ServerlessOps in Plain Language for Safety-Minded Aviation Teams

Safety-minded people in aviation share a few reflexes: report what went wrong, find the cause, fix the process and check that the fix held. Those same reflexes turn out to be the heart of a software practice called ServerlessOps, which is about keeping small, event-triggered cloud programs reliable. You do not need a programming background to follow it. This explainer uses ordinary language and a hangar-and-cockpit vocabulary to describe what the practice is, what goes wrong without it, and which five habits do most of the work.

First, what is a “serverless function”?

Picture a ground crew member who only appears when an aircraft arrives, does one defined job and leaves. Nobody pays for the crew member to stand idle, and if ten aircraft arrive at once, ten crew members appear. A serverless function is the software version: a short piece of code that the cloud runs when something happens, such as a form being submitted or a file being uploaded, then switches off. You pay for the moments it actually runs.

The convenience is real, and so is the catch. Because so many tiny pieces start on their own, you need rules about what happens when one of them stumbles, when two start for the same reason, or when a hundred start at once. Those rules are the practice. A reference entry from the Ops Atlas, the page that documents ServerlessOps, is a handy place to see the full list.

Is it an official standard?

No. The Atlas is clear on this: it notes that the name was used mostly between about 2016 and 2020, that no standards body or certificate sits behind it, and that its recommendations are collected from platform documentation and published engineering experience. Think of it as shared good practice, much like the informal habits pilots pass to each other, rather than a regulation.

Habit one: make repeats harmless

Cloud event systems tend to deliver a message at least once, which means occasionally twice. Imagine a work order that is accidentally printed two times. If both copies trigger a part replacement, you have fitted two parts. The fix is to give each job a unique, meaningful reference, such as the work-order number, and to check it before acting.

The Atlas stresses one subtlety. The number the messaging system stamps on a message is the wrong reference, because a message sent again after a failure gets a different stamp. Use the number that belongs to the real-world job.

Habit two: give failures a shelf and a person

When a part fails inspection, it does not simply disappear; it goes to a tagged quarantine area, and someone is responsible for it. Software needs the same thing. A failure queue collects work that could not be completed, so that people can examine it, fix the cause and send it again. The Atlas warns against the unowned version, a pile with no alarm and no reader. It also says the shelf should keep items longer than the original queue, so that a long outage does not wipe the evidence.

Habit three: set limits before the rush

Airports meter arrivals because runways have capacity. Cloud accounts carry a shared ceiling on simultaneous function runs; the Atlas mentions a default of 1,000 for a regional pool on the main provider. If one busy function eats the allowance, the others starve. Setting a limit per function, and staying under roughly 70 percent of the total at peak, keeps one runaway from grounding everything.

Habit four: keep the clocks nested

A pre-flight sequence has time boxes that fit inside each other: the boarding window sits inside the departure slot, which sits inside the crew duty period. Software timeouts should work the same way. A function’s request to another system ought to be abandoned earlier than the function’s own deadline, otherwise the function is cut off with no time to report what happened. The Atlas also suggests that a queue’s hide-the-message window be about six times the function’s time limit, so a slow run does not get duplicated by an impatient queue.

Habit five: watch the right gauges

A pilot does not stare at one gauge. The Atlas lists a small panel for functions: how many jobs truly achieved their business result, how old the waiting backlog is, how old the oldest failed item is, how often calls were throttled, how much spare parallel capacity exists, and what a million events cost. A single “errors” gauge can read zero while items are silently being sent back for another try, so it is never enough alone.

A tiny scenario

Suppose an academy’s scheduling system sends a notification whenever a lesson is moved. A glitch causes the same event to arrive three times. Without habit one, a student receives three messages and a refund routine runs three times. With a work reference, the second and third copies do nothing. If the notification service is down, habit two keeps the messages on a shelf with an alert. If a hundred lessons are moved at once after bad weather, habit three stops the notification function from using up the account’s shared capacity. None of this needs clever technology, only agreement beforehand.

Common slips beginners make

Beginners tend to trip on three things. One is trusting a platform’s marketing phrase, such as exactly-once delivery, without testing what happens on a repeat. Another is wiring several layers to retry on their own, which multiplies traffic against a system that is already struggling; the Atlas quotes five layers with three attempts each producing 243 times the load. The last is waiting for a real incident to discover that nobody knows who reads the failure shelf.

A useful corrective is a rehearsal. Send the same test event twice, slow one dependency on purpose, and watch what the system does. Aviation calls this practising the abnormal in a safe place.

What to do next

  1. List the automated actions in your system that affect a real person or a real record.
  2. Take each action in turn and consider the outcome if it repeats, arrives late or never fires.
  3. Name a person who reads the failure shelf every week.
  4. Pick two gauges beyond “errors” and look at them regularly.

The Atlas also offers a five-level maturity ladder and a ninety-day plan, but both are its own suggestions. The four questions above are a sufficient starting point, and many teams never get that far.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top