Kubernetes

There is a configuration that appears in virtually every Kubernetes cluster: a HorizontalPodAutoscaler targeting 70% CPU utilization and 70% memory utilization. It looks reasonable. It follows the examples in the official documentation. And in many cases, it silently causes more harm than good.

Related reading: Kubernetes HPA on memory: working examples and troubleshooting.

Related reading: liveness probe anti-patterns.

Related reading: node autoscaling with Cluster Autoscaler or Karpenter.

The problems surface in predictable ways: workloads that do nothing get scaled up because their memory footprint is naturally high. Latency-sensitive APIs scale too slowly because the CPU spike is already over by the time new pods are ready. Batch jobs oscillate between scaling up and down during normal operation. And teams spend hours debugging autoscaling behavior that should have been straightforward.

This article is about understanding why the default HPA configuration fails, under which exact conditions memory-based HPA is appropriate (and when it is not), and what alternative metrics — custom metrics, event-driven triggers, and external signals — produce autoscaling behavior that actually matches workload demand.

How HPA Actually Decides to Scale

Before diagnosing the problems, it is worth understanding the mechanics precisely. HPA computes a desired replica count using this formula:

desiredReplicas = ceil(currentReplicas × (currentMetricValue / desiredMetricValue))

For a CPU target of 70%, with 2 replicas currently consuming an average of 140% of their CPU request, HPA computes ceil(2 × (140 / 70)) = 4 replicas. This is conceptually simple but has a critical dependency that most configurations ignore: the metric value is expressed relative to the resource request, not the resource limit.

This distinction is fundamental to understanding every failure mode that follows. If a container has a CPU request of 100m and a limit of 2000m, and it is currently consuming 80m, HPA sees 80% utilization — even though the container is using only 4% of its allowed ceiling. Set an HPA threshold of 70% on a container with a CPU request of 100m and any nontrivial workload will trigger scaling immediately.

The HPA controller polls metrics every 15 seconds by default (--horizontal-pod-autoscaler-sync-period). Scale-up happens quickly — within one to three polling cycles when the threshold is consistently exceeded. Scale-down is deliberately slow: by default the controller waits 5 minutes (--horizontal-pod-autoscaler-downscale-stabilization) before reducing replicas, to avoid thrashing. This asymmetry matters when debugging oscillation.

CPU-Based HPA: When It Works and When It Doesn’t

CPU is a compressible resource. When a container hits its CPU limit, the kernel throttles it — the process slows down but does not crash or get evicted. This property makes CPU a reasonable proxy for load in many, but not all, scenarios.

Where CPU HPA Works Well

Stateless request-processing workloads are the sweet spot for CPU-based HPA. If your service does CPU-bound work per request — REST APIs performing data transformation, compute-heavy business logic, image processing — then CPU utilization correlates strongly with request volume. More requests means more CPU consumed, which means HPA adds replicas, which distributes the load.

The key prerequisites for CPU HPA to work correctly are:

• Accurate CPU requests. Set requests to the actual sustained consumption of the workload under normal load, not a low placeholder. Use VPA in recommendation mode or historical Prometheus data to right-size requests before enabling HPA.
• Reasonable request-to-limit ratio. A ratio of 1:4 or less keeps HPA thresholds meaningful. A container with request 100m and limit 4000m makes percentage-based thresholds nearly useless.
• CPU consumption that tracks user load linearly. If your service does CPU-heavy background work independent of incoming requests, CPU utilization will trigger scaling regardless of actual demand.

Where CPU HPA Fails

Latency-sensitive services with sharp traffic spikes. HPA reacts to average CPU utilization measured over the polling window. For a service that handles traffic bursts — a flash sale, a cron-triggered batch of API calls, a notification broadcast — by the time the HPA controller detects the spike, queues new pods, and those pods pass readiness checks, the burst may already be over. The result is replicas added after the damage is done, with the added cost of a scale-down cycle afterward.

I/O-bound workloads. A service that spends most of its time waiting on database queries, external API calls, or message queue reads will show low CPU utilization even under heavy load. HPA will not add replicas while the service is degraded — it sees idle CPUs while goroutines or threads are blocked waiting on I/O.

Workloads with cold-start costs. If a new replica takes 30-60 seconds to warm up (loading ML models, establishing connection pools, populating caches), scaling decisions need to happen earlier — before CPU peaks — not in reaction to it.

Memory-Based HPA: Why It Almost Always Breaks

Memory is an incompressible resource. Unlike CPU — which can be throttled without killing a process — when a container exhausts its memory limit, the OOM killer terminates it. This single property cascades into a set of fundamental problems with using memory as an HPA trigger.

The Core Problem: Memory Doesn’t Naturally Correlate With Load

For most well-architected services, memory consumption is relatively stable. A Go service allocates memory at startup for its runtime structures, connection pools, and caches — and then maintains roughly that footprint regardless of traffic. A JVM application allocates a heap at startup and uses garbage collection to manage it. In both cases, memory usage under 10 requests per second and under 10,000 requests per second may be nearly identical.

This means a memory-based HPA with a 70% threshold will either:

• Never trigger, because the workload’s memory is stable and always below the threshold — rendering the HPA useless.
• Always trigger, because the workload’s baseline memory consumption is naturally above the threshold — causing the workload to scale out permanently and never scale back in.

Neither outcome corresponds to actual scaling need.

The Request Misconfiguration Trap

This is the failure mode the user mentioned, and it is the most common cause of “my workload scales up for no reason.” Consider a Java service that needs 512Mi of heap to run normally. The team sets memory request to 256Mi — too conservative, either to save cost or because the initial estimate was wrong. The service immediately consumes 200% of its memory request just by being alive. An HPA with a 70% memory target will scale this workload to maximum replicas within minutes of deployment, and it will stay there forever.

The fix is never “adjust the HPA threshold.” The fix is right-sizing the memory request. But this reveals the deeper issue: memory-based HPA is extremely sensitive to the accuracy of your resource requests, and most teams do not have accurate requests — especially for newer workloads or after code changes that alter memory footprint.

JVM and Go Runtime Memory Behavior

JVM workloads are particularly problematic. By default, the JVM allocates heap up to a maximum (-Xmx) and then holds that memory — it does not release heap back to the OS aggressively, even after garbage collection. A JVM service that handles one request per hour will show nearly the same memory footprint as one handling thousands of requests per minute. Furthermore, the JVM’s garbage collector introduces memory spikes during collection cycles that are unrelated to load.

In containerized JVM environments, you also need to account for the container memory limit aware flag (-XX:+UseContainerSupport, enabled by default since JDK 11) which affects how the JVM calculates its heap ceiling relative to the container limit. Without proper tuning, the JVM may allocate a heap that fills 80-90% of the container’s memory limit — immediately triggering any memory-based HPA.

Go workloads behave differently but also poorly with memory HPA. Go’s garbage collector is designed to maintain low latency rather than minimal memory use. The runtime may hold memory above what is strictly needed, and the memory footprint can vary based on GC tuning parameters (GOGC, GOMEMLIMIT) in ways that are not correlated with incoming request load.

