Slopegraph · Diagram Design

One service went the other way

p95 response time per service before and after the caching layer, both axes on one 100-to-550ms scale. Four services got faster. The recommender got 81% slower, and being the only line going up it crosses three of them.

One service went the other way Slopegraph of p95 response time by service before and after a caching layer, on one shared 100-to-550-millisecond scale; four services got faster and the recommender got slower, crossing three of them. P95 RESPONSE TIME, MS BEFORE AFTER Search 512 208 Search Catalog 376 164 Catalog Checkout 291 143 Checkout Auth 154 121 Auth Recommender 238 431 Recommender LEGEND P95 RESPONSE TIME IN MS, LOWER IS BETTER · SAME 100-550 SCALE ON BOTH AXES · ILLUSTRATIVE FIGURES, NOT A MEASURED RELEASE Recommender · focal, the only regression Other services · strongest tone was slowest before

FOCAL · RECOMMENDER

The recommender went from fourth-slowest to slowest

238ms to 431ms. It is the only line that rises, so it crosses three of the other four — all but Auth, which stays below it throughout. Where it crosses on the page is an artifact of joining two endpoints with a straight line: this chart cannot say when the regression started, only that it is there.

Search won the most and is still the second slowest

512ms to 208ms, down 304 — the steepest slope here. It started so far above everything else that the largest absolute win still leaves it near the top. Level and change share one scale, which is what lets you read both at once.

Auth barely moved because it had nothing to win

154ms to 121ms, down 33. Nearly flat, and low on the axis — a shallow slope down here means a service that was already fast, while the same shallow slope near the top would mean a cache that did not work.

p95 response time · before and after the caching layer · illustrative example · diagram design