MongoDB Monitoring and Observability

researched 2026-05-28T15:30:15.471Z· 0 sources · 0 concepts · skill mongodb-monitoring-observability

Comprehensive reference for monitoring MongoDB deployments — from Atlas built-in dashboards through third-party integrations, CLI tools, and low-level FTDC diagnostics.

MongoDB Monitoring and Observability

Comprehensive reference for monitoring MongoDB deployments - from Atlas built-in dashboards through third-party integrations, CLI tools, and low-level FTDC diagnostics. [source]
When to use this skill: When answering questions about Atlas metrics, alert configuration, third-party monitoring integrations (Datadog, New Relic, Prometheus), FTDC diagnostics, slow query analysis, replication lag, connection pool behavior, or Atlas maintenance windows and planned operations. [source]
When not to use: For Atlas Search index tuning (use mongodb-search-ai), Atlas cost optimization (use mongodb-cost-optimization), or backup/restore planning (use mongodb-backup-restore). [source]
Required roles for most monitoring operations: clusterMonitor role on the admin database (self-managed), or Atlas Project Read Only / Project Data Access Read Only (Atlas UI). Third-party integrations (Datadog, Prometheus, New Relic) require Atlas Project Owner or Organization Owner to configure. [source]
Jump to: Quick Reference Tool Matrix [source]

1. Atlas Cloud Monitoring — Built-in Metrics and Dashboard Customization

Atlas provides real-time and historical metrics for every cluster tier M10 and above. Free/shared-tier clusters have reduced metric granularity (5-minute resolution vs. 1-minute for dedicated tiers). [source]

Key metric categories available in Atlas UI

Opcounters - insert, query, update, delete, getmore, command rates (ops/sec) [source]
CPU / System - process CPU, system CPU, I/O wait broken by read/write [source]
Memory - resident, virtual, mapped, cache (WiredTiger block cache, dirty bytes) [source]
Disk I/O - IOPS read/write, I/O utilization, disk queue depth [source]
Network - bytes in/out, number of requests [source]
Connections - current, available, total created [source]
Replication - oplog window hours, replication headroom, replication lag per secondary [source]
Query targeting - scanned/returned ratio (key indicator of missing indexes) [source]
Tickets - WiredTiger concurrent read/write tickets in use vs. available [source]

Dashboard customization

Atlas dashboards are pre-built per cluster but allow: [source]
Pin metric charts to a custom "Metrics" view for side-by-side comparison across nodes [source]
Toggle between individual node view (per-host) and cluster aggregate view [source]
Adjust time range (1h, 8h, 24h, 48h, 1w, custom) [source]
Use the Real-Time Performance Panel (RTPP) for 1-second granularity on live traffic - available on M10+ in the Atlas UI under the cluster's Real Time tab [source]
The RTPP shows: opcounters, read/write tickets, connections, network, logical size, and an interactive currentOp view showing the slowest in-flight operations per namespace. [source]

2. Ops Manager and Cloud Manager — Self-Managed Deployments

> Deep reference: see mongodb-ops-manager for full coverage of App DB sizing/HA, Backup Daemon placement, automation goal-state, air-gap/Local Mode, Kubernetes Operator, federation, and Live Migration to Atlas. This section covers the monitoring agent surface only. [source]
MongoDB Ops Manager is the on-premises deployment of MongoDB's management platform for teams running MongoDB in their own data centers or private clouds. MongoDB Cloud Manager is the hosted SaaS version of the same platform - it provides identical monitoring, automation, and backup capabilities without requiring you to host the Ops Manager application yourself. Both share the same agent architecture described below. [source]

Monitoring agent behavior

Runs as a daemon alongside your MongoDB processes [source]
Polls serverStatus, replSetGetStatus, dbStats, collStats, currentOp (filtered) at configurable intervals [source]
Stores time-series data in Ops Manager's own MongoDB backing store (separate from your application data) [source]
Sends alerts through Ops Manager's alert notification system - same alert types as Atlas [source]

Ops Manager / Cloud Manager dashboards

Both platforms replicate Atlas-style metric dashboards inside the web UI. The topology view shows replica set health, node states (PRIMARY/SECONDARY/ARBITER), and replication lag per member. The Hardware tab surfaces CPU, disk IOPS, and memory at host level for correlation with MongoDB behavior. [source]

Alert scope levels

Project-level alerts - apply to all clusters in a project (e.g., CPU > 80% on any node) [source]
Cluster-level alerts - scoped to a specific cluster [source]
Billing alerts - monthly spend thresholds, data transfer thresholds [source]

Alert tuning best practices

Set delay intervals (e.g., "notify if condition persists for 5 minutes") to suppress transient spikes - CPU can spike briefly during flushes without being actionable [source]
CPU alert baseline: M10–M30 should alert at 75%; M50+ with sustained IOPS-heavy workloads often benefit from 85% thresholds with short delay [source]
Replication lag: alert at 10–15 seconds for most OLTP workloads; 60 seconds for batch-heavy pipelines [source]
Oplog window: never let it drop below 4 hours; alert at 48 hours to give time to investigate before backup windows are at risk [source]
Connection count: alert at 80% of the cluster's maxIncomingConnections; calculate max from db.adminCommand({getCmdLineOpts:1}) or Atlas connection string parameters [source]

