Observability Is Converging. Humans Aren’t the Only Ones Querying It Anymore
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
Unified observability storage combining metrics, logs, and traces in a single columnar store is now mature technology, with SigNoz, ClickStack, Honeycomb, and Grafana's LGTM stack each taking different architectural approaches. Historical context traces the three-pillar model's origins to independent tool lineages (RRDtool/Prometheus for metrics, Splunk/Elasticsearch for logs, Dapper/Zipkin for traces) and the unification arguments from Peter Bourgon and Charity Majors dating back to 2017. The piece then argues the real frontier is agentic access: SigNoz, ClickHouse, Grafana, and Honeycomb have all shipped MCP servers and AI-native tooling in 2026, and ClickHouse found that purpose-built semantic tools for agents outperform raw SQL access. Open questions remain about how agent-driven exploratory query patterns should reshape data layout, query planning, and where observability semantics should live in the stack.
Table of contents
Unified storage is no longer a proposalTwo engineering choices, and one assumption that no longer holdsTo understand this, look at how it split apartGet Greptime ’s stories in your inboxThe unification argument goes back eight yearsBy summer 2026, everyone had seen agents comingHow far down does the change go?Questions this post answers
What is ClickStack and when did ClickHouse release it?
ClickStack is an open-source observability stack from ClickHouse that packages the OpenTelemetry Collector, ClickHouse itself, and a query experience into one system for exploring metrics, logs, and traces together. It followed ClickHouse's acquisition of HyperDX in March 2025, with ClickStack shipping that same May. Full-text search in ClickHouse is generally available, though it does not implement BM25-style relevance scoring; the TimeSeries engine and PromQL support remain experimental. Track how observability stacks like ClickStack evolve by following backend infrastructure coverage on daily.dev.
Why do coding agents perform worse with a general-purpose ClickHouse MCP server than with specialized observability tools?
Models perform noticeably better against structured investigative primitives than against raw SQL exploration, according to ClickHouse's internal evaluations. While building AI Notebooks, ClickHouse found that wrapping log pattern analysis, trace outlier investigation, and cross-signal correlation into semantic tools for agents reduced tool calls and improved consistency compared to their existing general-purpose MCP server. Developers wiring agents into observability platforms can compare MCP server designs on daily.dev.
Who first proposed unifying metrics, logs, and traces into a single observability system?
Peter Bourgon proposed the idea in his August 2018 post 'Observability signals,' suggesting an uber-system that takes raw events at the front door and de-multiplexes them by shape into purpose-driven backends. He had earlier popularized the three-pillars framing in a February 2017 post after the Distributed Tracing Summit, making him both an originator of the pillar model and one of its first critics. Engineers debating observability architecture can dig into this history on daily.dev before picking a stack.