Observability

Zipkin17.4k★

Zipkin pioneered distributed tracing — it's the tool Twitter built to understand how requests flow through microservices. Send spans from your services, and Zipkin gives you a visual trace showing exactly where latency lives. If you're debugging slow requests across microservices, you need distributed tracing, and Zipkin is the battle-tested option. Jaeger (from Uber, now CNCF) is the more modern alternative with better Kubernetes integration and a richer UI. OpenTelemetry is the standard for instrumentation and can export to either. Commercially, Datadog APM, New Relic, and Honeycomb offer tracing with much better UX. The simplicity is appealing. Zipkin runs as a single Java jar, accepts multiple wire formats, and the dependency graph view is genuinely useful for understanding service topology. The catch: Zipkin shows its age. The UI feels dated compared to Jaeger or commercial tools. The query capabilities are limited — complex trace analysis requires exporting data elsewhere. Most teams starting fresh in 2026 should look at Jaeger or go straight to OpenTelemetry Collector with a backend that supports trace analytics. Zipkin earned its place in history, but the ecosystem has moved on.

VictoriaMetrics16.6k★

VictoriaMetrics is Prometheus on a diet that somehow got stronger. It's a time-series database and monitoring solution that uses up to 10x less RAM and disk than Prometheus, while staying fully compatible with PromQL. Drop-in replacement, better performance. If you're running Prometheus and hitting memory limits or storage costs, VictoriaMetrics is the upgrade path. It handles long-term storage, downsampling, and multi-tenant setups that vanilla Prometheus struggles with. Thanos adds similar capabilities to Prometheus but with more moving parts. Mimir (Grafana) is the enterprise alternative. InfluxDB serves different use cases — general time-series vs. metrics-focused. Best for teams already invested in the Prometheus ecosystem who need it to scale. Solo developers probably don't need it — Prometheus alone is fine for small setups. The catch: the cluster version's advanced features (downsampling, multi-tenancy) are enterprise-only. The single-node version is fully open source and covers most needs, but if you're scaling horizontally, the free tier has limits. Also, the query language additions beyond PromQL create vendor lock-in.

Thanos14.0k★

Thanos takes Prometheus and makes it work at scale. Long-term storage in object stores (S3, GCS), global query across multiple Prometheus instances, downsampling for historical data, and high availability. It's Prometheus without the "what happens when the disk fills up" problem. If you're running Prometheus and hitting storage limits or need to query across clusters, Thanos is the standard upgrade path. Cortex solves the same problem with a different architecture (multi-tenant by design). Mimir from Grafana Labs is the newer option with better performance. Commercially, Grafana Cloud manages this stack for you. The sidecar approach is elegant — add Thanos to existing Prometheus instances without changing your setup. Compaction and downsampling keep long-term storage costs manageable. The catch: Thanos adds operational complexity on top of an already complex Prometheus stack. Object store configuration, compactor tuning, and query latency for historical data all need attention. If you're running a single Prometheus instance and it's working fine, you don't need Thanos yet. And Mimir is increasingly the recommended choice for new deployments in the Grafana ecosystem, making Thanos feel like the "previous generation" solution.

OpenTelemetry Collector6.8k★

The OpenTelemetry Collector is the great equalizer of observability — a vendor-neutral pipeline that collects logs, metrics, and traces from your apps and routes them anywhere. Prometheus, Jaeger, Datadog, Grafana Cloud — send the same telemetry to multiple backends simultaneously. Swap vendors by changing a config file, not your code. If you're tired of vendor lock-in from Datadog or New Relic, the OTel Collector is your escape hatch. Deploy it as a sidecar or gateway, and your instrumentation stays portable forever. Datadog's agent is simpler but sends data to Datadog only. Prometheus handles metrics with a pull model. Vector is the faster alternative for log-specific pipelines. The catch: OTel Collector is a pipeline, not a backend. You still need somewhere to send data — and running Jaeger plus Prometheus plus Loki is real infrastructure. The configuration is YAML-heavy with receivers, processors, and exporters that can get complex fast. And "vendor neutral" means no vendor's support team helps you debug your Collector config.