Datadog is a commercial software-as-a-service observability platform that collects metrics, logs, traces and other telemetry from servers, containers, cloud services and applications, and correlates them in one shared interface for monitoring, alerting and troubleshooting. In Observability within the Cloud field, it is one of the most widely used all-in-one monitoring products, competing with open-source stacks built on Prometheus and Grafana.
Its central idea is correlation through shared tags: every data point carries labels such as host, service, environment and version, so an engineer can move from a latency spike on a dashboard to the related traces and log lines in a few clicks.
How Datadog collects and correlates telemetry
Datadog gathers data through three main paths: a local agent, direct cloud integrations and instrumentation libraries inside applications. All three send data to the hosted backend, where it is indexed, stored and made queryable.
The Datadog Agent is a lightweight process installed on hosts or deployed as a DaemonSet in Kubernetes. It collects system metrics, tails log files, runs checks for databases and middleware and forwards traces from application libraries. Cloud integrations pull metrics from provider APIs, such as AWS CloudWatch or Azure Monitor, without any agent. Tracing libraries and OpenTelemetry SDKs instrument code so that each request produces spans for distributed tracing.
The table below shows how each telemetry type enters the platform and what question it answers.
| Telemetry type | Collected by | Question it answers |
|---|---|---|
| Metrics | Agent, cloud integrations, custom code | How much and how fast over time |
| Logs | Agent log tailing, forwarders, APIs | What exactly happened in one event |
| Traces | Tracing libraries, OpenTelemetry | Where a request spent its time |
| Profiles | Continuous profiler in the runtime | Which code consumes CPU or memory |
| Real user data | Browser and mobile SDKs | What end users actually experience |
| Synthetic results | Scripted tests from global locations | Whether key journeys work right now |
Unified service tagging with the reserved tags env, service and version ties these streams together. When tags are applied consistently, a single service page shows its request rate, error rate, latency, recent deployments and related logs; when tags are inconsistent, correlation silently breaks.
Main product areas and what they cover
Datadog is sold as a set of separately priced products on a common data platform. Organizations usually start with infrastructure monitoring and add further modules as their needs grow.
| Product area | Main function | Typical users |
|---|---|---|
| Infrastructure monitoring | Host, container and cloud metrics | Operations and platform teams |
| APM | Request traces and service maps | Backend developers, SRE teams |
| Log management | Ingestion, search and archiving | Developers, security analysts |
| Real user monitoring | Front-end performance and errors | Web and mobile teams |
| Synthetic monitoring | Scheduled API and browser tests | Quality and reliability teams |
| Cloud security | Threat detection and posture checks | Security operations teams |
On top of the data, Datadog provides dashboards, monitors and service level objectives. Monitors evaluate metric thresholds, anomalies, forecasts or log patterns and notify on-call tools. SLO tracking expresses reliability targets based on a service level indicator and shows the remaining error budget over a rolling window, typically 7, 30 or 90 days.
Datadog compared with open-source observability stacks
The main alternative to Datadog is a self-managed open-source stack, commonly Prometheus for metrics, Grafana for visualization, Loki or Elasticsearch for logs and Jaeger or Tempo for traces, often fed by OpenTelemetry. The choice is largely a trade-off between operating effort and license cost.
| Aspect | Datadog | Open-source stack |
|---|---|---|
| Setup time | Hours to days | Days to weeks |
| Operations burden | Handled by the vendor | Carried by the internal team |
| Cost model | Per host, per GB, per span | Infrastructure and staff time |
| Correlation across signals | Built in through tags | Requires careful configuration |
| Data control | Stored in vendor regions | Stored in own environment |
| Vendor lock-in | Higher with proprietary agents | Lower with open standards |
Datadog accepts OpenTelemetry data, which reduces lock-in at the instrumentation level, but dashboards, monitors and query syntax remain proprietary. Teams that expect to switch vendors later often instrument with OpenTelemetry from the start and send data to Datadog through the OpenTelemetry Collector.
Pricing, cost control and metric cardinality
Datadog cost is driven by usage-based billing dimensions, and uncontrolled growth in any of them is the most frequent complaint about the platform. Infrastructure is billed per monitored host, logs per ingested and indexed volume, APM per host plus indexed spans, and custom metrics per unique time series.
Metric cardinality is the critical factor for custom metrics. Each unique combination of metric name and tag values counts as a separate series, so adding a tag such as user_id or request_id can multiply one metric into hundreds of thousands of series and a matching bill.
- Use Logging without Limits style pipelines: ingest broadly, but index only the logs needed for search and archive the rest cheaply.
- Apply retention filters for traces so that errors and slow requests are kept while routine requests are sampled.
- Keep high-cardinality identifiers out of metric tags and put them in logs or span attributes instead.
- Review the usage attribution reports monthly to see which teams and services drive cost.
Common mistakes when adopting Datadog
Most problems with Datadog come from configuration and governance, not from the product itself. The table lists frequent symptoms with their causes and fixes.
| Symptom | Cause | Fix |
|---|---|---|
| Unexpectedly high invoice | Unbounded tags or indexed logs | Set cardinality limits and exclusion filters |
| Logs not linked to traces | Missing trace IDs in log records | Enable log injection in tracing libraries |
| Too many pages at night | Static thresholds on noisy metrics | Alert on symptoms and SLO burn rates |
| Dashboards nobody trusts | Inconsistent service and env tags | Enforce unified service tagging |
| Blind spots in short-lived workloads | No agent on serverless functions | Use the serverless integration or extension |
Alert fatigue deserves special attention. Alerting on every CPU spike trains engineers to ignore notifications; alerting on user-facing golden signals such as latency, errors, traffic and saturation keeps pages rare and meaningful.
The idea in brief
Modern software runs on many machines and services at once, so when something goes wrong it is hard to see where. Datadog works like a control room with one wall of screens for all of them.
Small programs and connectors send it numbers, messages and request timings from everywhere. Datadog puts them side by side, marks them with the same labels and raises an alarm when something looks wrong. Engineers find the cause faster because every clue is already in one place, but the service is paid for by usage, so careless data collection becomes expensive.
A practical scenario
An online travel agency runs about 40 microservices on Kubernetes in a public cloud. The platform team deploys the Datadog Agent as a DaemonSet, enables APM in each service and applies the env, service and version tags in the deployment pipeline. Monitors alert when the booking service misses its 99.9% availability SLO at a fast burn rate.
One evening the booking error rate rises to 4% within ten minutes of a release. The on-call engineer opens the service page, sees that errors started with the new version, follows a failed trace to a payment adapter and reads the linked log line showing a changed API field. The team rolls back in under fifteen minutes, and the incident consumes only a small part of the monthly error budget instead of the full night it would have taken without correlated data.