""


ALERT_RULES = [
    AlertRule(
        name="query_latency_p95",
        metric="neo4j_query_latency_p95_ms",
        warning_threshold=500,
        urgent_threshold=2000,
        page_threshold=5000,
        sustained_minutes=5,
        description="Query latency p95 exceeds threshold"
    ),
    AlertRule(
        name="cache_hit_ratio",
        metric="neo4j_page_cache_hit_ratio",
        warning_threshold=90,   # Alert when BELOW threshold
        urgent_threshold=80,
        page_threshold=60,
        description="Page cache hit ratio below threshold"
    ),
    AlertRule(
        name="cdc_consumer_lag",
        metric="kafka_consumer_lag_seconds",
        warning_threshold=300,
        urgent_threshold=900,
        page_threshold=3600,
        sustained_minutes=2,
        description="CDC pipeline falling behind"
    ),
    AlertRule(
        name="disk_utilization",
        metric="disk_utilization_percent",
        warning_threshold=70,
        urgent_threshold=85,
        page_threshold=95,
        description="Disk space running low"
    ),
    AlertRule(
        name="supernode_degree",
        metric="neo4j_max_node_degree",
        warning_threshold=10_000,
        urgent_threshold=50_000,
        page_threshold=200_000,
        description="Supernode detected — high relationship count"
    ),
]


def evaluate_alert(rule: AlertRule, current_value: float) -> AlertLevel:
    """Evaluate a metric against alert thresholds."""
    # For "lower is worse" metrics like cache hit ratio
    if rule.name in ("cache_hit_ratio",):
        if current_value < rule.page_threshold:
            return AlertLevel.PAGE
        if current_value < rule.urgent_threshold:
            return AlertLevel.URGENT
        if current_value < rule.warning_threshold:
            return AlertLevel.WARNING
        return AlertLevel.INFO

    # For "higher is worse" metrics
    if current_value >= rule.page_threshold:
        return AlertLevel.PAGE
    if current_value >= rule.urgent_threshold:
        return AlertLevel.URGENT
    if current_value >= rule.warning_threshold:
        return AlertLevel.WARNING
    return AlertLevel.INFO

04. Backup and Disaster Recovery

Backup Strategy

Backup Script

#!/bin/bash
# neo4j-backup.sh — daily backup with rotation

set -euo pipefail

BACKUP_DIR="/backups/neo4j"
RETENTION_DAYS=14
TIMESTAMP=$(date +%Y%m%d_%H%M%S)
BACKUP_NAME="neo4j_backup_${TIMESTAMP}"

echo "Starting backup: ${BACKUP_NAME}"

# Online backup (Enterprise Edition)
neo4j-admin database backup \
    --to-path="${BACKUP_DIR}" \
    --type=full \
    neo4j

# Compress
tar -czf "${BACKUP_DIR}/${BACKUP_NAME}.tar.gz" \
    -C "${BACKUP_DIR}" "${BACKUP_NAME}"

# Upload to cloud storage
aws s3 cp "${BACKUP_DIR}/${BACKUP_NAME}.tar.gz" \
    "s3://company-backups/neo4j/${BACKUP_NAME}.tar.gz"

# Clean up local backup directory
rm -rf "${BACKUP_DIR}/${BACKUP_NAME}"

# Rotate old backups
find "${BACKUP_DIR}" -name "*.tar.gz" \
    -mtime +${RETENTION_DAYS} -delete

echo "Backup complete: ${BACKUP_NAME}.tar.gz"

Disaster Recovery Plan

The last row is the ultimate safety net of the sidecar pattern. Because the relational database is the system of record, you can always rebuild the graph from scratch.

05. Scaling Patterns

Read Replicas

For read-heavy workloads, which graph databases almost always are:

                    ┌──────────────────┐
                    │   Application    │
                    └────────┬─────────┘
                             │
                    ┌────────▼─────────┐
                    │  Load Balancer   │
                    └──┬─────┬─────┬───┘
                       │     │     │
            ┌──────────▼┐ ┌─▼──────▼──┐
            │  Primary  │ │  Replica   │
            │  (writes) │ │  (reads)   │
            └───────────┘ └────────────┘

from neo4j import GraphDatabase, READ_ACCESS, WRITE_ACCESS


def create_driver_with_routing(
    primary_uri: str,
    user: str,
    password: str
) -> GraphDatabase.driver:
    """Create a driver that routes reads to replicas."""
    # Use neo4j:// protocol for automatic routing
    driver = GraphDatabase.driver(
        primary_uri.replace("bolt://", "neo4j://"),
        auth=(user, password),
        max_connection_pool_size=100,
        connection_acquisition_timeout=30
    )
    return driver


def read_query(driver, cypher: str, params: dict = None):
    """Execute a read query — routed to a replica."""
    with driver.session(
        default_access_mode=READ_ACCESS
    ) as session:
        result = session.run(cypher, params or {})
        return [dict(r) for r in result]


def write_query(driver, cypher: str, params: dict = None):
    """Execute a write query — always goes to primary."""
    with driver.session(
        default_access_mode=WRITE_ACCESS
    ) as session:
        result = session.run(cypher, params or {})
        return result.consume()

Connection Pooling

Connection pool exhaustion is a common production issue. Configure these settings before you go live:

When to Consider Sharding

Most graph databases, including Neo4j, do not natively shard like relational databases. Before you consider sharding, exhaust these options in order:

1. Index everything you query by. Most latency problems are missing indexes, not capacity problems.
2. Add read replicas. If reads are the bottleneck, add replicas.
3. Optimize queries. Profile with EXPLAIN and PROFILE in Cypher. Fix Cartesian products and unbounded variable-length paths.
4. Increase page cache. If the cache hit ratio is below 95%, give Neo4j more RAM.
5. Separate workloads. Run analytics queries on a replica and operational queries on the primary.