When Memory HPA Is Actually Appropriate

There are narrow cases where memory-based HPA makes sense:

• Workloads where memory consumption genuinely tracks with load linearly. Some data processing pipelines, in-memory caches that grow with request volume, or streaming applications that buffer data proportionally to throughput. If you can demonstrate from metrics that memory and load have a strong linear correlation, memory HPA is defensible.
• As a safety valve alongside CPU HPA. Using memory as a secondary metric (not primary) to protect against memory leaks or runaway allocations in a service that normally scales on CPU. In this case, set the memory threshold high — 85-90% — so it only triggers in genuine overconsumption scenarios.
• Caching services where eviction is not desirable. If a service uses memory as a performance cache and you want to scale out before memory pressure causes cache eviction, memory utilization can be a useful trigger — provided requests are accurately sized.

Outside these specific cases, removing memory from your HPA spec and relying on the signals below will produce better behavior in virtually every scenario.

Kubernetes HPA Memory Example: Complete Working Manifests

The two manifests people actually search for, ready to apply. First, an HPA scaling on memory only — appropriate for workloads whose memory tracks load and is released when load drops (queues, caches, most Go/Rust services; not JVM heaps):

apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: worker-memory-hpa
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: queue-worker
  minReplicas: 2
  maxReplicas: 10
  metrics:
  - type: Resource
    resource:
      name: memory
      target:
        type: Utilization
        averageUtilization: 75    # % of the pod's memory REQUEST
  behavior:
    scaleDown:
      stabilizationWindowSeconds: 300

Second, the more common production shape: CPU and memory combined. The HPA computes a desired replica count for each metric and applies the highest — memory pressure can hold replicas up even when CPU is idle:

apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: api-cpu-memory-hpa
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: api-server
  minReplicas: 3
  maxReplicas: 20
  metrics:
  - type: Resource
    resource:
      name: cpu
      target:
        type: Utilization
        averageUtilization: 65
  - type: Resource
    resource:
      name: memory
      target:
        type: Utilization
        averageUtilization: 80
  behavior:
    scaleUp:
      stabilizationWindowSeconds: 0
      policies:
      - type: Percent
        value: 100        # at most double the replicas...
        periodSeconds: 60 # ...per minute
    scaleDown:
      stabilizationWindowSeconds: 300
      policies:
      - type: Pods
        value: 2          # drop at most 2 pods...
        periodSeconds: 120 # ...every 2 minutes

Three details that make or break these examples in practice. Utilization is computed against requests, not limits — a pod requesting 512Mi at 75% target scales when average usage crosses ~384Mi, regardless of the limit. Both metrics must be resolvable: if metrics-server cannot read memory for any pod (init containers, sidecars without requests), the HPA marks the metric invalid and logs FailedGetResourceMetric. And memory-based scale-down only works if your runtime returns memory to the OS — JVM workloads typically hold their heap, so replicas scale up under pressure and never come back down; for those, scale on CPU or a custom metric instead.

Right-Sizing Requests Before You Add HPA

No HPA strategy works correctly without accurate resource requests. Before adding any autoscaler — CPU, memory, or custom metrics — run your workload under representative load and measure actual consumption. The easiest way to do this is with VPA in recommendation mode:

apiVersion: autoscaling.k8s.io/v1
kind: VerticalPodAutoscaler
metadata:
  name: my-service-vpa
spec:
  targetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: my-service
  updatePolicy:
    updateMode: "Off"   # Recommendation only — don't auto-apply

After 24-48 hours of traffic, check the VPA recommendations:

kubectl describe vpa my-service-vpa

The lowerBound, target, and upperBound values give you a data-driven baseline for setting requests. Set your requests at or near the VPA target value before configuring HPA. This single step eliminates the most common cause of HPA misbehavior.

Note that VPA and HPA cannot both manage the same resource metric simultaneously. If VPA is set to auto-update CPU or memory, and HPA is also scaling on those metrics, the two controllers will fight each other. The safe combination is: HPA on CPU/memory + VPA in recommendation-only mode, or HPA on custom metrics + VPA on CPU/memory in auto mode. See the Kubernetes VPA guide for the full details.

Better Signals: What to Scale On Instead

The fundamental shift is moving from resource consumption metrics (which describe the past) to demand metrics (which describe what the workload is being asked to do right now or will be asked to do in seconds).

Requests Per Second (RPS)

For HTTP services, requests per second per replica is usually the most accurate proxy for load. Unlike CPU, it measures demand directly — not a side-effect of demand. An HPA that maintains 500 RPS per replica will scale predictably as traffic grows, regardless of whether the service is CPU-bound, memory-bound, or I/O-bound.

RPS is available as a custom metric from your service mesh (Istio exposes it as istio_requests_total), from your ingress controller (NGINX exposes request rates via Prometheus), or from your application’s own Prometheus metrics. Configuring HPA on custom metrics requires the Prometheus Adapter or a compatible custom metrics API implementation.

apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: my-service-hpa
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: my-service
  minReplicas: 2
  maxReplicas: 20
  metrics:
  - type: Pods
    pods:
      metric:
        name: http_requests_per_second
      target:
        type: AverageValue
        averageValue: "500"   # 500 RPS per replica

Queue Depth and Lag

For consumer workloads — services reading from Kafka, RabbitMQ, SQS, or any message queue — the right scaling signal is consumer lag: how many messages are waiting to be processed. A lag of zero means consumers are keeping up; a growing lag means you need more consumers.

CPU will not give you this signal reliably. A consumer blocked on a slow database write will show low CPU but growing lag. An idle consumer will show low CPU even if the queue contains millions of unprocessed messages. Scaling on lag directly solves both problems.

This is precisely the use case that KEDA was built for. KEDA’s Kafka scaler, for example, reads consumer group lag directly and scales replicas to maintain a configurable lag threshold — no custom metrics pipeline required.

Latency

P99 latency per replica is an excellent scaling signal for latency-sensitive services. If your SLO is a 200ms P99 response time and latency starts climbing toward 400ms, that is a direct signal that the service is overloaded — regardless of what CPU or memory shows.

Latency-based autoscaling requires custom metrics from your service mesh or APM tool, but the added complexity is often justified for user-facing APIs where latency directly impacts experience.

Scheduled and Predictive Scaling

For workloads with predictable traffic patterns — business-hours services, weekly batch jobs, end-of-month processing peaks — proactive scaling outperforms reactive scaling by definition. Rather than waiting for CPU to spike and then scrambling to add replicas, you pre-scale before the expected load increase.

KEDA’s Cron scaler enables this pattern declaratively, defining scale rules based on time windows rather than observed metrics.

HPA Configuration Best Practices

Always Set minReplicas ≥ 2 for Production

A minReplicas: 1 HPA means your service has a single point of failure during scale-in events. When HPA scales down to 1 replica and that pod is evicted for node maintenance, your service has zero available instances for the duration of the new pod’s startup time. For any production workload, set minReplicas: 2 as a baseline.

Tune Stabilization Windows

The default 5-minute scale-down stabilization window is too aggressive for many workloads. A service that processes jobs in 3-minute batches will show a predictable CPU trough between batches — HPA will attempt to scale down, only to scale back up when the next batch arrives. Increase the stabilization window to match your workload’s natural cycle:

apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: my-service-hpa
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: my-service
  minReplicas: 2
  maxReplicas: 20
  metrics:
  - type: Resource
    resource:
      name: cpu
      target:
        type: Utilization
        averageUtilization: 60
  behavior:
    scaleDown:
      stabilizationWindowSeconds: 600   # 10 minutes
      policies:
      - type: Percent
        value: 25                        # Scale down max 25% of replicas at once
        periodSeconds: 60
    scaleUp:
      stabilizationWindowSeconds: 0     # Scale up immediately
      policies:
      - type: Percent
        value: 100
        periodSeconds: 30

The behavior block (available in HPA v2, GA since Kubernetes 1.23) gives you independent control over scale-up and scale-down behavior. Aggressive scale-up with conservative scale-down is the right default for most production services.

