Opens in a new tab
Guide

Customer reliability begins with the upload that did not finish

In brief

Keep the customer’s operation intact as a report becomes a ticket, an investigation and an answer about whether to wait or retry.

5 min read

Is this the right starting point?
Before you read
No specialist engineering background; ordinary customer-support or service-work familiarity is sufficient.
What you will take away
Turn a customer failure report into scoped evidence, accepted investigation ownership and an update that accounts for unfinished customer work.
Sources
An amber inset in a green mosaic contains two coral pieces separated by a jagged gap; an amber line connects it to a card with two similarly shaped pieces.
On this page3 sections

In this hypothetical case, a customer is trying to upload a large file. It fails, so they contact the team running the service. Someone opens the dashboard and finds a healthy overall success rate. The file is still missing. The chart has not answered the customer's question.

Both accounts can be accurate. A small group of failed uploads can barely disturb a total dominated by successful requests. Customer reliability engineering brings the reported experience into the investigation: what this person tried, when they tried it and what happened. The customer does not need to name the broken component before engineering has somewhere useful to start.

The detail that survives the ticket

Start with the attempted operation, approximate time, time zone and interval over which the problem occurred. Add an identifier that safely locates the request, if one is available, and the scope the customer can describe. These details help the next person find the work without asking the customer to become familiar with the service's internal machinery.

Keep the customer's words alongside any standard category used to route the ticket. An illustrative report such as “Checkout appeared complete but no order arrived” tells us that something seemed to succeed before the expected result went missing. Reducing it to “latency” loses that sequence. A category can help sort a queue while the original description helps someone investigate it.

Use an approved channel and collect only what the investigation needs. A bounded request identifier may be enough to find the event; credentials and broad log dumps are not the price of being taken seriously. Preserve relevant records before they expire. It is much easier to work from them than to ask a customer to reconstruct yesterday's attempts from memory.

What to keepWhat it lets the next person do
Customer or tenant identifierLocate the affected account or group, with appropriate access controls
Operation and impact windowSelect the relevant measurement and period
Observed consequenceUnderstand what the customer could not finish
Evidence and unknownsKeep reported facts separate from possible explanations
Investigation ownerReach a staffed team and know how to escalate
Next communicationExplain what is known and when another update will arrive

This short record serves two people. The engineer needs enough detail to investigate the failure. The person keeping the customer informed needs a staffed route into that investigation and a time for the next update. If all they inherit is “possible outage,” both have to rebuild the conversation before they can help.

Give the upload team a question it can investigate

The team responsible for the user operation can begin while the cause is still unknown. It can inspect its service and involve the teams behind its dependencies. Waiting for the customer to prove which component failed delays that work and gives the least-equipped person the diagnostic task. Routing early does not mean declaring a cause; it means putting the unresolved question with people who can examine it.

For our hypothetical upload, the useful details are that large files fail in one region. Keep both. They identify a group of requests the engineers can inspect or try to reproduce. Remove object size and region, and the next person may return to the same broad chart that looked reassuring at the start.

Where the available system measurements and privacy constraints allow, compare the relevant endpoint, the service entry point handling the operation, along with region, version and workload. If this customer's requests make up only a small share of all traffic, failures concentrated there may scarcely change the overall average. Narrowing the view makes the reported pattern easier to see. It also narrows the conclusion: finding a problem in those uploads does not establish that every customer is affected.

From the reported upload to the work behind it
The customer symptom leads to the affected population, then a service investigation and customer update. Each step retains the operation and its context.
This illustrative investigation follows the reported operation into the relevant group of requests, then brings the findings back to the customer. It does not infer a cause from the report alone.
Follow the report through the investigation

The customer’s operation, time and impact identify the starting question. Tenant, region or request path narrow the group to examine. Relevant logs, traces and known changes inform the investigation; the update explains progress and returns to the original symptom. This is a conceptual path, not a completed incident. Diagram labels: Customer symptom: Operation, time and impact; Affected population: Tenant, region or request path; Service investigation: Relevant logs, traces and known changes; Customer update: Explain progress and confirm the original symptom.

The customer-facing update can now describe real progress: whether the team reproduced the symptom and which question it is checking next. Wider incident coordination can organize recovery while that owner explains the verified findings. A time for the next update is useful even when a restoration time is not yet known. Guessing when the service will recover adds a promise the investigation has not earned.

What should the customer do with the file now?

Suppose the investigation finds that the service accepted the upload, but a later processing step failed. The acceptance response at the application's interface, or API, described an earlier event. It did not mean the customer's work had finished. Explain that difference in terms of the file: what is working again, what happened to the earlier submission and whether any work is still pending.

That last question changes what the customer should do next. Waiting may be appropriate if the existing submission is still being processed. Repeating an action may create a duplicate if the service has already accepted it. Neither “try again” nor “everything is green” answers this safely without checking the affected operation and accounting for earlier submissions.

Once the immediate question is resolved, repeated reports can reveal work the service still needs. They should inform the service objective, the reliability target for the operation, and the record of remaining risk. The gap may call for measurements that cover this group of uploads, a better fallback when processing fails, or a clearer commitment about what the service provides. Closing the ticket alone leaves the next discovery waiting for another customer to speak up.

Our original file has led to a much more useful question than whether the dashboard is generally healthy. Can large uploads in that region finish, and what happens to the ones already submitted? Answering it helps the person who called. Making that failure visible also gives the next engineer a place to look before the next customer has to explain it all again.

Source context

This article does not include external reference links. Read it as the author’s perspective and evaluate the guidance against your environment.

Report an error or outdated detail

Keep reading

Where to go next