Picture a small online bakery that takes pre-orders for weekend boxes. On Friday evening the order form hums along, confirmation emails go out, and the kitchen printer spits tickets. On Monday morning the owner discovers that eleven orders never produced a ticket. Nobody saw an error. The orders simply vanished somewhere between the form and the kitchen.
That small disaster is the reason this guide exists. It walks through, in order, how to build a safety net for background jobs, which is the most practical slice of ServerlessOps for anyone running a modest website. Following the steps requires no engineering background, although whoever maintains your site will need to do the wiring, so pass the list along.
Why background jobs fail quietly
When a customer clicks a button and the page waits for a result, a failure is loud. When the work is handed off to run later, a failure is silent unless you have arranged otherwise. The Ops Atlas entry on operating function-based systems describes the platform behaviour in numbers: for background invocations, events are held for up to six hours and retried up to two times. After that, without a destination for failures, they are simply gone.
The fix is not clever. It is a sequence of small, boring decisions, and the rest of this page lists them in the order we would make them.
Step 1: List every handoff
Grab a sheet of paper and write down each place where your site says “I will do this later”. Typical entries: send a confirmation email, update stock, create a delivery label, notify a shipping partner, store a receipt. Our own story about mapping a services roadmap for a new client begins the same way, with an inventory before any build work.
Count them. Most small shops land between five and fifteen. The Atlas calls this exercise the first stretch of a ninety-day plan: inventory triggers, failure queues and timeouts before you change anything.
Step 2: Give each handoff a holding place for failures
For every item on your list, ask where a failed attempt will rest. The answer should be a dedicated queue, usually called a dead-letter queue, or a failure destination configured on the function. Think of it as the tray on the bakery counter where a ticket goes when the kitchen cannot read it. Without the tray, the ticket falls on the floor.
Step 3: Name an owner
A tray that nobody checks is a bin. The Atlas lists the unowned dead-letter queue among its anti-patterns, and its remedy is a named person or team for each one. Write the name next to the handoff on your paper. If it is you, put a recurring reminder in your calendar; if it is a freelancer, put it in the contract.
Step 4: Wire an alarm
Owners are busy, so the tray should shout when it fills. Ask your developer to raise an alert when anything lands in a failure queue, and a second one when the oldest item has been sitting too long. The Atlas names backlog age and the age of the oldest failed message among the signals it highlights, and it points out a trap: alerting only on the count of errors can miss failures of individual items inside a batch, because those are not reported as errors by the platform.
Step 5: Keep failures longer than the source keeps its work
Retention is the quiet hero. If your main queue holds a message for four days and your failure queue only holds it for one, a long weekend can erase the evidence. The Atlas example pairs a fourteen-day failure queue with a four-day source. Check the numbers on your own setup and make sure the safety net always outlives the thing it protects.
Step 6: Make the retry harmless
Once failed items can be replayed, you need to be sure that replaying does not double up an effect. In our bakery, a replayed order must not print two tickets or charge twice. The Atlas advises storing the order reference as the deduplication key and handing it on to anything downstream. It warns against using the platform’s own message identifier, because a redrive gives a fresh one. A convenient side effect is that redriven messages restart their retention clock, so replays are safe from a premature expiry.
A friendly way to test: replay the same ten orders on a copy of your site and count the tickets. The right answer is ten.
Step 7: Handle partial failures in batches
Some jobs arrive in groups. If one item in a group of ten fails, you do not want the other nine retried forever. The Atlas recommends partial batch responses, where the function reports exactly which items failed so that only those return to the queue. For stream-style sources it also suggests finite retries and a maximum record age, since the defaults it mentions retry indefinitely.
One more point on batches: the bakery owner found that a single malformed order, with an empty delivery date, had been poisoning a whole group of ten. Once only the broken item went back to the tray, the other nine flowed through and the owner could fix the single bad record by hand in a couple of minutes.
Step 8: Put a ceiling on parallel work
Our bakery story has a sequel. A holiday promotion sent a thousand orders in ten minutes, the ticket printer service choked and every order retried at once, making things worse. A cap on parallel work would have spared the downstream service. The Atlas suggests keeping peak demand under seventy percent of the available limit and notes that a platform-wide pool of 1,000 is a typical default. If you have not seen how traffic surges stress a stack, our companion read on building a blueprint for digital growth sets the scene.
It also helps to think about what the cap does to customers. Orders above the ceiling wait in line a little longer instead of tipping the printer service over, and waiting is almost always kinder than failing. A brief delay is invisible to a shopper; a lost order is not.
Step 9: Rehearse a bad day
The Atlas treats a deliberate rehearsal as the threshold of its third maturity level: a duplicate-storm game day in which the same messages are sent repeatedly and you confirm there are zero double effects. Schedule one for a quiet afternoon. Send the duplicates, break the downstream partner on purpose, and watch whether the tray fills, the alarm rings and the owner reacts.
Step 10: Review the numbers monthly
Set a thirty-minute slot each month to look at four things: how many items entered failure queues, how old the oldest one was, how many requests were refused for exceeding limits, and what a million events would cost at current volume. In the Atlas worked example, the whole path cost about $3.75 per million events, or $112.58 a month at its sample volume. Your numbers will differ, but the habit of looking is the point.
What a long outage teaches
The entry also draws a lesson from a widely reported cloud disruption in October 2025, in which processing of queue messages and background invocations was deliberately slowed to protect interactive traffic and the backlog cleared the following afternoon, Pacific time. The takeaway for small operators is modest and useful: choose retention and duplicate-check expiry that comfortably cover an outage of many hours, so that everything queued during the trouble can still be processed afterwards. If you want a wider view of how disruptions affect plans, our road-trip piece on weighing vehicle options for a Colombian itinerary makes the same argument about spare time in a schedule, just with fewer queues.
Back to the bakery
Eleven missing orders led the owner to do the ten steps above over two weekends. The next promotion produced a fat batch of orders, one downstream hiccup and a single alert at lunchtime. Two tickets sat in the tray; the owner replayed them after the printer service recovered, and nothing was printed twice. That is what ServerlessOps looks like when it works: not a dramatic rescue, just a tray, a name, an alarm and a calm afternoon.