If you have exhausted all of these and still need more capacity, consider domain-based partitioning (separate graphs per tenant or business domain), Neo4j 5+ Fabric for cross-database queries, or purpose-built distributed graph databases (TigerGraph, Amazon Neptune, JanusGraph) for truly massive scale.

06. Cost Management

Track actual against estimated cost monthly. Set a budget alert at 80% of your estimate.

07. Runbook: Common Issues and Fixes

The 8 most common operational issues. Put this in your team's wiki where on-call engineers can find it at 3am.

Issue 1: Query Latency Spike

Symptoms: p95 latency jumps from 200ms to 5 seconds or more. Users report timeouts.

Diagnosis:

-- Find the slowest currently running queries
CALL dbms.listQueries() YIELD queryId, query, elapsedTimeMillis
WHERE elapsedTimeMillis > 5000
RETURN queryId, query, elapsedTimeMillis
ORDER BY elapsedTimeMillis DESC

Common causes and fixes:

1. Missing index: CREATE INDEX FOR (n:Label) ON (n.property)
2. Cartesian product: Add relationship patterns to connect disconnected MATCH clauses.
3. Supernode hit: Add degree filter WHERE size((n)--()) < 10000
4. Unbounded path: Change [:REL*] to [:REL*..5] with an explicit depth limit.

Issue 2: CDC Pipeline Stopped

Symptoms: Kafka consumer lag growing, graph data is stale.

Diagnosis:

# Check consumer group lag
kafka-consumer-groups.sh --bootstrap-server localhost:9092 \
    --group graph-sync-consumer --describe

# Check connector status
curl -s http://localhost:8083/connectors/pg-graph-connector/status

Common causes and fixes:

1. Consumer crashed: Restart consumer, check dead letter queue for failed events.
2. Connector lost: Restart Debezium connector, verify PostgreSQL replication slot exists.
3. Schema change: Update CDC handler for new or changed columns, replay from last committed offset.
4. Kafka disk full: Increase retention, add brokers, or reduce topic retention time.

Issue 3: Out of Memory

Symptoms: GC pauses over 1 second, heap usage above 95%, queries timing out.

Diagnosis:

# Check JVM heap usage
curl -s http://localhost:2004/metrics | grep heap

# Check for memory-hungry queries
CALL dbms.listQueries() YIELD query, allocatedBytes
ORDER BY allocatedBytes DESC LIMIT 5

Common causes and fixes:

1. Page cache too small: Increase server.memory.pagecache.size.
2. Heap too small: Increase server.memory.heap.max_size.
3. Query returning too much data: Add LIMIT, filter earlier in the query.
4. Graph outgrew instance: Upgrade instance size or add read replicas.

Issue 4: Data Inconsistency Between PostgreSQL and Neo4j

Symptoms: Application shows different data depending on which database it queries.

Diagnosis:

# Run the consistency check from Chapter 13
results = validate_sync_consistency(pg_conn, neo4j_session)
for check in results:
    if not check["pass"]:
        print(f"FAIL: {check}")

Common causes and fixes:

1. CDC lag: Wait for consumer to catch up and monitor lag.
2. Dropped events: Check dead letter queue, re-sync affected tables.
3. Handler bug: Fix the CDC handler, replay events from before the bug.
4. Schema drift: PostgreSQL column renamed but handler not updated.

Issue 5: Supernode Causing Slow Queries

Symptoms: Specific traversal queries suddenly slow after a data load.

Diagnosis:

-- Find nodes with highest degree
MATCH (n)
WITH n, size([(n)--() | 1]) AS degree
WHERE degree > 1000
RETURN labels(n) AS labels, n.name AS name, degree
ORDER BY degree DESC
LIMIT 20

Common causes and fixes:

1. Data modeling issue: Break the supernode into subgroups (by time period or category).
2. Query issue: Filter relationships by type or property before traversing from the supernode.
3. Legitimate high-degree node: Add a degree check in the query and handle high-degree nodes separately.

Issue 6: Backup Failure

Symptoms: Backup job exited with error. No recent backup available.

Diagnosis: Check backup job logs, verify disk space on backup target.

Fixes: Free space or increase the backup volume. Verify the backup user has correct filesystem permissions. Retry and verify cloud credentials for cloud backups.

Issue 7: Read Replica Lag

Symptoms: Read replica returns stale data, replication lag metrics climbing.

Fixes: Check bandwidth between primary and replica. Reduce query load on the replica or add another. Wait for any large transaction on the primary to complete. Restart the replica if lag does not recover.

Issue 8: Connection Pool Exhaustion

Symptoms: "Unable to acquire connection" errors in application logs.

Fixes: Increase pool size if too small for the workload. Fix connection leaks — ensure every session is closed with with blocks. Reduce query time to free connections faster. Add a connection acquisition timeout to fail fast rather than queue indefinitely.

08. Chapter Checklist

Before you move on, verify:

• You are monitoring query latency (p50/p95/p99), graph size, and pipeline health
• Alert thresholds are set and you have tested that alerts fire correctly
• Backup runs daily and you have tested restore at least once
• You have a disaster recovery plan that includes "rebuild from relational source"
• Connection pool is sized correctly and you monitor pool utilization
• The on-call team has a runbook they can find at 3am
• Cost monitoring is in place with budget alerts
• You have identified any supernodes and have queries that handle them gracefully