The global API management market reached $10.02 billion in 2025 and is on track to hit $108.61 billion by 2033, a compound annual growth rate of 34.7%. The numbers describe a category that has moved from developer tool to core enterprise infrastructure in less than a decade. APIs now sit underneath payments, logistics, mobile apps, B2B integrations, and the entire generation of cloud-native software being built on top of them. The question that gets less attention is what it takes to keep these platforms running when traffic surges, regions go dark, and a single misbehaving endpoint can ripple across a customer’s business.
That is the territory where Khushan Adatiya has spent more than a decade. A Senior Software Engineer at Google with prior experience at AWS, He has worked on infrastructure sustaining hundreds of thousands of transactions per second, with a specific focus on observability, governance, and security at scale. He also serves as a judge for the Globee Awards’ Artificial Intelligence. That experience offers a close view into the operational realities of API platforms, having contributed to systems operating at immense scale from within.
The Cost Curve Buried in Every API Platform
Gartner forecasts worldwide end-user spending on public cloud services to reach $723.4 billion in 2025, up 21.5% from 2024. Inside that total is a smaller, faster-growing line item that few executives notice until it gets uncomfortable: the bill for observing the cloud. API gateways, monitoring pipelines, telemetry stores, and the analytics layered on top of them are now meaningful cost centers in their own right. The harder problem is that telemetry traffic tends to grow faster than the business it monitors, which means a successful product launch can produce an unwelcome surprise on the next infrastructure invoice.
Adatiya was the core engineer behind Apigee Edge API Monitoring, the system known internally as Sonar. The product gives enterprise customers a real-time view into API traffic across their global footprint, and behind it sit thirty-two Dataflow pipelines and Bigtable clusters absorbing forty thousand writes per second at peak. His responsibilities ran from the dashboards customers actually saw, down through the resource decisions on the pipelines feeding them. The team’s working principle was that monitoring infrastructure could not be allowed to become more expensive to run than the API platform it watched.
“Observability is the only piece of infrastructure that gets blamed both when it costs too much and when it tells you too little,” Khushan Adatiya says. “Engineers have to design against both pressures at once. The moment you cut something to save money, you will be asked why you did not catch an incident sooner. Designing for that tradeoff is the actual work.”
From Alert Fatigue to Resolution
A New Relic study of 1,700 IT and engineering executives put the median cost of an operational IT outage at $33,333 per minute, with median annual losses from unplanned downtime totaling $76 million. The number explains why the language of monitoring has shifted from uptime to mean time to resolution. Customers no longer ask whether you have a dashboard. They ask how quickly your dashboard tells the on-call engineer where to look and what specifically to do next. The economics turn on minutes, not features.
Inside Apigee, Adatiya led the development of the “Distribution by MP Host” feature, which reduced triage time for internal teams from five or six minutes down to a single minute, a fivefold improvement in how fast on-call engineers could isolate where errors were originating. He also drove test coverage on the eventlog-dataflow-pipeline from thirty percent line coverage to eighty-seven percent, and from fourteen percent branch coverage to fifty-one percent. The combination is not glamorous engineering. It is the kind of work that quietly stops bad days from becoming worse ones.
“There is a temptation to design alerts around what is interesting,” Adatiya notes. “The harder discipline is designing them around what is actionable for the person paged at three in the morning. Half of the triage time disappears the moment an alert tells someone exactly which proxy, which region, and which playbook to open. The rest is engineering taste.”
Inside Apigee’s Quiet Cost-Cutting Operation
Cost discipline in cloud infrastructure has become a quiet competitive advantage in 2025 and 2026, with cloud finance practices moving from the FinOps team into the engineering organization itself. Observability costs in particular have grown in ways that are difficult to predict, driven by organic business growth, telemetry complexity, and rising expectations on what monitoring should cover. The companies pulling ahead are the ones treating cloud spend as an engineering responsibility rather than an accounting problem. That means engineers willing to read the bill and own what it says.
Adatiya, who also serves as a judge for the Gemini Live Agent Challenge, brought the same critical-evaluation discipline to the cost side of his Apigee work. He identified disk allocation patterns on Dataflow pipelines that were over-provisioned for typical workload and worked directly with the underlying Dataflow infrastructure team to tune them. He decommissioned redundant proof-of-concept Bigtable instances and built a process for proactively scaling down infrastructure after high-traffic events like Black Friday and Cyber Monday. The cumulative impact was four hundred fifty thousand dollars per year in direct infrastructure savings, with an additional twenty-five thousand dollars per day reclaimed during off-peak periods.
“Cloud bills do not negotiate,” Adatiya observes. “If a disk is over-provisioned, it is over-provisioned every hour of every day, and the only way to find that is to actually read the resource allocation patterns and ask whether they match what the workload needs. Engineers who can do that math save companies real money. It is not a glamorous specialty, but it is a serious one.”
The Workloads Pushing Observability to Its Limits
Observability has expanded well beyond backend services responding to user requests. The workloads stressing modern monitoring systems are different in kind. Event-driven architectures generate telemetry with no fixed schema. Streaming pipelines produce metrics that change shape as they flow. Generative AI workloads, now one of the fastest-growing segments in cloud, create traces that mean nothing in isolation but everything when correlated across context. The shape of what needs to be watched has outrun the tools built to watch it.
Adatiya has seen this evolution from inside one of the largest API platforms in the world. The Apigee work was, in its way, an early version of the problem: how to monitor systems whose traffic patterns changed faster than the monitoring infrastructure could adapt. What he took away from that work were operating principles. How to write tests against data that is itself in motion. How to design pipelines that degrade gracefully when downstream services slow. How to maintain test coverage when the system under test has hundreds of moving parts. None of these problems went away when the workload shifted. They got harder.
“The hardest part of observability has never been collecting data,” Adatiya explains. “It has always been deciding what is worth collecting, then trusting that what you collected actually represents what happened. That problem scales as fast as your workload does, and right now your workload is scaling faster than it has in a decade.”
The Next Monitoring Problem Is Already Here
The next decade of observability work will not look like the last one. Enterprises are now deploying AI agents that act on behalf of users, write to production systems, and coordinate across teams of other agents, and the monitoring problem these systems present looks nothing like what came before. Traditional observability tracks errors and latency. Agentic observability has to track reasoning chains, distinguish a slow plan from a wrong plan, and detect when an agent’s intended trajectory has quietly diverged from what its principal actually wanted. The discipline is being invented in real time.
Adatiya is one of the engineers thinking through what that next surface looks like. He recently published a HackerNoon analysis arguing that 2026 is the year enterprise software pivots from SaaS dashboards to Agents-as-a-Service, examining the architectural decisions and operational tradeoffs that come with treating agents as production services. The thesis sits naturally alongside his Apigee work. Once agents become billable, monitored, and SLA-bound services, every problem he spent a decade solving for API platforms returns in a more demanding form. Observability is no longer something AI teams will get to later. It is the substrate that decides whether the entire architecture is operable in production.
“Observability used to mean asking whether a service was up,” Adatiya reflects. “It will soon mean asking whether a system of autonomous agents is actually doing what you meant. The second question is harder, and the engineers who learn to answer it will be the ones who already learned to answer the first one carefully.”
Last updated: June 5, 2026
