ResearchMay 1, 2026 · 10 min read

Davson's Distinction: Primary-State Observability for Distributed Systems

Cyril W. Davson, the English engineer who in 1955 published The Physics of the Primary State of Matter, was the contemporary who took Karl Schappeller seriously enough to systematize him. The book is widely remembered for its English-language exposition of Raumkraft. Its more durable contribution is a clean operational distinction between two kinds of measurement — and that distinction, transposed into telemetry vocabulary, is the most useful frame the team has found for thinking about observability in distributed systems.

The standard taxonomy of telemetry — metrics, logs, traces — is taught in every observability course and is essentially correct as a description of what data exists. It is essentially wrong as a guide to what to measure. The wrongness is not in the taxonomy; it is in the level of analysis. Metrics, logs, and traces are all downstream of the actual quantity an SRE cares about, which is the state of the medium through which the system is operating.

This is exactly the distinction Davson was trying to draw, in different vocabulary, in 1955.

What Davson Wrote

Davson's distinction is set out clearly in the opening chapter of The Physics of the Primary State of Matter:

It is necessary to distinguish carefully between two classes of measurement. The first class records phenomena occurring within material substance and is the proper subject of secondary physics. The second class concerns the state of the medium itself prior to its perturbation by material substance, and is the proper subject of primary physics. The error of the present age is to mistake the first for the second.Davson, 1955, paraphrased

For Davson, primary measurements were measurements of the medium's intrinsic state — its tension, its permeability, its capacity for storing latent energy. Secondary measurements were of the matter that perturbed the medium. The two were related but not interchangeable, and a science that confused them was a science that would forever be chasing its own perturbations without ever reaching the underlying state.

Two ornate Victorian-era brass-and-copper observatory telescopes mounted side by side on a dark obsidian platform under a starlit night sky — left telescope's eyepiece glows warm amber as it points slightly upward, right telescope's eyepiece glows cool teal as it points slightly downward, with a small distant observatory dome silhouette in the background — Fig. 1 — Davson's distinction in one image. Two instruments, two classes of measurement, two different things to look for.

The Telemetry Translation

Substitute for “medium” the word “system” — meaning the running production environment, the network, the machines, the dependencies, the queues — and the distinction is exact. Most telemetry is secondary: it measures what the application emits in response to load. Primary telemetry measures what the system could do under load, before the load arrives.

Concrete examples of secondary measurements, all useful, all secondary:

Request rate, error rate, p99 latency at the load balancer.
CPU utilization, memory residency, GC pauses on the application tier.
Dapper-style trace spans showing which downstream services were called and how long each took.

The primary equivalents — measurements of the medium itself — that almost nobody collects:

Headroom-to-saturation across all queues, in seconds. Not the queue depth (secondary), but the time until saturation at the current arrival rate (primary). A queue at 80% depth is not interesting; a queue with 8 seconds until saturation is.
Service-time distribution under counterfactual load. Not the observed p99 (secondary), but a periodic shadow trace through the system at a specified load level that records what the p99 would be at that load. The right cadence is workload-dependent; what matters is that the trace exists at all.
Dependency liveness as a vector field. Not the binary up/down of each dependency (secondary), but the probability of degradation of each dependency conditional on the recent state of every other dependency. A small graph embedding refreshed at a useful cadence captures the structure, drawing on the causal-inference tradition for the conditional shape.
Configuration drift relative to the canonical state. Not the current configuration (secondary), but the diff between the running configuration on every node and the configuration the deployment system thinks is running. The diff is non-zero in every production fleet anyone has ever measured; the question is how big it is and how fast it grows.

Eight ornate brass and copper analog pressure gauges mounted in a 4-by-2 grid on a dark obsidian dashboard panel — top row gauges glow with safe teal needles in the middle of their arcs, bottom row gauges have needles deep into amber warning zones and coral red danger zones — Fig. 2 — Headroom-to-saturation, instrumented as a primary measurement. Each gauge reads *seconds until this queue saturates at the current arrival rate*, not *current depth*.

What Changes

An observability stack built around primary measurements has a qualitatively different operational character from one built around secondary measurements. Three differences are worth naming:

The dashboards stop being about incidents. Secondary metrics are most informative during an incident. Primary metrics are most informative the hour, day, or week before an incident. The dashboard's purpose shifts from triage to forecast.
The alert rules stop firing on symptoms. A traditional alert fires when the p99 latency exceeds a threshold. A primary-state alert fires when the system's capacity to maintain a target p99 has degraded — which is a strictly earlier signal and which is almost always actionable while the secondary alert is almost always a notification of an outage already in progress. The Google SRE book calls these symptom-based versus cause-based alerts; the primary-state framing extends the terminology one layer further.
The post-mortem changes shape. A secondary-stack post-mortem is dominated by the question what happened. A primary-stack post-mortem is dominated by the question why did the medium permit this. The second question generates better fixes.

A sparse graph of 30+ service-mesh nodes connected by edges across a deep navy field, with a flowing vector field of tiny curved amber arrows showing the dependency-conditional liveness gradient curling around degraded clusters and streaming away from healthy ones — Fig. 3 — Dependency-conditional liveness as a vector field across the service mesh. The arrows are the gradient of the conditional degradation probability; the heatmap is the integrated risk.

Where This Sits in the Stack

The vocabulary belongs in research notes, not in a shipped sidecar product. The four primary-state metrics — headroom-to-saturation, counterfactual p99, dependency-conditional liveness, configuration drift — can be computed from the same OpenTelemetry signals an operator already exports, with sufficient additional instrumentation to maintain the counterfactual shadow trace and the configuration diff. The implementation question is unglamorous and load-bearing: who owns the per-service primary-state computation, where does it run, and what is the storage tier for the resulting time series. Bryan Cantrill and the broader SRE tradition have been arguing some version of this for fifteen years; the Davson framing is the cleanest single-sentence statement of the principle the field has converged toward.

The Davson Lineage

The methodological correction Davson proposed for physics — measure the medium, not the perturbation — is exactly the correction operations engineering needs in 2026, and the discipline of asking what does the medium permit rather than what just happened generates strictly better fixes. Davson's 1955 exposition remains the most coherent English-language treatment of the primary/secondary distinction we have found, and the Bahn collection of correspondence preserved at Rex Research extends the methodological argument with the engineering detail Davson omitted from the book itself. The book is in the public domain. The vocabulary is free. The discipline costs whatever the additional instrumentation costs, and that cost is the right thing to pay.