Skip to content

Eight ServerlessOps Signals Worth Tracking, in Order of Value

Published November 16, 2026 by Beena Yadav

Dashboards for function-based systems tend to grow like attics. Every incident adds a graph, nobody removes one, and eventually the page that should answer “are we fine?” needs a guided tour. This list takes the opposite approach. It names eight signals, ranks them by how much they tell you per minute of attention, and says how to calculate each one and where it can mislead. The ordering is my own judgement, so argue with it freely; the underlying set is taken from the metrics list in the Atlas’s ServerlessOps reference page.

A note on reading the list. A signal that costs nothing to collect but changes nobody’s behaviour does not earn a place. Each item below is paired with the decision it should trigger.

1. Business success rate

What it tells you: whether the outcome the business wanted actually happened. An order was confirmed, a document was filed, a payment settled exactly once.

How to compute it: count completed business outcomes and divide by outcomes that were expected. The numerator comes from the system of record, not from function logs, since a function can finish without error and still have achieved nothing.

Where it misleads: if the denominator is also derived from the same pipeline, lost inputs vanish from both sides and the rate stays at 100 percent. Take the denominator from the entry point, such as the front-end’s record of submitted requests.

Decision it triggers: any sustained drop is an incident, whatever the platform graphs say.

2. Duplicate side-effect rate

What it tells you: how often something irreversible happened more than once, such as a double email, a repeated charge or a second shipment.

How to compute it: count repeated side effects per business key over a period, divided by total side effects. Your duplicate-guard table is the natural source.

Where it misleads: it only counts duplicates you can recognise. If your guard keys on the wrong identifier, repeats look unique. The Atlas warns about keying on the platform’s message identifier, because replays receive fresh ones.

Decision it triggers: a non-zero figure on a critical path means the idempotency design has a gap; the target in the Atlas’s maturity ladder for its third level is zero.

3. Failure-queue inflow and oldest age

What it tells you: the volume of work the system gave up on, and how long it has sat untouched.

How to compute it: two numbers, messages arriving per hour and the age of the oldest message, both available from standard queue metrics.

Where it misleads: a quiet queue is good news only if every asynchronous path actually has one. A path with no failure destination produces no inflow because the events simply vanish. Audit coverage before celebrating a flat line.

Decision it triggers: a named owner triages whenever inflow is non-zero, and anything approaching the retention limit gets replayed or discarded on purpose.

4. Backlog age

What it tells you: how stale the oldest unprocessed item is. Platforms expose it under names such as iterator age for streams and async event age for function events.

How to compute it: read it straight from the platform metric and alarm on its maximum, not its average.

Where it misleads: averages hide a single stuck record. Also, a healthy-looking function error count can coexist with a growing backlog, because item failures reported through partial responses are not counted as function errors.

Decision it triggers: when age rises past a threshold tied to a business promise, such as “customers hear back within ten minutes”, scale, fix or shed load.

5. Throttle ratio

What it tells you: how often the platform refused work because a limit was hit.

How to compute it: throttled attempts divided by total attempts. The Atlas illustrates with a modelled day on a shop-style workload: 63,000 throttled attempts among 31.5 million, which is 0.20 percent.

Where it misleads: a tiny ratio over a day can hide a severe burst at a single minute. Look at the peak minute as well as the daily figure.

Decision it triggers: a rising ratio means raise the cap, smooth the arrival or protect the neighbours with reserved capacity.

6. Concurrency headroom

What it tells you: how far the busiest moment sits from the ceiling.

How to compute it: one minus peak concurrent executions over the applicable cap. The Atlas proposes a warning when the remaining share dips under thirty percent, and a design target of keeping peak use at about seventy percent of a cap.

Where it misleads: the cap that matters may belong to a downstream system and not to the platform. A function with enormous headroom can still overwhelm a database that takes only forty connections.

Decision it triggers: shrink the function’s own ceiling to match the weakest dependency.

7. Cold-start share

What it tells you: the fraction of runs that paid a start-up delay.

How to compute it: runs flagged with an initialisation duration, divided by all runs, taken from the platform’s report lines.

Where it misleads: it is easy to over-react. The Atlas says the share is small on the provider it examines, below one percent of invocations, so a team seeing ten times that figure has a specific cause such as a large package or heavy start-up code, and should investigate that cause, not add warm-up pings.

Decision it triggers: only act if the latency promise to users is actually affected.

8. Cost per million events

What it tells you: what each unit of work costs across the entire path, functions, queues, storage and data transfer included.

How to compute it: total monthly cost of the path divided by events processed, multiplied to a million. The Atlas’s illustrative path comes to $3.75 per million, near $112.58 a month at the volume it assumes, though your figure will depend on your design.

Where it misleads: it can improve while the real picture worsens if retries or duplicates are quietly inflating the denominator. Pair it with attempts per event, which the same worked example puts at 1.05.

Decision it triggers: a monthly review, and a conversation about whether a steady workload belongs on a different compute model.

How to adopt the list without drowning

Start with the first four. They answer the “are we fine?” question for most teams. Add signals five and six once concurrency limits are in place, and seven and eight in the monthly review, not the daily dashboard. Engineers charting a career in this area can fold the list into their growth plan. A piece on progressing from junior to senior engineer shows where operational literacy starts to matter, while a skills-and-projects map for the engineer’s journey leaves space for it too. Anyone coming from support work, as described in an account of crossing over from sysadmin and helpdesk roles, has probably watched queue depth for a living already. Several of these signals sit beside the pipeline-level monitoring those roles own.

One caution on thresholds. Do not import numbers blindly. A warning at thirty percent headroom is a sensible starting suggestion, not a law. Calibrate against your own history for a month, then set alarms that fire rarely and always mean something.

A weekly rhythm for the first four

Signals one to four deserve a standing slot. Spend ten minutes on Monday asking three questions of each: did it move, why, and what are we doing about it. Write the answers down in a running note, even if the answer is that nothing changed. After a couple of months that note becomes the most useful incident-prevention document your team owns, because it records which small drifts later turned into outages and which turned out to be noise.

For signals five to eight, a monthly review is enough, provided the alarms exist for the sudden cases. Keep a short list of decisions made in that review, such as raising a cap, shortening a timeout or moving a workload, so that the reasons survive staff changes.

Reading list for the curious

The Atlas also names a longer vocabulary: telemetry conventions that are still in development, an event specification that has reached graduation, and a number of anti-patterns that these metrics help detect, such as alarms limited to the error count. Reading the entry end to end takes about half an hour and pays for itself the first time an alarm stays silent during a real problem.