Use a Lower CPU Threshold Than You Think

A CPU target of 70% sounds like it leaves headroom, but it does not account for the time required to scale. If your service takes 45 seconds to pass readiness checks after a new pod starts, and you scale at 70% CPU, the existing pods will be at 100%+ CPU (throttled) for 45 seconds before relief arrives. Set CPU targets at 50-60% for services where scale-up latency matters. This keeps more headroom available during the scaling reaction window.

Combine HPA with PodDisruptionBudgets

HPA scale-down terminates pods. Without a PodDisruptionBudget, HPA can terminate multiple replicas simultaneously during a scale-down event, potentially taking your service below its minimum healthy instance count during cluster maintenance. Always pair an HPA with a PDB:

apiVersion: policy/v1
kind: PodDisruptionBudget
metadata:
  name: my-service-pdb
spec:
  minAvailable: "50%"
  selector:
    matchLabels:
      app: my-service

Don’t Mix VPA Auto-Update with HPA on the Same Metric

If VPA is set to auto-update CPU or memory requests, and HPA is also scaling on CPU or memory utilization, you create a control loop conflict. VPA changes the request (the denominator of the utilization calculation), which immediately changes the apparent utilization, which triggers HPA to change replica count, which changes the per-pod load, which triggers VPA again. Use VPA in Off or Initial mode when HPA is managing the same workload on resource metrics.

Decision Framework: Which Autoscaler for Which Workload

Use this as a starting point when configuring autoscaling for a new workload:

Going Beyond HPA: KEDA and Custom Metrics

Once you outgrow what HPA v2 can express — particularly for event-driven architectures, external system triggers, or composite scaling conditions — KEDA provides a Kubernetes-native autoscaling framework that extends the HPA model without replacing it.

KEDA works by implementing a custom metrics API that HPA can consume, plus its own ScaledObject CRD that abstracts the configuration of over 60 built-in scalers: Kafka, RabbitMQ, Azure Service Bus, AWS SQS, Prometheus queries, Datadog metrics, HTTP request rate, and more. The important architectural point is that KEDA does not replace HPA — it feeds it. Under the hood, KEDA creates and manages an HPA resource targeting the scaled deployment. You get HPA’s stabilization windows, replica bounds, and Kubernetes-native behavior, driven by signals that HPA itself cannot access natively.

For a detailed walkthrough of KEDA scalers and real-world event-driven patterns, see Event-Driven Autoscaling in Kubernetes with KEDA.

For workloads where the right scaling signal comes from a Prometheus metric — request rates, custom business metrics, queue sizes exposed via exporters — the Kubernetes Autoscaling 1.26 and HPA v2 article covers how the custom metrics API pipeline works and how changes in Kubernetes 1.26 affected KEDA behavior.

Frequently Asked Questions

Further Reading

• Event-Driven Autoscaling with KEDA — queue depth, Kafka lag, and external triggers as scaling signals
• Kubernetes VPA Explained — right-sizing resource requests before enabling HPA
• HPA v2 and Kubernetes Autoscaling 1.26 — custom metrics API, KEDA integration, and behavioral changes
• Kubernetes HPA Official Documentation
• KEDA Concepts — Official Docs

Every production Kubernetes cluster talks to the outside world. Your services call payment APIs, connect to managed databases, push events to SaaS analytics platforms, and reach legacy systems that will never run inside the mesh. By default, Istio lets all outbound traffic flow freely — or blocks it entirely if you flip outboundTrafficPolicy to REGISTRY_ONLY. Neither extreme gives you what you actually need: selective, observable, policy-controlled access to external services.

That is exactly what Istio ServiceEntry solves. It registers external endpoints in the mesh’s internal service registry so that Envoy sidecars can apply the same traffic management, security, and observability features to outbound calls that you already enjoy for east-west traffic. No new proxies, no egress gateways required for the basic case — just a YAML resource that tells the mesh “this external thing exists, and here is how to reach it.”

In this guide, I will walk through every field of the ServiceEntry spec, explain the four DNS resolution modes with real-world use cases, and show production-ready patterns for external APIs, databases, TCP services, and legacy workloads. We will also cover how to combine ServiceEntry with DestinationRule and VirtualService to get circuit breaking, retries, connection pooling, and even sticky sessions for external dependencies.

What Is a ServiceEntry

Istio maintains an internal service registry that merges Kubernetes Services with any additional entries you declare. When a sidecar proxy needs to decide how to route a request, it consults this registry. Services inside the mesh are automatically registered. Services outside the mesh are not — unless you create a ServiceEntry.

A ServiceEntry is a custom resource that adds an entry to the mesh’s service registry. Once registered, the external service becomes a first-class citizen: Envoy generates clusters, routes, and listeners for it, which means you get metrics (istio_requests_total), access logs, distributed traces, mTLS origination, retries, timeouts, circuit breaking — the full Istio feature set.

Without a ServiceEntry, outbound traffic to an external host either passes through as a raw TCP connection (in ALLOW_ANY mode) with no telemetry, or gets dropped with a 502/503 (in REGISTRY_ONLY mode). Both outcomes are undesirable in production. The ServiceEntry bridges that gap.

ServiceEntry Anatomy: All Fields Explained

Let us look at a complete ServiceEntry and then break down each field.

apiVersion: networking.istio.io/v1
kind: ServiceEntry
metadata:
  name: external-api
  namespace: production
spec:
  hosts:
    - api.stripe.com
  location: MESH_EXTERNAL
  ports:
    - number: 443
      name: https
      protocol: TLS
  resolution: DNS
  exportTo:
    - "."
    - "istio-system"

hosts

A list of hostnames associated with the service. For external services, this is typically the DNS name your application uses (e.g., api.stripe.com). For services using HTTP protocols, the hosts field is matched against the HTTP Host header. For non-HTTP protocols and services without a DNS name, you can use a synthetic hostname and pair it with addresses or static endpoints.

addresses

Optional virtual IP addresses associated with the service. Useful for TCP services where you want to assign a VIP that the sidecar will intercept. Not required for HTTP/HTTPS services that use hostname-based routing.

ports

The ports on which the external service is exposed. Each port needs a number, name, and protocol. The protocol matters: setting it to TLS tells Envoy to perform SNI-based routing without terminating TLS. Setting it to HTTPS means HTTP over TLS. For databases, you’ll typically use TCP.

location

MESH_EXTERNAL or MESH_INTERNAL. Use MESH_EXTERNAL for services outside your cluster (third-party APIs, managed databases). Use MESH_INTERNAL for services inside your infrastructure that are not part of the mesh — for example, VMs running in the same VPC that do not have a sidecar, or a Kubernetes Service in a namespace without injection enabled. The location affects how mTLS is applied and how metrics are labeled.

resolution

How the sidecar resolves the endpoint addresses. This is the most critical field and I will dedicate the next section to it. Options: NONE, STATIC, DNS, DNS_ROUND_ROBIN.

endpoints

An explicit list of network endpoints. Required when resolution is STATIC. Optional with DNS resolution to provide labels or locality information. Each endpoint can have an address, ports, labels, network, locality, and weight.

exportTo

Controls the visibility of this ServiceEntry across namespaces. Use "." for the current namespace only, "*" for all namespaces. In multi-team clusters, restrict exports to avoid namespace pollution.