Atlas Custom Metrics

Atlas supports custom metric alerts via the Atlas Administration API (/api/atlas/v2/groups/{groupId}/alertConfigs). The metricName field accepts any metric Atlas exposes - including metrics not shown by default in the UI. Full metric name catalog: https://www.mongodb.com/docs/atlas/reference/alert-conditions/ [source]

$currentOp polling for application-level insight

For application-level custom metrics, poll $currentOp on a schedule. Note: run this query from an admin-context connection - the $all field was deprecated in MongoDB 4.0 and removed in favor of the admin-context currentOp command directly: [source]
Key fields: secs_running, op, ns, planSummary, waitingForLock, msg, locks. [source]

Application-level metrics to track

Query latency percentiles (p50, p95, p99) per collection [source]
Error rates by MongoDB error code [source]
Connection pool waitQueueSize - rising queue = pool exhaustion signal [source]
Retry attempt counts - spike in retries indicates transient elections or network partitions [source]

Setup

Atlas Datadog integration requires M10+ clusters and a Datadog API key. Configure via Atlas UI: Project → Integrations → Datadog. Select region (US1, US3, US5, EU1, AP1, US1_FED) to match your Datadog account region. [source]

Datadog Database Monitoring (DBM) for Atlas

Separate from the metrics integration - requires Datadog Agent with MongoDB integration. Provides query-level explain plan capture, wait event analysis, query normalization and fingerprinting. Configure via conf.d/mongo.d/conf.yaml with a clusterMonitor role user. [source]

6. New Relic Integration

Configure via Atlas UI: Project → Integrations → New Relic. Metrics ship under MongoDBAtlas.* namespace. Primary value: APM-to-database correlation - New Relic links slow transaction traces in application code directly to slow MongoDB operations when using the New Relic APM agent. [source]
For self-managed MongoDB, use nri-mongodb with the New Relic Infrastructure agent (EXTENDED_METRICS: true, COLLECTION_METRICS: true). [source]

Atlas managed endpoint (M10+ only)

Enable via Atlas UI: Project → Integrations → Prometheus. Scrape URL: https://cloud.mongodb.com/prometheus/v1.0/groups/{groupId}/metrics. Auth: HTTP Basic with Atlas programmatic API key pair. [source]

Self-managed Prometheus

Use mongodb_exporter (Percona) on port 9216. Search "MongoDB Overview Percona" in the Grafana dashboard library for a production-ready starting point. [source]

8. FTDC (Full Time Diagnostic Capture)

FTDC is MongoDB's always-on internal diagnostic system (enabled by default since MongoDB 3.2). It is the first artifact MongoDB Support requests for any performance investigation. [source]
Samples every second: full serverStatus, replSetGetStatus, oplog metadata, system CPU/memory, WiredTiger internal stats. [source]
Samples every 200ms: lighter CPU/I/O subset for sub-second spike reconstruction. [source]
Location: <dbPath>/diagnostic.data/ - files rotate at ~10 MB. Atlas retains FTDC automatically; for self-managed, copy the entire directory while mongod is live (safe - FTDC uses its own write path). [source]

FTDC diagnostic questions

Checkpoint stall? → WiredTiger checkpoint duration spike [source]
CPU saturated? → system CPU counters at 100% [source]
Connection spike before incident? → connections.current time series [source]
Replication lag gradual or sudden? → replSetGetStatus.members[].optimeDate delta [source]
Cache eviction pressure? → cache dirty % over time [source]

mongostat

Key columns: insert/query/update/delete (ops/sec), dirty (WT dirty cache %), used (WT cache %), qrw/arw (queue/active read-write), conn, repl. [source]
When to use: quick snapshot of server load; real-time cache utilization; spotting queue buildup. [source]

mongotop

Shows per-collection total/read/write ms per interval. When to use: identify hottest collection during a performance issue. [source]

db.currentOp()

When to use: real-time slow op investigation; finding lock waiters (waitingForLock: true). [source]

Atlas Profiler and Performance Advisor

Cluster → Profiler tab: near-real-time slow queries (~2 min pipeline latency) [source]
Cluster → Performance Advisor: automatic index recommendations ranked by avg execution time × frequency [source]
Default slow threshold: 100ms (configurable to 0ms) [source]

system.profile

Key fields: millis, planSummary (IXSCAN vs COLLSCAN), keysExamined, docsExamined, queryHash. Caution: profiling level 2 has measurable overhead - use level 1 with tuned slowms in production. [source]

Root causes

Secondary under-resourced (upgrade tier or distribute reads) [source]
Flow control (MongoDB 4.2+) - check replSetGetStatus.flowControl.isLagged [source]
Chained replication - check rs.status().syncSourceHost [source]
Long-running transactions on secondary [source]
Network partition/bandwidth saturation [source]

Atlas connection limits by tier

Connections are per-node. A 3-node M30 replica set has 9,000 total across all nodes. [source]

