Compute

Compute means the resources and execution capability used to perform work on data. In a platform diagram, it includes tasks such as filtering records, joining tables, training models, and calculating predictions. CPUs or accelerators execute instructions, while memory holds the working state.

Follow the work, not just the machine

A daily report might read order files, group them by store, and write totals. Storage retains the files and totals. Compute performs the grouping. Serving lets a reader query the accepted totals. These roles can share a machine or be provided by different services.

AWS's compute overview describes several ways to provide execution resources. A virtual machine, a container service, and a managed job service expose different operating responsibilities; none removes the need to understand the workload.

More capacity helps only the work it can accelerate

Suppose a request spends 80 ms waiting for a feature service and 20 ms calculating a score. Halving only the calculation makes this sequential path 90 ms instead of 100 ms. It does not make the feature service faster. These are illustrative values with network work outside the stated path excluded.

Processing speed, memory limits, input-output waits, and queues can each constrain a task. Measure where the time goes before adding resources. More compute also does not repair missing input records or an incorrect sales definition.

Reference: AWS: Compute.