Resolution Types: NONE vs STATIC vs DNS vs DNS_ROUND_ROBIN

The resolution field determines how Envoy discovers the IP addresses behind the service. Getting this wrong is the number one cause of ServiceEntry misconfigurations. Here is a clear breakdown.

When to Use NONE

Use NONE when you want to register a range of external IPs or wildcard hosts without Envoy performing any address resolution. This is common for broad egress policies: “allow traffic to *.googleapis.com on port 443.” Envoy will simply forward traffic to whatever IP the application resolved via kube-dns. The downside: Envoy has limited ability to apply per-endpoint policies.

When to Use STATIC

Use STATIC when the external service has known, stable IP addresses that rarely change. This avoids DNS dependencies entirely. You define the IPs in the endpoints list. Classic use case: a legacy Oracle database on a fixed IP in your data center.

When to Use DNS

Use DNS for most external API integrations. Envoy performs asynchronous DNS resolution and creates a cluster endpoint for each returned IP address. This enables per-endpoint health checking and circuit breaking — critical for production reliability. This is the mode you want for services like api.stripe.com or your RDS instance endpoint.

When to Use DNS_ROUND_ROBIN

Use DNS_ROUND_ROBIN when the external hostname returns many IPs and you do not need per-IP circuit breaking. Envoy treats all resolved IPs as a single logical endpoint and round-robins across them. This is lighter weight than DNS mode and avoids creating a large number of endpoints in Envoy’s cluster configuration.

Practical Patterns

Pattern 1: External HTTP API (api.stripe.com)

The most common ServiceEntry pattern. Your application calls a third-party HTTPS API. You want Istio telemetry, and optionally retries and timeouts.

apiVersion: networking.istio.io/v1
kind: ServiceEntry
metadata:
  name: stripe-api
  namespace: payments
spec:
  hosts:
    - api.stripe.com
  location: MESH_EXTERNAL
  ports:
    - number: 443
      name: tls
      protocol: TLS
  resolution: DNS

Note the protocol is TLS, not HTTPS. Since your application initiates the TLS handshake directly, Envoy handles this as opaque TLS using SNI-based routing. If you were terminating TLS at the sidecar and doing TLS origination via a DestinationRule, you would set the protocol to HTTP and handle the upgrade separately — but for most external APIs, let the application manage its own TLS.

Pattern 2: External Managed Database (RDS / CloudSQL)

Managed databases expose a DNS endpoint that resolves to one or more IPs. During failover, the DNS record changes. You need Envoy to respect DNS TTLs and route to the current primary.

apiVersion: networking.istio.io/v1
kind: ServiceEntry
metadata:
  name: orders-database
  namespace: orders
spec:
  hosts:
    - orders-db.abc123.us-east-1.rds.amazonaws.com
  location: MESH_EXTERNAL
  ports:
    - number: 5432
      name: postgres
      protocol: TCP
  resolution: DNS

For TCP services, Envoy cannot use HTTP headers to route, so it relies on IP-based matching. The DNS resolution mode ensures Envoy periodically re-resolves the hostname and updates its endpoint list. This is critical for RDS multi-AZ failover scenarios where the DNS endpoint flips to a new IP.

Pattern 3: Legacy Internal Service Not in the Mesh

You have a monitoring service running on a set of VMs at known IP addresses inside your VPC. It is not part of the mesh, but your meshed services need to talk to it.

apiVersion: networking.istio.io/v1
kind: ServiceEntry
metadata:
  name: legacy-monitoring
  namespace: observability
spec:
  hosts:
    - legacy-monitoring.internal
  location: MESH_INTERNAL
  ports:
    - number: 8080
      name: http
      protocol: HTTP
  resolution: STATIC
  endpoints:
    - address: 10.0.5.10
    - address: 10.0.5.11
    - address: 10.0.5.12

Key differences: location is MESH_INTERNAL because the service lives inside your network, and resolution is STATIC because we know the IPs. The hostname legacy-monitoring.internal is synthetic — your application uses it, and Istio’s DNS proxy (or a CoreDNS entry) resolves it to one of the listed endpoints.

Pattern 4: TCP Services with Multiple Ports

Some external services expose multiple TCP ports — for example, an Elasticsearch cluster with both data (9200) and transport (9300) ports.

apiVersion: networking.istio.io/v1
kind: ServiceEntry
metadata:
  name: external-elasticsearch
  namespace: search
spec:
  hosts:
    - es.example.com
  location: MESH_EXTERNAL
  ports:
    - number: 9200
      name: http
      protocol: HTTP
    - number: 9300
      name: transport
      protocol: TCP
  resolution: DNS

Each port gets its own Envoy listener configuration. The HTTP port benefits from full Layer 7 telemetry and traffic management. The TCP port gets Layer 4 metrics and connection-level policies.

Combining ServiceEntry with DestinationRule

A ServiceEntry alone registers the external service. To apply traffic policies — connection pooling, circuit breaking, TLS origination, load balancing — you pair it with a DestinationRule. This is where things get powerful.

Connection Pooling and Circuit Breaking

External APIs have rate limits. Your managed database has a maximum connection count. Protecting these dependencies at the mesh level prevents cascading failures.

apiVersion: networking.istio.io/v1
kind: ServiceEntry
metadata:
  name: stripe-api
  namespace: payments
spec:
  hosts:
    - api.stripe.com
  location: MESH_EXTERNAL
  ports:
    - number: 443
      name: tls
      protocol: TLS
  resolution: DNS
---
apiVersion: networking.istio.io/v1
kind: DestinationRule
metadata:
  name: stripe-api-dr
  namespace: payments
spec:
  host: api.stripe.com
  trafficPolicy:
    connectionPool:
      tcp:
        maxConnections: 50
        connectTimeout: 5s
      http:
        h2UpgradePolicy: DO_NOT_UPGRADE
        maxRequestsPerConnection: 100
    outlierDetection:
      consecutive5xxErrors: 3
      interval: 30s
      baseEjectionTime: 60s
      maxEjectionPercent: 100

This configuration caps outbound connections to Stripe at 50, sets a 5-second connection timeout, and ejects endpoints that return 3 consecutive 5xx errors. In production, this prevents a degraded third-party API from consuming all your connection slots and causing a domino effect across your services.

TLS Origination

Sometimes your application speaks plain HTTP, but the external service requires HTTPS. Instead of modifying application code, you can offload TLS origination to the sidecar.

apiVersion: networking.istio.io/v1
kind: ServiceEntry
metadata:
  name: external-api
  namespace: default
spec:
  hosts:
    - api.external-service.com
  location: MESH_EXTERNAL
  ports:
    - number: 80
      name: http
      protocol: HTTP
    - number: 443
      name: https
      protocol: TLS
  resolution: DNS
---
apiVersion: networking.istio.io/v1
kind: DestinationRule
metadata:
  name: external-api-tls
  namespace: default
spec:
  host: api.external-service.com
  trafficPolicy:
    portLevelSettings:
      - port:
          number: 443
        tls:
          mode: SIMPLE

Your application sends HTTP to port 80. A VirtualService (shown in the next section) redirects that to port 443. The DestinationRule initiates TLS to the external endpoint. The application never knows TLS happened.

