PromptSharp › Prompt Library › Dev & Engineering › Performance regression hunt: from a slower endpoint to the change that caused it

Performance regression hunt: from a slower endpoint to the change that caused it

An endpoint got noticeably slower after last week's releases. Narrow it down to the change and the mechanism before anyone starts optimizing at random.

The prompt — copy and run it

You are a performance engineer investigating a latency regression. I will paste before-and-after timings, the list of changes deployed in the window, and any profiles or traces I have.

Produce:

A) REGRESSION SHAPE — from my numbers: is it every request or a tail, constant or growing with load or data size, one endpoint or several. State what the shape suggests.

B) SUSPECT CHANGES — the deployed changes ranked by how plausibly each could cause this shape, with the specific line or behavior that makes it a suspect.

C) NARROWING PLAN — the cheapest steps to isolate the cause: bisect the deploy window, toggle a flag, compare query plans, profile one request before and after.

D) MECHANISM CHECK — for the top suspect, the measurement that would prove the mechanism (extra query per item, lost cache hit, larger payload, lock contention).

E) GUARD — once fixed, the benchmark or alert that would catch this class of regression next time.

Inputs: [PASTE BEFORE / AFTER TIMINGS WITH PERCENTILES] · [PASTE DEPLOYED CHANGES IN THE WINDOW] · [PASTE PROFILES / TRACES IF ANY] · [TRAFFIC + DATA GROWTH IN THE SAME WINDOW]

Rules: Do not invent timings, profile results, or query plans — reason only from my paste and mark guesses as unconfirmed. Verify the mechanism with a measurement before shipping any optimization. Strip customer data, credentials, and internal identifiers from traces before pasting.

Why this prompt works

Optimizing without isolating the cause often makes code more complex while the regression stays. Reading the shape of the slowdown first, ranking deployed changes by fit, and proving the mechanism with one measurement converts a vague 'it got slower' into a specific culprit. The closing guard means the same class of regression shows up in a benchmark rather than in user complaints.

Frequently asked

Related Dev & Engineering prompts