Read a sample: 1. Start With the Application
A team is moving an internal document-processing application to Azure. Its first request sounds straightforward: “We need storage for our files.”
That sentence does not say whether the application opens files through a shared directory, uploads documents through an API, or stores temporary work items while another process runs. It does not say how quickly a document must appear, what happens if a region becomes unavailable, or whether an accidentally deleted file must be recoverable. Each answer could change the design.
The first useful storage decision is therefore not a service selection. It is a description of what the application and its operators need storage to do.
This chapter gives you a requirements worksheet for gathering that description. You can use it for a new application, a migration, an architecture discussion, or an exam scenario. The aim is not to produce a perfect specification before asking any technical questions. It is to make important assumptions visible early enough that they can be tested.
Azure offers storage services with different interfaces and operating characteristics. Choosing among them requires more than matching a word in a request to a product name. “Files” might mean documents that users upload and download, a shared directory that an application expects to mount, or files inside a virtual machine’s operating system. Those are related ideas, but they are not necessarily the same storage problem.
Start by describing the workload in ordinary terms:
- What information does it create or use?
- Which people or components read and change that information?
- How does the application expect to interact with it?
- What happens when the information is slow, unavailable, damaged, or deleted?
- How much information exists now, and how might that change?
A useful description is specific enough to test. “The application needs fast storage” is difficult to evaluate. “The application must open a document for an interactive user, and the team wants most opens to complete within two seconds under the expected workload” is more useful. It identifies an operation, an audience, and a target to investigate.
Even that statement needs context. Does the two-second target include network travel, authentication, application processing, and rendering? Storage contributes to the result, but it may not be responsible for all of it. A requirement should identify the experience being measured and, where possible, the part of the system responsible for meeting it.
The following worksheet is a working document, not a form that must be completed in one meeting. Record what is known, mark what is uncertain, and identify the person or evidence that can resolve each open question.
Two columns beyond the table are valuable in practice: confidence and owner. A statement supported by measured production logs has a different status from a guess made during a workshop. An assigned owner makes an unanswered question actionable.
You do not need a separate worksheet for every file or message. Group information by workload when it has similar behavior and requirements. Separate it when the differences matter. An application’s temporary processing messages, user-uploaded documents, and virtual machine system data may belong to the same business process, but they do not necessarily need the same storage model or recovery treatment.
Data shape describes what the application handles and how those pieces relate. A collection of independent images has a different shape from a directory tree that several applications navigate. A request waiting for a worker to process it has a different purpose from the business record that remains after processing.
Ask about size distribution, not just total capacity. One large dataset could contain a small number of large files or a very large number of small items. The total amount stored might be identical, while the number of operations, listing behavior, and transfer patterns differ considerably. Record a typical item size, the range, and any unusually large items if that information is available.
Then ask what happens to those items. For example:
- Are documents uploaded once and read many times, or edited repeatedly?
- Must a client list directories, rename files, or write to part of a file?
- Do several users or processes need access to the same data?
- Are work items processed asynchronously, with retries when processing fails?
- Does an application running in a virtual machine expect a disk and a file system?
These questions reveal the interface the application needs, not merely the kind of content it contains. A photo is a file in everyday language, but an application might handle it as an object uploaded and retrieved through an API. Conversely, a legacy application might require a shared file system even if its content could theoretically be stored as objects after a redesign.
Do not treat the existing interface as automatically permanent or automatically disposable. Record whether it is a hard constraint, a preference, or an assumption. If changing it is possible, note the development effort, testing, and migration risk that would come with the change.
Latency and throughput answer different questions. Latency is the time an operation takes to complete. Throughput describes how much work is completed over a period of time, such as data transferred or operations handled. A workload may need low latency for individual requests, high throughput for batch processing, or both.
Consider a document application. A user opening one small document may care about response time. A nightly import may care more about completing thousands of documents…