Combining ServiceEntry with VirtualService

VirtualService gives you Layer 7 traffic management for external services: retries, timeouts, fault injection, header-based routing, and traffic shifting. This is invaluable when you are migrating between API providers or need resilience policies for unreliable external dependencies.

Retries and Timeouts

apiVersion: networking.istio.io/v1
kind: VirtualService
metadata:
  name: stripe-api-vs
  namespace: payments
spec:
  hosts:
    - api.stripe.com
  http:
    - route:
        - destination:
            host: api.stripe.com
            port:
              number: 443
      timeout: 10s
      retries:
        attempts: 3
        perTryTimeout: 3s
        retryOn: connect-failure,refused-stream,unavailable,cancelled,retriable-status-codes
        retryRemoteLocalities: true

This applies a 10-second overall timeout with up to 3 retry attempts (3 seconds each) for specific failure conditions. Note that this only works for HTTP-protocol ServiceEntries. For TLS-protocol entries where Envoy cannot see the HTTP layer, you are limited to TCP-level connection retries configured via the DestinationRule.

Traffic Shifting Between External Providers

Migrating from one external API to another? Use weighted routing to shift traffic gradually.

apiVersion: networking.istio.io/v1
kind: ServiceEntry
metadata:
  name: geocoding-primary
  namespace: geo
spec:
  hosts:
    - geocoding.internal
  location: MESH_EXTERNAL
  ports:
    - number: 443
      name: tls
      protocol: TLS
  resolution: STATIC
  endpoints:
    - address: api.old-geocoding-provider.com
      labels:
        provider: old
    - address: api.new-geocoding-provider.com
      labels:
        provider: new
---
apiVersion: networking.istio.io/v1
kind: DestinationRule
metadata:
  name: geocoding-dr
  namespace: geo
spec:
  host: geocoding.internal
  trafficPolicy:
    tls:
      mode: SIMPLE
  subsets:
    - name: old-provider
      labels:
        provider: old
    - name: new-provider
      labels:
        provider: new
---
apiVersion: networking.istio.io/v1
kind: VirtualService
metadata:
  name: geocoding-vs
  namespace: geo
spec:
  hosts:
    - geocoding.internal
  http:
    - route:
        - destination:
            host: geocoding.internal
            subset: old-provider
          weight: 80
        - destination:
            host: geocoding.internal
            subset: new-provider
          weight: 20

This sends 80% of geocoding traffic to the old provider and 20% to the new one. Adjust the weights as you gain confidence. Fully reversible — just set the old provider back to 100%.

DNS Resolution Patterns: Istio DNS Proxy vs kube-dns

Istio DNS resolution for external services involves two layers: how your application resolves the hostname (kube-dns / CoreDNS), and how the sidecar resolves the hostname (Envoy’s async DNS or Istio’s DNS proxy). Understanding the interplay is crucial for reliable Istio DNS behavior.

Default Flow (Without Istio DNS Proxy)

Your application calls api.stripe.com. kube-dns resolves it to an IP. The application opens a connection to that IP. The sidecar intercepts the connection and — if the ServiceEntry uses DNS resolution — Envoy independently resolves api.stripe.com to determine its endpoint list. Two separate DNS lookups happen, which can lead to inconsistencies if DNS records change between the two resolutions.

With Istio DNS Proxy (dns.istio.io)

Istio’s sidecar includes a DNS proxy that intercepts DNS queries from the application. When enabled (via meshConfig.defaultConfig.proxyMetadata.ISTIO_META_DNS_CAPTURE and ISTIO_META_DNS_AUTO_ALLOCATE), the proxy can:

• Auto-allocate virtual IPs for ServiceEntry hosts that do not have addresses defined, which is critical for TCP ServiceEntries that need IP-based matching.
• Resolve ServiceEntry hosts directly, avoiding the round-trip to kube-dns for known mesh services.
• Ensure consistency between the application’s DNS resolution and the sidecar’s endpoint resolution.

In modern Istio installations (1.18+), DNS capture is enabled by default. Verify with:

istioctl proxy-config bootstrap <pod-name> -n <namespace> | grep -A2 "ISTIO_META_DNS"

When DNS Proxy Matters Most

The DNS proxy is especially important for TCP ServiceEntries without an explicit addresses field. Without a VIP, Envoy cannot match an incoming TCP connection to the correct ServiceEntry because there is no HTTP Host header to inspect. The DNS proxy solves this by auto-allocating a VIP from the 240.240.0.0/16 range and returning that VIP when the application resolves the hostname. The sidecar then intercepts traffic to that VIP and routes it to the correct external endpoint.

Sticky Sessions with ServiceEntry

Some external services require session affinity — for example, a legacy service that stores session state in memory, or a WebSocket endpoint that must maintain a persistent connection to the same backend. Istio supports sticky sessions for external services through consistent hashing in a DestinationRule.

apiVersion: networking.istio.io/v1
kind: ServiceEntry
metadata:
  name: legacy-session-service
  namespace: default
spec:
  hosts:
    - legacy-session.internal
  location: MESH_INTERNAL
  ports:
    - number: 8080
      name: http
      protocol: HTTP
  resolution: STATIC
  endpoints:
    - address: 10.0.1.10
    - address: 10.0.1.11
    - address: 10.0.1.12
---
apiVersion: networking.istio.io/v1
kind: DestinationRule
metadata:
  name: legacy-session-dr
  namespace: default
spec:
  host: legacy-session.internal
  trafficPolicy:
    loadBalancer:
      consistentHash:
        httpCookie:
          name: SERVERID
          ttl: 3600s

This configuration hashes on an HTTP cookie named SERVERID. If the cookie does not exist, Envoy generates one and sets it on the response so that subsequent requests from the same client stick to the same endpoint. You can also hash on:

• HTTP header: consistentHash.httpHeaderName: "x-user-id" — useful when your application sends a user identifier in every request.
• Source IP: consistentHash.useSourceIp: true — simplest option but breaks in environments with NAT or shared egress IPs.
• Query parameter: consistentHash.httpQueryParameterName: "session_id" — for REST APIs that include a session identifier in the URL.

Sticky sessions with ServiceEntry work identically to in-mesh sticky sessions. The key requirement is that the ServiceEntry must use STATIC or DNS resolution (not NONE) so that Envoy has multiple endpoints to hash across. With DNS_ROUND_ROBIN, there is only one logical endpoint, so consistent hashing has no effect.

