10 APM Tools DevOps Teams Use for Tracing and Latency
10 APM tools compared: Datadog, Dynatrace, New Relic, Splunk AppDynamics, Elastic, Grafana Tempo, Honeycomb, Instana, App Insights, Jaeger.
Application performance monitoring (APM) tools trace requests through your code and services so you can see where latency and errors come from. Ten tools DevOps teams commonly use are Datadog APM, Dynatrace, New Relic, Splunk AppDynamics, Elastic APM, Grafana Cloud with Tempo, Honeycomb, IBM Instana, Azure Application Insights and the open-source Jaeger with OpenTelemetry.
This post focuses on APM specifically: distributed tracing, service maps, code-level timing and the golden signals of latency, traffic, errors and saturation. For the wider monitoring stack (metrics, logs, dashboards and alerting), see our pillar guide to observability tools for DevOps monitoring. Features were checked against each vendor's documentation in November 2026. Pricing models differ and change often, so prices are left out.
Key takeaways
- APM answers "which service, endpoint or query made this request slow", while infrastructure monitoring answers "which host or pod is unhealthy".
- OpenTelemetry is now supported by every tool on this list, so you can instrument once and keep the option to change backends.
- Commercial suites (Datadog, Dynatrace, New Relic, AppDynamics, Instana) bundle tracing with infrastructure, logs and real user monitoring; open-source stacks need more assembly.
- Sampling decides cost and usefulness. Tail-based sampling keeps the slow and failed traces you actually need.
- Pick a tool by where you run (Kubernetes, VMs, serverless, one cloud), the languages you use, and how much you want to operate yourself.
What an APM tool does
An APM agent or SDK records a span for each unit of work: an incoming HTTP request, a database query, a call to another service. Spans share a trace ID that travels in request headers (the W3C traceparent header defined in the Trace Context recommendation), so the backend can assemble the full path of one request across services. From those traces the tool builds service maps, latency percentiles per endpoint, error rates and flame graphs.
Most products add continuous profiling (which functions use CPU and memory), real user monitoring (page loads in real browsers) and synthetic checks. If you are new to the difference between monitoring and observability, read observability vs traditional monitoring first.
The 10 APM tools compared
| # | Tool | Model | Instrumentation | Best fit |
|---|---|---|---|---|
| 1 | Datadog APM | SaaS | Datadog tracers, OpenTelemetry | Teams that want one SaaS for traces, metrics, logs and RUM |
| 2 | Dynatrace | SaaS or Managed | OneAgent auto-instrumentation, OpenTelemetry | Large estates that want automatic discovery and root-cause analysis |
| 3 | New Relic | SaaS | Language agents, OpenTelemetry | Teams wanting full-stack APM with a usage-based free tier |
| 4 | Splunk AppDynamics | SaaS or on-premises | AppDynamics agents | Enterprises with Java and .NET business transactions, hybrid estates |
| 5 | Elastic APM | Self-managed or Elastic Cloud | Elastic APM agents, OpenTelemetry (EDOT) | Teams already on Elasticsearch for logs |
| 6 | Grafana Cloud / Tempo | Open source or SaaS | OpenTelemetry, Grafana Alloy | Prometheus and Grafana users who want traces in the same UI |
| 7 | Honeycomb | SaaS | OpenTelemetry | High-cardinality debugging of distributed systems |
| 8 | IBM Instana | SaaS or self-hosted | Single host agent, OpenTelemetry | Automatic tracing in Kubernetes and IBM-heavy environments |
| 9 | Azure Application Insights | Azure managed | Azure Monitor OpenTelemetry Distro, autoinstrumentation | Applications hosted on Azure |
| 10 | Jaeger + OpenTelemetry | Open source (CNCF) | OpenTelemetry SDKs and Collector | Self-hosted tracing with no vendor licence |
The tools in detail
1. Datadog APM
Datadog APM traces requests with Datadog's language tracers or OpenTelemetry and links each trace to host metrics, container metrics, logs and profiles in the same platform. Its service catalog, service map and Watchdog anomaly detection help teams find which deploy or dependency caused a latency jump. Deployment tracking tags traces with the version, so you can compare error rates before and after a release.
It suits teams that already use Datadog for infrastructure and want everything in one place. Watch ingestion and retention settings, because span volume drives cost. See Datadog APM docs.
2. Dynatrace
Dynatrace installs one OneAgent per host or as a Kubernetes operator, which discovers processes and instruments supported runtimes automatically. Its Smartscape topology maps dependencies, and the Davis AI engine groups related anomalies into a single problem with a probable root cause. Grail, its data lakehouse, stores traces, logs and metrics for querying with DQL.
Dynatrace fits large organisations that want deep automatic coverage with little manual instrumentation. It also accepts OpenTelemetry data. See Dynatrace documentation and our Dynatrace interview questions.
