One date anchors the entire discipline: 1968, when Melvin Conway observed that a company’s systems end up shaped like the ways its people talk to each other. Everything now marketed as modern delivery practice — DevOps included — is an engagement with that observation. The evidence base has since matured from anecdote to measurement, and for aviation software teams, whose booking platforms and operational tools sit under unusual public scrutiny, the measured version of the field is the only one worth arguing about. This piece summarises where the evidence stands as of 2026, and where it is moving.
From a 1968 Insight to a Measured Discipline
The timeline matters because it explains why current practice looks the way it does. Conway’s 1968 observation established that structure and software are coupled. Decades of lean manufacturing thinking supplied flow concepts. Then came the research era: the Accelerate findings published in 2018 demonstrated statistically that delivery performance predicts organisational performance, and The DevOps Handbook, in its expanded 2021 edition, pulled the playbooks together. By 2025, the DORA research programme was reporting findings from the AI era, examining how machine assistance changes — and sometimes fails to change — delivery outcomes. Aviation adopted these ideas unevenly, as it adopts most things, constrained by certification cycles and legacy integration, but the direction of travel has been consistent for a decade. It is worth stating plainly for newcomers: nothing in the 2018 findings has been overturned. The correlation between delivery performance and organisational outcomes has been re-tested on larger samples, in more industries, and has held. What has changed since is the granularity — finer metrics, clearer definitions, and harder questions about causation.
What DevOps Research Measures Now
The current framework tracks five delivery metrics rather than four, a change practitioners should treat as settled.
- Lead time — from commitment to running in production.
- Deployment frequency — the cadence at which code reaches production.
- Time-to-restore — the speed with which a disrupted service comes back.
- Change failure rate — what fraction of releases degrade service.
- Rework rate — added in 2024, capturing delivered work that must be redone.
Two definitional changes deserve note. In 2023, recovery was redefined to count only change-caused failures, separating self-inflicted incidents from external ones. In 2024, rework joined the set, replacing the older “four key metrics” framing. Rework is the metric with the sharpest aviation application: an airline platform that ships quickly but repeatedly rebuilds its own fare logic is converting speed into churn, and no frequency dashboard will show it. Teams reporting only their flattering measures should expect that gap to be noticed.
How DevOps Adoption Actually Succeeds
The adoption evidence favours focus over scale. In one widely cited comparison, a pilot value stream coached by an enabler earned 4.25 of a possible 5; the central toolchain-first programme earned 2.40; the organisation-wide mandate 2.50. The gap between a coached pilot and a decree is the difference between learning and announcing. The maturity ladder in common use begins with siloed teams, climbs past automated islands, then integrated delivery, then continuous delivery, and ends with a learning system at the summit — and most aviation software organisations surveyed informally sit at the second or third rung, automating within teams while remaining disconnected across them. The ladder is climbed a rung at a time, which frustrates executives and reassures engineers. The ladder also explains a common measurement trap: organisations on the automated-islands rung report strong tool adoption and weak outcomes, because automation without integration moves local speed while global flow stays stuck. Rung, not budget, is what the outcomes track.
The 2025 Findings and the AI Question
The most recent research cycle turned to machine assistance, and its early findings are usefully deflating. Adoption of AI tools is nearly universal among surveyed teams; the gains are not. The pattern in the data suggests that individual productivity rises while systemic throughput barely moves, because the surrounding process — review queues, approval gates, brittle test suites — absorbs whatever the tools free up. For aviation software, the lesson lands neatly: an assistant that writes code faster does not help a team that deploys quarterly, for the same reason a faster engine does not help an aircraft waiting for a runway slot. The organisations benefiting most from AI-era tooling are, so far, the ones that had already shortened their feedback loops. The tools amplify a flow that exists.
The Platform Turn
Team Topologies reframed Conway’s law as a design tool rather than a curse: if structure shapes systems, structure can be chosen deliberately. The stream-aligned team owns a passenger-facing flow end to end; the platform team reduces its cognitive load by providing self-service capability. Aviation examples are easy to construct — a platform team offering certified deployment pipelines so each booking-market stream team does not rebuild its own. This is the organisational trend with the strongest current momentum, and it is organisational before it is technical, which is precisely why tool purchases keep disappointing the executives who make them. The practical test of a platform team is whether a stream team can go from idea to production without filing a ticket to another human. If it cannot, the platform is a queue with better branding.
What to Watch Through 2026
Three developments look set to shape the next survey cycle. Platform engineering will keep consolidating, as the gap between enabled pilots and decrees keeps being rediscovered at enterprise scale. Regulatory frameworks for software in transport will continue moving toward continuous-evidence models, which suits automated pipelines and penalises manual paperwork. And the definition churn in the metrics themselves — recovery redefined in 2023, rework added in 2024 — should stabilise, giving the industry its first long run of comparable numbers since the research began. Teams comparing themselves to published benchmarks should check which definition a given figure uses; mixing eras produces confident nonsense. A fuller orientation to the field’s vocabulary and evidence sits in the Ops Atlas reference on DevOps.
What It Means for Aviation Software
Three implications follow directly from the measured record. First, regulated contexts are not exempt from high performance: automated controls — access kept to a minimum, credentials that expire within the hour, expand-and-contract migrations, quarterly sampling of roughly ten random changes — produce audit evidence as a byproduct of delivery. Second, resilience is rehearsed, never assumed; the documented contrast between a 190-minute unrehearsed rollback and a 20-minute practised one is the single most persuasive number in the change-management conversation. Third, the J-curve is a planning parameter, not a surprise: automation raises the test bar before it raises throughput, and programmes that pre-agree an acceptable dip duration survive to see the climb.
A fourth implication is quieter and worth naming for training organisations. As delivery work becomes more measurable, the skills that distinguish candidates shift from tool familiarity toward systems thinking — reading a value stream, designing feedback loops, running a blameless review. Curricula that still centre on individual tools will age faster than the ones built around the measures. The operational skill set behind that shift — the practical site reliability practices — deserves its own treatment and more space than a state-of-the-field summary can give it.
The synthesis is straightforward. The discipline’s centre of gravity has moved from tools to systems thinking, and the organisations that treat delivery performance as a measurable property of structure — not of individual heroics — are the ones whose software keeps pace with their aircraft. The evidence base will keep moving; the habit of measuring it should not.