Operations teams that run online appointment and messaging platforms tend to describe trouble in terms of symptoms: the monthly cloud invoice looks wrong, reminders arrive twice, or a slow afternoon turns into a full outage. This article takes four such symptoms and traces each to a likely cause and a remedy, using the vocabulary of ServerlessOps, the practice of keeping small event-triggered cloud functions dependable. It is strictly about administration and software operations. Nothing here touches clinical subjects, and none of it is a basis for health or care choices.
How the diagnosis works
Good troubleshooting separates what you can observe from what you assume. Each section below starts with an observation, lists the cause that the published evidence supports, and ends with a change you can verify. The evidence comes from the Ops Atlas write-up on this practice, a public reference whose figures are mostly platform documentation and worked examples, and which is explicit that the term has no certifying body. The entry is available as the Atlas’s ServerlessOps reference, should you want to check any number cited here.
Symptom 1: the bill grew far faster than traffic
Observation. Request volume rose modestly, yet function charges multiplied.
Likely cause. A downstream dependency slowed down. Functions that call another service and wait are billed for the waiting. The Atlas includes a lab where a dependency became ten times slower per call. Monthly compute for the test workload went from under sixty dollars to over five hundred, and the busiest moment needed 840 parallel executions where 84 had sufficed, hard against a ceiling of one thousand. A tenfold slowdown produced roughly a ninefold bill and nearly exhausted the allowance.
Remedy. Shorten the timeout on the downstream call so it fails well before the function’s own limit, avoid chains in which one function waits on another, and track concurrency headroom with a warning once it drops below about 30 percent.
Symptom 2: a slow afternoon became an outage
Observation. One partner system degraded, and soon unrelated features stopped working.
Likely cause. Parallel capacity is shared across an account. If one function is allowed to grow without a ceiling, a spike takes the whole pool and starves its neighbours. Retries add fuel: with five layers each trying three times, the Atlas cites advice from an Amazon engineer that load on the struggling dependency can reach 243 times normal.
Remedy. Put an explicit cap on the spiking function, keep peak use under about 70 percent of the account total, and retry in one layer only, with growing pauses and a finite number of attempts.
Symptom 3: people received the same message twice
Observation. A reminder or receipt went out two or three times.
Likely cause. Event platforms deliver at least once, and some trigger types retry on timeout. The protection, if any, was keyed to the wrong identifier. The Atlas points out that a message sent again from a failure queue receives a new message ID, so a guard built on that ID is bypassed exactly when it is needed. Its test harness replays ten thousand orders seeded with some duplicates and late timeouts. A design that merely logged what it had seen still let 295 double effects through; logging combined with a business reference let none through.
Remedy. Choose a business reference such as an appointment number, check it before any side effect, pass it on to the next system, and keep the record for as long as your longest plausible outage.
Symptom 4: the error chart is quiet but customers are complaining
Observation. Dashboards show a low error rate, yet some notifications never arrive.
Likely cause. When a function handles a group of items, it can report success for the group while sending individual items back for another try. Those item failures do not count as errors. Meanwhile a failure queue with no alarm quietly fills. The Atlas lists “alarms on errors only” and “dead-letter queue as an unowned bin” among its anti-patterns.
Remedy. Raise alerts when the oldest waiting item gets stale or when the failure queue starts growing, name an owner who reads the queue, and enable partial batch responses so only genuine failures are retried.
A worked example of reading the signals
Suppose a team reviews a month of data and finds thirty million booking events handled through 31.5 million function runs, with 63 thousand calls throttled along the way. Dividing runs by events gives 1.05 attempts per event, so retries are modest. Dividing throttled calls by runs gives a throttle ratio of 0.20 percent, which is comfortable. These are the Atlas’s own sample numbers, and they show how a few divisions replace a vague feeling that things are fine.
Had the attempts-per-event figure crept toward two or three, that would point back to symptom two and stacked retries. Had the throttle ratio climbed, the cause might be a function missing its cap. The signals are cheap to compute and they tell you which section of this article to reread.
Checking that the remedies worked
A diagnosis is only complete when the symptom is gone for a measurable reason. Pick one number per symptom and look at it weekly: cost per million events, concurrency headroom, duplicate side-effect rate and failed-item age. The Atlas gives a worked cost example of $3.75 per million events along a sample path, which is useful as a shape to imitate rather than a target to copy.
A five-level maturity scale is also in the entry, which describes it as the authors’ own suggestion. The level labelled “safe” is reached when a deliberate duplicate-storm drill produces zero double effects. That is a clean, testable standard that any operations team can apply without extra tooling.
Related reading on this site
Cost is the thread that connects symptoms one and four, and our write-up on keeping cloud spend under control across several providers explores the budgeting side in more depth. For a lighter parallel on discipline with recurring commitments, a piece on splitting shared bills with a recurring payment tracker makes the same point at household scale: write the rules down before the money moves.