Workload

A workload is the description of the demand a system must handle: who uses it, how often, how many at once, with what data, and when. It is the condition under which quality requirements such as response time are measured. A performance figure reported without its workload cannot be compared with another figure, because the two may have been produced under different demands.

What a workload description contains

Users and their roles; how many are active at once; the request rate and how it varies across the day, the month, and short bursts; the volume of records and its growth; the size of documents or requests; how often data is updated; whether caches are cold or warm; and the period observed. Separate measured current demand from forecast demand, and say which one a test used.

For an invoice-evidence lookup, an illustrative workload is 20 clerks, roughly 0.67 requests per second on an ordinary day and 2 per second at month end, 400,000 receipt records growing by about 8,000 a month, documents up to 20 MB, and a cold cache after each nightly refresh. A vendor test of one request per second on 1,000 small cached documents is a different workload, and its result does not transfer.

Averages hide the moments that matter

Google’s site-reliability guidance points out that even a simple measure such as requests per second aggregates over a window and can hide much higher instantaneous rates in bursts lasting a few seconds. State the aggregation interval alongside the rate, and include the peak period the system must survive, not only the typical hour.

A workload is a description of demand, not of capacity. Whether a system can handle it is the question the test answers. Workload isolation is a related but different idea: it concerns how one workload’s demand affects another sharing the same resources.

Figures here are illustrative, not benchmarks.

References:

• Google SRE: Service level objectives