Troubleshooting Common Issues

503 Errors When Calling External Services

The most common ServiceEntry issue. Start with this diagnostic sequence:

# Check if the ServiceEntry is applied and visible to the proxy
istioctl proxy-config cluster <pod-name> -n <namespace> | grep <external-host>

# Check the listeners
istioctl proxy-config listener <pod-name> -n <namespace> --port <port>

# Look at Envoy access logs for the specific request
kubectl logs <pod-name> -n <namespace> -c istio-proxy | grep <external-host>

Common causes of 503 errors:

• Wrong protocol: Setting protocol: HTTPS when your application initiates TLS. Use TLS for pass-through; use HTTP only if the sidecar does TLS origination.
• Missing ServiceEntry in REGISTRY_ONLY mode: If outboundTrafficPolicy is REGISTRY_ONLY, any host without a ServiceEntry is blocked.
• exportTo restriction: The ServiceEntry is in namespace A, exported only to ".", and the calling pod is in namespace B.
• DNS resolution failure: Envoy cannot resolve the hostname. Check that the DNS servers are reachable from the pod.

DNS Resolution Failures

When Envoy’s async DNS resolver fails, you will see UH (upstream unhealthy) or UF (upstream connection failure) flags in access logs.

# Verify DNS works from inside the sidecar
kubectl exec <pod-name> -n <namespace> -c istio-proxy -- \
  pilot-agent request GET /dns_resolve?proxyID=<pod-name>.<namespace>&host=api.stripe.com

# Check Envoy cluster health
istioctl proxy-config endpoint <pod-name> -n <namespace> | grep <external-host>

If the endpoint shows UNHEALTHY, Envoy resolved the DNS but the outlier detection ejected the host. If no endpoint appears at all, DNS resolution is failing. Common fix: ensure your pods can reach an external DNS server, or that CoreDNS is configured to forward queries for the external domain.

TLS Origination Not Working

If you configured TLS origination via a DestinationRule but traffic still fails:

• Ensure the ServiceEntry port protocol is HTTP, not TLS. If you set it to TLS, Envoy treats the connection as opaque TLS pass-through and will not apply the DestinationRule’s TLS settings.
• Verify the DestinationRule’s host field exactly matches the ServiceEntry’s hosts entry.
• Check that the VirtualService (if used) routes to the correct port number.

TCP ServiceEntry Not Intercepting Traffic

For TCP-protocol ServiceEntries without the DNS proxy, Envoy cannot match traffic by hostname. You must either:

• Set an explicit addresses field with a VIP that your application targets.
• Enable Istio’s DNS proxy to auto-allocate VIPs.
• Ensure the destination IP matches what the ServiceEntry resolves to.

Without one of these, TCP traffic goes through the PassthroughCluster and bypasses your ServiceEntry entirely.

Frequently Asked Questions

Wrapping Up

ServiceEntry is one of the most practical Istio resources you will use in production. It transforms opaque outbound connections into managed, observable, policy-controlled traffic — and it does so without requiring changes to your application code. Start with the basics: create a ServiceEntry for each external dependency, set the correct resolution type, and pair it with a DestinationRule for connection limits and circuit breaking. As you mature, add VirtualServices for retries and timeouts, configure sticky sessions where needed, and enable the DNS proxy for seamless TCP service integration.

The pattern is always the same: register the service, apply policies, observe the traffic. Every external dependency you formalize with a ServiceEntry is one fewer blind spot in your production mesh.

Most observability stacks that have been running in production for more than a year end up with alerting spread across two systems: Prometheus Alertmanager handling metric-based alerts and Grafana Alerting managing everything else. Engineers add a Slack integration in Grafana because it is convenient, then realize their Alertmanager routing tree already covers the same service. Before long, the on-call team receives duplicated pages, silencing rules live in two places, and nobody is confident which system is authoritative.

Related reading: OpenSearch vs Loki vs Quickwit vs ClickHouse.

This is the alerting consolidation problem, and it affects teams of every size. The question is straightforward: should you standardize on Prometheus Alertmanager, move everything into Grafana Alerting, or deliberately run both? The answer depends on your datasource mix, your GitOps maturity, and how your organization manages on-call routing. This guide breaks down the architecture, features, and operational trade-offs of each system so you can make a deliberate choice instead of drifting into accidental complexity.

Alertmanager vs Grafana Alerting: the short answer

Prometheus Alertmanager only routes and delivers alerts that Prometheus has already evaluated — it is a single YAML file, GitOps-native, lightweight, and unbeatable for a Prometheus-only metrics stack. Grafana Alerting evaluates the rules itself against any datasource (Loki, Elasticsearch, CloudWatch, SQL, Prometheus), adds a UI, native multi-tenancy and mute timings, and is the only choice when you alert on more than metrics. If your alerts are all PromQL and you value declarative config, use Alertmanager; if you alert on logs, databases or want UI-driven management, use Grafana Alerting; many teams deliberately run both — Alertmanager for the metric pipeline, Grafana for everything else. That is the same question as “Grafana alerts vs Prometheus alerts”, and the rest of this guide backs up each side with architecture, a feature table, and a decision framework.

Architecture Overview

Before comparing features, you need to understand how each system fits into the alerting pipeline. They occupy the same logical space — “receive a condition, route a notification” — but they get there from fundamentally different starting points.

Prometheus Alertmanager: The Standalone Receiver

Alertmanager is a dedicated, standalone component in the Prometheus ecosystem. It does not evaluate alert rules itself. Instead, Prometheus (or any compatible sender like Thanos Ruler, Cortex, or Mimir Ruler) evaluates PromQL expressions and pushes firing alerts to the Alertmanager API. Alertmanager then handles deduplication, grouping, inhibition, silencing, and notification delivery. Those metrics reach Prometheus in the first place through the scrape targets you declare as ServiceMonitor or PodMonitor CRDs.

# Simplified Prometheus → Alertmanager flow
#
# [Prometheus] --evaluates rules--> [firing alerts]
#        |
#        +--POST /api/v2/alerts--> [Alertmanager]
#                                      |
#                          +-----------+-----------+
#                          |           |           |
#                       [Slack]    [PagerDuty]  [Email]

The entire configuration lives in a single YAML file (alertmanager.yml). This includes the routing tree, receiver definitions, inhibition rules, and silence templates. There is no database, no UI-driven state — just a config file and an optional local storage directory for notification state and silences. This makes it trivially reproducible and ideal for GitOps workflows.

For high availability, you run multiple Alertmanager instances in a gossip-based cluster. They use a mesh protocol to share silence and notification state, ensuring that failover does not result in duplicate or lost notifications. The HA model is well-understood and has been stable for years.

Grafana Alerting: The Integrated Platform

Grafana Alerting (sometimes called “Grafana Unified Alerting,” introduced in Grafana 8 and significantly matured through Grafana 11 and 12) takes a different architectural approach. It embeds the entire alerting lifecycle — rule evaluation, state management, routing, and notification — inside the Grafana server process. Under the hood, it actually uses a fork of Alertmanager for the routing and notification layer, but this is an implementation detail that is invisible to users.

# Simplified Grafana Alerting flow
#
# [Grafana Server]
#   ├── Rule Evaluation Engine
#   │     ├── queries Prometheus
#   │     ├── queries Loki
#   │     ├── queries CloudWatch
#   │     └── queries any supported datasource
#   │
#   ├── Alert State Manager (internal)
#   │
#   └── Embedded Alertmanager (routing + notifications)
#           |
#           +-----------+-----------+
#           |           |           |
#        [Slack]    [PagerDuty]  [Email]

The critical distinction is that Grafana Alerting evaluates alert rules itself, querying any configured datasource — not just Prometheus. It can fire alerts based on Loki log queries, Elasticsearch searches, CloudWatch metrics, PostgreSQL queries, or any of the 100+ datasource plugins available in Grafana. Rule definitions, contact points, notification policies, and mute timings are stored in the Grafana database (or provisioned via YAML files and the Grafana API).