3. New Relic
New Relic provides language agents for Java, .NET, Node.js, Python, Ruby, Go and PHP, plus native OpenTelemetry ingest. Everything lands in one telemetry database you query with NRQL, which makes ad hoc questions such as "p95 latency by customer tier for the checkout endpoint" quick to answer. Errors Inbox groups errors across services, and change tracking marks deployments on charts.
New Relic's free tier includes a monthly data allowance, which makes it practical for learning and small projects. See New Relic APM docs.
4. Splunk AppDynamics
AppDynamics models applications as business transactions, such as "login" or "checkout", and baselines each one so it can alert on deviation instead of fixed thresholds. Cisco acquired AppDynamics in 2017, then completed its acquisition of Splunk in 2024, and the product is now sold as Splunk AppDynamics. It is strongest for Java and .NET monoliths and hybrid estates where on-premises deployment still matters.
Splunk also sells Splunk Observability Cloud with its own OpenTelemetry-native APM, so check which product a job description means. See the Splunk AppDynamics product page.
5. Elastic APM
Elastic APM stores traces in Elasticsearch and shows them in Kibana's Observability app beside logs and metrics. You can use Elastic's own agents or the Elastic Distributions of OpenTelemetry. Teams that already centralise logs in the Elastic Stack get trace-to-log correlation without adding a vendor.
Self-managing Elasticsearch at trace volume takes planning for shards, retention and index lifecycle. Elastic Cloud removes that work. See Elastic APM docs.
6. Grafana Cloud with Tempo
Grafana Tempo is an open-source trace backend that stores traces cheaply in object storage and is queried with TraceQL. Grafana Cloud bundles Tempo with Mimir (metrics), Loki (logs) and Pyroscope (profiles), and its Application Observability view gives RED metrics per service built from OpenTelemetry data. Grafana Alloy is Grafana's distribution of the OpenTelemetry Collector.
This is the natural choice for teams already running Prometheus and Grafana. See Grafana Tempo docs and our guide to using Prometheus and Grafana together.
7. Honeycomb
Honeycomb stores wide events with many attributes and lets you slice them by any field, even ones with millions of distinct values such as user ID or build ID. BubbleUp compares slow requests against normal ones and shows which attributes differ. It is built around OpenTelemetry and promotes SLO-based alerting.
It suits teams debugging unknown problems in distributed systems more than teams wanting packaged infrastructure dashboards. See Honeycomb docs.
8. IBM Instana
Instana uses one agent per host that discovers technologies and traces every request without sampling by default, with one-second metric granularity. It maps Kubernetes clusters, services and endpoints automatically and supports OpenTelemetry. IBM acquired Instana in 2020.
It fits teams that want automatic coverage with minimal configuration, especially in Kubernetes and in environments with IBM middleware. See IBM Instana documentation.
9. Azure Application Insights
Application Insights is the APM feature of Azure Monitor. Microsoft recommends the Azure Monitor OpenTelemetry Distro for .NET, Java, Node.js and Python, and offers autoinstrumentation for App Service, Functions and AKS. Data lands in a Log Analytics workspace, where you query it with KQL. Microsoft retired classic (non-workspace) Application Insights resources in February 2024.
For applications on Azure it is the lowest-effort option and integrates with Azure alerts and workbooks. See Application Insights overview.
10. Jaeger with OpenTelemetry
Jaeger is a CNCF graduated tracing project. Jaeger v2 is built on the OpenTelemetry Collector and stores traces in backends such as Elasticsearch, OpenSearch or Cassandra. Paired with OpenTelemetry SDKs, it gives you distributed tracing with no licence cost, at the price of running and scaling the storage yourself.
It is a common choice for learning tracing and for teams that must keep data in their own environment. See Jaeger documentation.
Instrumenting once with OpenTelemetry
Whatever backend you choose, the OpenTelemetry Collector lets you send the same traces to one or more destinations. A minimal Collector config that receives OTLP from applications, batches spans and exports them looks like this:
receivers:
otlp:
protocols:
grpc:
endpoint: 0.0.0.0:4317
http:
endpoint: 0.0.0.0:4318
processors:
batch: {}
exporters:
otlp:
endpoint: tempo.observability.svc:4317
tls:
insecure: true
service:
pipelines:
traces:
receivers: [otlp]
processors: [batch]
exporters: [otlp]
Ports 4317 (gRPC) and 4318 (HTTP) are the OTLP defaults listed in the OTLP specification. For many languages you can start with zero-code instrumentation. Example for a Java service:
export OTEL_SERVICE_NAME=checkout
export OTEL_EXPORTER_OTLP_ENDPOINT=http://otel-collector:4318
java -javaagent:opentelemetry-javaagent.jar -jar checkout.jar
In production, use tls with real certificates and add tail-based sampling in the Collector so you keep every error and slow trace while dropping most routine ones.