Tuning recommendations

Single MongoClient per process (most common leak: new client per request) [source]
Lambda/serverless: maxPoolSize=5–10, maxIdleTimeMS=60000 [source]
Enable waitQueueTimeoutMS to surface exhaustion quickly rather than hanging [source]

Free and shared tier clusters (M0, M2, M5)

M0, M2, and M5 clusters do not support configurable maintenance windows. Atlas manages all maintenance entirely, with no operator control over timing. These clusters may be restarted at any time. Upgrade to M10 or higher for maintenance window control. [source]
This is a common point of confusion - the project-level maintenance window setting applies only to dedicated-tier clusters (M10+). [source]

Maintenance window configuration

Atlas maintenance windows are configured at the project level and apply to all dedicated-tier (M10+) clusters within that project. [source]
Location: Atlas UI → Project Settings → Maintenance Window [source]
Default behavior: When no custom window is configured, Atlas selects the window (commonly Tuesday 10:00–12:00 UTC for many regions). Configure an explicit window aligned with your lowest-traffic period for production workloads. [source]
Configuring a custom window: [source]
Choose day of week (Sunday through Saturday; Sunday=1 in the API/CLI, matching the integer table below) [source]
Choose start hour in UTC (0–23); the window is exactly 1 hour [source]
Changes take effect immediately and persist until cleared [source]
Important scope limitation: Project-scoped, not per-cluster. To set different windows for dev vs. prod clusters, place them in separate Atlas projects. [source]
Atlas CLI commands: [source]
Day-of-week values: Sunday=1, Monday=2, Tuesday=3, Wednesday=4, Thursday=5, Friday=6, Saturday=7. [source]

What triggers maintenance

Emergency security patches bypass the maintenance window entirely. Atlas notifies project and organization owners via email, but the window configuration does not constrain it. [source]

How Atlas performs rolling maintenance

Secondaries first - one at a time, waiting for each to rejoin and catch up before proceeding. [source]
Primary last - triggers a replica set election. [source]
Election window - typically 10–30 seconds; writes temporarily unavailable, reads fall back to secondaries. [source]
mongos nodes (sharded clusters only) - restarted last. Skip for replica-set-only deployments. [source]
Application impact: Drivers with retryable writes handle the election transparently. Applications without retryable writes may see one transient write failure. [source]
Alert during maintenance: The "Primary election" alert fires during every maintenance restart. Configure a lower-urgency channel for this alert type or correlate it with the maintenance window time. [source]
Total duration: 3-node replica set: 5–15 min. Sharded clusters: multiply per-shard restart time by shard count (30–60 min for large topologies). [source]

Deferring maintenance

Deferral postpones by exactly 7 days, once only per scheduled event [source]
After one deferral, maintenance executes at the rescheduled time - no further deferral [source]
Critical security patches cannot be deferred - attempting to defer returns an error [source]

Querying the maintenance window via API

Response fields: dayOfWeek (1–7, absent if no custom window), hourOfDay (0–23 UTC), startASAP (maintenance queued for next opportunity), autoDeferOnceEnabled. [source]

Emergency and critical security patches

Same-day or next-day notice for critical CVEs; 24–48 hours for lower-severity updates [source]
Notifications sent to all Project Owners and Organization Owners [source]
Cannot be deferred; rolling restart procedure still used to minimize impact [source]
Monitor Activity Feed (Atlas UI → Project → Activity) for maintenance start/completion timestamps [source]

Minimizing application impact

Why serverSelectionTimeoutMS=30000 matters: With 3000ms, a 12-second election causes ServerSelectionTimeoutError before the new primary is elected. [source]
Post-election warm-up: First queries to the new primary may be slower (30–60s) while connections re-establish and the WiredTiger cache warms. [source]

Atlas maintenance for sharded clusters

Sequence: config server replica set (CSRS) → shard replica sets (sequentially) → mongos routers (parallel, stateless). [source]
Balancer: suspended during maintenance; in-progress migrations complete, no new ones start. [source]
Duration estimate: per-shard restart time × shard count + CSRS + mongos. A 4-shard cluster at 10 min/shard ≈ 40–50 min for shards alone. [source]

Customer communication template

Replace all [bracketed] placeholders before sending. [source]
Placeholder guide: [HH:MM]–[HH:MM] UTC = configured 1-hour window (e.g., "02:00–03:00 UTC"). Always include: specific UTC time window, < 30s impact (not full restart duration), explicit no-data-loss statement, retryable writes note, escalation path. [source]

Children

Atlas Metrics and Dashboards (frontier)
Atlas Alerts (frontier)
Datadog Integration (frontier)
Prometheus Integration (frontier)
New Relic Integration (frontier)
FTDC Diagnostics (frontier)
Slow Query Monitoring (frontier)
Replication Lag Monitoring (frontier)
Connection Metrics (frontier)
Atlas Maintenance Windows

Frontier under this node: Atlas Alerts, Atlas Metrics and Dashboards, Connection Metrics, Datadog Integration, FTDC Diagnostics, New Relic Integration, Prometheus Integration, Replication Lag Monitoring, Slow Query Monitoring

← the whole tree · 3D view· how to read this page