For high availability in self-hosted environments, Grafana Alerting relies on a shared database and a peer-discovery mechanism between Grafana instances. In Grafana Cloud, HA is fully managed by Grafana Labs.

A note on the term “Grafana Alertmanager”

People search for “Grafana Alertmanager” and mean one of three different things, so it is worth separating them. First, the embedded Alertmanager: Grafana Alerting uses a fork of the Prometheus Alertmanager code internally for its routing and notification layer — when the Grafana UI shows an “Alertmanager”, this built-in engine is usually what it means. Second, an external Alertmanager: Grafana can connect to a standalone Prometheus Alertmanager as a datasource and act as a read-only viewer for its firing alerts and silences (see the hybrid section below). Third, Mimir/Cloud Alertmanager: Grafana Cloud exposes a managed, multi-tenant Alertmanager for teams that want Prometheus-native routing without self-hosting. So “Grafana Alertmanager” is not a separate product — it is one of these three, and which one you mean changes the answer to almost every configuration question.

Feature Comparison

The following table provides a side-by-side comparison of the capabilities that matter most in production alerting systems. Both systems are mature, but they prioritize different things.

Alertmanager Strengths

Alertmanager has been a production staple since 2015. Over a decade of use across thousands of organizations has made it one of the most battle-tested components in the CNCF ecosystem. Here is where it genuinely excels.

Declarative, GitOps-Native Configuration

The entire Alertmanager configuration is a single YAML file. There is no hidden state in a database, no click-driven configuration that someone forgets to document. You check it into Git, review it in a pull request, and deploy it through your CI/CD pipeline like any other infrastructure code. This is a significant operational advantage for teams that have invested in GitOps.

# alertmanager.yml — everything in one file
global:
  resolve_timeout: 5m
  slack_api_url: "https://hooks.slack.com/services/T00/B00/XXX"

route:
  receiver: platform-team
  group_by: [alertname, cluster, namespace]
  group_wait: 30s
  group_interval: 5m
  repeat_interval: 4h
  routes:
    - match:
        severity: critical
      receiver: pagerduty-oncall
      group_wait: 10s
    - match_re:
        team: "^(payments|checkout)$"
      receiver: payments-slack
      continue: true

receivers:
  - name: platform-team
    slack_configs:
      - channel: "#platform-alerts"
  - name: pagerduty-oncall
    pagerduty_configs:
      - service_key: ""
  - name: payments-slack
    slack_configs:
      - channel: "#payments-oncall"

inhibit_rules:
  - source_match:
      severity: critical
    target_match:
      severity: warning
    equal: [alertname, cluster]

Every change is auditable. Rollbacks are a git revert away. This matters enormously when you are debugging why an alert did not fire at 3 AM.

Lightweight and Single-Purpose

Alertmanager does one thing: route and deliver notifications. It has no dashboard, no query engine, no datasource plugins. This single-purpose design makes it operationally simple. Resource consumption is minimal — a small Alertmanager instance handles thousands of active alerts on a few hundred megabytes of memory. It starts in milliseconds and requires almost no maintenance.

Mature Inhibition and Routing

Alertmanager’s inhibition rules are first-class citizens. You can suppress downstream warnings when a critical alert is already firing, preventing alert storms from overwhelming your on-call team. The hierarchical routing tree with continue flags allows for nuanced delivery: send to the team channel AND escalate to PagerDuty simultaneously, with different grouping strategies at each level.

Proven High Availability

The gossip-based HA cluster has been stable for years. Running three Alertmanager replicas behind a load balancer (or using Kubernetes service discovery) gives you reliable notification delivery without shared storage. The protocol handles deduplication across instances automatically, which is the hardest part of distributed alerting.

Grafana Alerting Strengths

Grafana Alerting has matured considerably since its rocky introduction in Grafana 8. By Grafana 11 and 12, it has become a legitimate production alerting platform with capabilities that Alertmanager cannot match on its own.

Multi-Datasource Alert Rules

This is Grafana Alerting’s strongest differentiator. You can write alert rules that query Loki for error log spikes, CloudWatch for AWS resource utilization, Elasticsearch for application errors, or a PostgreSQL database for business metrics — all from the same alerting system. If your observability stack includes more than just Prometheus, this eliminates the need for separate alerting tools per datasource.

# Grafana alert rule provisioning example — alerting on Loki log errors
apiVersion: 1
groups:
  - orgId: 1
    name: application-errors
    folder: Production
    interval: 1m
    rules:
      - uid: loki-error-spike
        title: "High error rate in payment service"
        condition: C
        data:
          - refId: A
            datasourceUid: loki-prod
            model:
              expr: 'sum(rate({app="payment-service"} |= "ERROR" [5m]))'
          - refId: B
            datasourceUid: "__expr__"
            model:
              type: reduce
              expression: A
              reducer: last
          - refId: C
            datasourceUid: "__expr__"
            model:
              type: threshold
              expression: B
              conditions:
                - evaluator:
                    type: gt
                    params: [10]
        for: 5m
        labels:
          severity: warning
          team: payments

This is something Alertmanager simply cannot do. Alertmanager only receives pre-evaluated alerts — it has no concept of datasources or query execution.

Unified UI for Alert Management

Grafana provides a single pane of glass for alert rule creation, visualization, notification policy management, contact point configuration, and silence management. For teams where not every engineer is comfortable editing YAML routing trees, the visual notification policy editor significantly reduces the barrier to entry. You can see the state of every alert rule, its evaluation history, and the exact notification path it will take — all without leaving the browser.

Native Multi-Tenancy and RBAC

Grafana’s organization model and role-based access control extend naturally to alerting. Different teams can manage their own alert rules, contact points, and notification policies within their organization or folder scope, without seeing or interfering with other teams. Achieving this with standalone Alertmanager requires either running separate instances per tenant or using Mimir’s multi-tenant Alertmanager.

Mute Timings and Richer Scheduling

While Alertmanager supports silences (ad-hoc, time-limited suppressions), Grafana Alerting adds mute timings — recurring time-based windows where notifications are suppressed. This is useful for scheduled maintenance windows, business-hours-only alerting, or suppressing non-critical alerts on weekends. Alertmanager requires external tooling or manual silence creation for recurring windows.

Grafana Cloud as a Managed Option

For teams that want to avoid managing alerting infrastructure entirely, Grafana Cloud provides a fully managed Grafana Alerting stack. This includes HA, state persistence, and notification delivery without any self-hosted components. The Grafana Cloud alerting stack also includes a managed Mimir Alertmanager, which means you can use Prometheus-native alerting rules if you prefer that model while still benefiting from the managed infrastructure.