APM signals worth tracking
| Signal | What it tells you | Typical alert |
|---|---|---|
| Latency percentiles (p50, p95, p99) | How slow requests are for most users and for the unlucky few | p99 above the SLO target for 10 minutes |
| Error rate | Share of requests that fail (5xx, exceptions) | Above 1% for an endpoint after a deploy |
| Throughput | Requests per second per service and endpoint | Sudden drop, which often means upstream failure |
| Apdex or SLO burn rate | User satisfaction or error budget use over time | Burn rate high enough to exhaust the budget early |
| Downstream call time | Time spent in databases, caches and external APIs | One dependency taking most of the request time |
| Saturation | Thread pools, connection pools, queue depth | Pool exhausted or queue growing steadily |
Average latency hides problems. A service with a low average can still have a p99 of several seconds, and those slow requests are often the most complex orders. Alert on percentiles.
Common APM mistakes
- Instrumenting only the frontend service. Without spans from every hop, the trace shows a long gap where the real delay sits.
- Breaking context propagation. Message queues, thread pools and custom HTTP clients can drop the trace header. Check that one trace covers the whole request.
- Keeping 100% of traces forever. Costs grow quickly. Sample routine traffic and keep errors and slow requests.
- Putting personal data into span attributes. Request bodies, emails and tokens end up in a third-party system. Use Collector processors to drop or mask them.
- Inconsistent service names across environments. Set the
service.nameanddeployment.environmentresource attributes the same way everywhere so dashboards line up. - Treating APM as an operations tool only. Developers should look at traces for their own services during code review and incident reviews.
How to choose an APM tool
| If your situation is | Start by evaluating |
|---|---|
| Already using Prometheus and Grafana | Grafana Tempo or Grafana Cloud |
| Logs already in Elasticsearch | Elastic APM |
| Hosted mostly on Azure | Application Insights |
| Large hybrid estate, little time to instrument | Dynatrace or Instana |
| Java and .NET business-transaction monitoring | Splunk AppDynamics |
| Want one SaaS for everything | Datadog or New Relic |
| Hard-to-reproduce bugs in microservices | Honeycomb |
| Must self-host with no licence cost | Jaeger or Tempo with OpenTelemetry |
Run a two-week trial on one real service. Check how fast you can answer three questions: which endpoint is slowest at p99, which downstream call causes it, and which deploy changed it. Also measure the data volume your service generates, because that predicts the bill.
Using APM in the DevOps loop
- Mark every deployment in the APM tool so latency changes line up with releases. This feeds the post-deploy checks in our post-deployment validation steps.
- Alert on SLOs (for example, 99.5% of checkout requests under 400 ms) rather than on raw CPU.
- Link traces to logs with the trace ID in every log line, so one click goes from a slow span to its log entries.
- Use APM data to trigger automated rollback when error rate or latency crosses a threshold after a canary. See our deployment rollback strategies.
To practise building a metrics and tracing stack, the Prometheus and Grafana training runs in Kharadi, Pune and live online.
Official references
- OpenTelemetry documentation
- W3C Trace Context
- Datadog APM
- Dynatrace documentation
- Grafana Tempo
- Jaeger documentation
- Azure Application Insights
Frequently asked questions
What is the difference between APM and observability?
APM focuses on application behaviour: request traces, endpoint latency, errors and code-level timing. Observability is the broader ability to understand a system from its telemetry, combining traces with metrics, logs and profiles. Most APM products now sit inside larger observability platforms.
Which APM tool is best for Kubernetes?
Dynatrace, Datadog and Instana offer the most automatic Kubernetes coverage through operators or host agents. For open source, OpenTelemetry with Grafana Tempo or Jaeger works well, especially with the OpenTelemetry Operator for auto-instrumentation.
Is there a free APM tool?
Yes. Jaeger and Grafana Tempo are open source and free to run, though you pay for infrastructure. New Relic and Grafana Cloud have free tiers with monthly data limits that suit learning and small services.
Do I need OpenTelemetry if my vendor has its own agent?
Not strictly, but OpenTelemetry keeps your instrumentation portable. If you later change vendors, you change the Collector exporter instead of re-instrumenting every service. All ten tools here accept OpenTelemetry data.
What is distributed tracing?
Distributed tracing follows one request across multiple services. Each service records spans that share a trace ID passed in request headers, and the backend joins them into one timeline showing where time was spent.
What happened to AppDynamics?
Cisco bought AppDynamics in 2017 and completed its acquisition of Splunk in 2024. AppDynamics is now sold as Splunk AppDynamics, alongside Splunk Observability Cloud.
How does sampling affect APM?
Sampling keeps a fraction of traces to control cost. Head-based sampling decides at the start of a request; tail-based sampling decides after it finishes, so it can keep all errors and slow requests. Tail-based sampling is more useful but needs a Collector or backend that buffers traces.
What's Your Reaction?
Like
0
Dislike
0
Love
0
Funny
0
Angry
0
Sad
0
Wow
0