When to Use Prometheus Alertmanager

Alertmanager is the right choice when the following conditions describe your environment:

• Your metrics stack is Prometheus-native. If all your alert rules are PromQL expressions evaluated by Prometheus, Thanos Ruler, or Mimir Ruler, Alertmanager is the natural fit. There is no added value in routing those alerts through Grafana.
• GitOps is non-negotiable. If every infrastructure change must go through a pull request and be fully declarative, Alertmanager’s single-file configuration model is significantly easier to manage than Grafana’s database-backed state. Tools like amtool provide config validation in CI pipelines.
• You need fine-grained routing with inhibition. Complex routing trees with multiple levels of grouping, inhibition rules, and continue flags are more naturally expressed in Alertmanager’s YAML format. The routing logic has been stable and well-documented for years.
• You run microservices with per-team routing. If each team owns its routing subtree and the routing logic is complex, Alertmanager’s hierarchical model scales better than UI-driven configuration. Teams can own their section of the config file via CODEOWNERS in Git.
• You want minimal operational overhead. Alertmanager is a single binary with minimal resource requirements. There is no database to back up, no migrations to run, and no UI framework to keep updated.

When to Use Grafana Alerting

Grafana Alerting is the right choice when these conditions apply:

• You alert on more than just Prometheus metrics. If you need alert rules based on Loki logs, Elasticsearch queries, CloudWatch metrics, or database queries, Grafana Alerting is the only option that handles all of these natively. The alternative is running separate alerting tools per datasource, which is worse.
• Your team prefers UI-driven configuration. Not every engineer wants to edit YAML routing trees. If your organization values a visual interface for managing alerts, contact points, and notification policies, Grafana’s UI is a major productivity advantage.
• You are using Grafana Cloud. If you are already on Grafana Cloud, using its built-in alerting is the path of least resistance. You get HA, managed notification delivery, and a unified experience without running any additional infrastructure.
• Multi-tenancy is a requirement. If multiple teams need isolated alerting configurations with RBAC, Grafana’s native organization and folder-based access model is significantly easier to set up than running per-tenant Alertmanager instances.
• You want mute timings for recurring maintenance windows. If your team regularly needs to suppress alerts during scheduled windows (deploy windows, batch processing hours, weekend non-critical suppression), Grafana’s mute timings feature is more ergonomic than creating and managing recurring silences in Alertmanager.

Running Both Together: The Hybrid Pattern

In practice, many production environments run both Alertmanager and Grafana Alerting. This is not necessarily a mistake — it can be a deliberate architectural choice when done with clear boundaries.

Common Hybrid Architecture

The most common pattern looks like this:

• Prometheus Alertmanager handles all metric-based alerts. PromQL rules are evaluated by Prometheus or a long-term storage ruler (Thanos, Mimir). Alertmanager owns routing, grouping, and notification for these alerts.
• Grafana Alerting handles non-Prometheus alerts: log-based alerts from Loki, business metrics from SQL datasources, and cross-datasource correlation rules.

The key to making this work without chaos is establishing clear ownership rules:

# Ownership boundaries for hybrid alerting
#
# Prometheus Alertmanager owns:
#   - All PromQL-based alert rules
#   - Infrastructure alerts (node, kubelet, etcd, CoreDNS)
#   - Application SLO/SLI alerts based on metrics
#
# Grafana Alerting owns:
#   - Log-based alert rules (Loki, Elasticsearch)
#   - Business metric alerts (SQL datasources)
#   - Cross-datasource correlation rules
#   - Alerts for teams that prefer UI-driven management
#
# Shared:
#   - Contact points / receivers use the same Slack channels and PagerDuty services
#   - On-call rotations are managed externally (PagerDuty, Grafana OnCall)

Both systems can deliver to the same notification channels. The critical discipline is ensuring that silencing and maintenance windows are applied in both systems when needed. This is the primary operational cost of the hybrid approach.

Grafana as a Viewer for Alertmanager

Even if you use Alertmanager exclusively for routing and notification, Grafana can serve as a read-only viewer. Grafana natively supports connecting to an external Alertmanager datasource, allowing you to see firing alerts, active silences, and alert groups in the Grafana UI. This gives you the operational visibility of Grafana without moving your alerting logic into it.

# Grafana datasource provisioning for external Alertmanager
apiVersion: 1
datasources:
  - name: Alertmanager
    type: alertmanager
    url: http://alertmanager.monitoring.svc:9093
    access: proxy
    jsonData:
      implementation: prometheus

Migration Considerations

If you are moving from one system to the other, here are the practical considerations to plan for.

Migrating from Alertmanager to Grafana Alerting

• Rule conversion. Your PromQL-based recording and alerting rules defined in Prometheus rule files need to be recreated as Grafana alert rules. Grafana provides a migration tool that can import Prometheus-format rules, but complex expressions may need manual adjustment.
• Routing tree translation. Alertmanager’s hierarchical routing tree maps to Grafana’s notification policies, but the semantics are not identical. Test the notification routing thoroughly — the continue flag behavior and default routes may differ.
• Silence and inhibition migration. Active silences are ephemeral and do not need migration. Inhibition rules need to be recreated in Grafana’s format. Recurring maintenance windows should be converted to mute timings.
• Run in parallel first. The safest migration strategy is to run both systems in parallel for two to four weeks, sending notifications from both, then cutting over when you have confidence in the Grafana setup. Accept the temporary noise of duplicate alerts — it is far cheaper than missing a critical page during migration.

Migrating from Grafana Alerting to Alertmanager

• Datasource limitation. You can only migrate alerts that are based on Prometheus-compatible datasources. Alerts querying Loki, Elasticsearch, or SQL datasources have no equivalent in Alertmanager — you will need an alternative solution for those.
• Rule export. Export Grafana alert rules and convert them to Prometheus-format rule files. The Grafana API (GET /api/v1/provisioning/alert-rules) provides structured output that can be transformed with a script.
• Contact point mapping. Map Grafana contact points to Alertmanager receivers. The configuration format is different, but the concepts are equivalent.
• State loss. Alertmanager does not carry over Grafana’s alert evaluation history. You start fresh. Plan for a brief period where alerts may re-fire as Prometheus evaluates rules that were previously managed by Grafana.

Decision Framework

If you want a quick decision path, use this framework:

Start here:
│
├── Do you alert on non-Prometheus datasources (Loki, ES, SQL, CloudWatch)?
│   ├── YES → Grafana Alerting (at least for those datasources)
│   └── NO ↓
│
├── Is GitOps/declarative config a hard requirement?
│   ├── YES → Alertmanager
│   └── NO ↓
│
├── Do you need multi-tenancy with RBAC?
│   ├── YES → Grafana Alerting (or Mimir Alertmanager)
│   └── NO ↓
│
├── Are you on Grafana Cloud?
│   ├── YES → Grafana Alerting (path of least resistance)
│   └── NO ↓
│
└── Default → Alertmanager (simpler, lighter, well-proven)

For many teams, the honest answer is “both” — Alertmanager for the Prometheus-native metric pipeline, Grafana Alerting for everything else. That is a valid architecture as long as the ownership boundaries are documented and the on-call team knows where to look.

Frequently Asked Questions

Final Thoughts

The Alertmanager vs Grafana Alerting debate is not really about which tool is better — it is about which tool fits your operational context. Alertmanager is simpler, lighter, and more GitOps-friendly. Grafana Alerting is more versatile, more accessible to UI-oriented teams, and the only option if you need multi-datasource alerting. Running both is perfectly valid when the boundaries are clear.

The worst outcome is not picking the “wrong” tool. The worst outcome is running both accidentally, with overlapping coverage, duplicated notifications, and no clear ownership. Whatever you choose, document the decision, define the ownership boundaries, and make sure your on-call team knows exactly where to go when they need to silence an alert at 3 AM.