The Haywards Heath projectKeep asking, or wait to be told?
A year of quarter-hour rainfall, three agents that try to notice it, and what each approach spends.
The Environment Agency publishes rainfall from a gauge it names Haywards Heath, station E9350, open since October 2001. For every quarter of an hour it records how many millimetres fell. In calendar year 2025 that came to 35,040 readings with none missing, though nearly all are marked unchecked by the Agency, so we treat them as provisional. Rain fell in 2,330 of those quarter-hours. The learner groups back-to-back wet quarter-hours into spells and finds 724 of them. Half the spells, 363, lasted a single quarter-hour; the longest ran for 43 in a row, just under eleven hours.
Now imagine an agent whose job is to notice rain, perhaps to warn a sports club or pause a garden watering system. The first version polls: every so many minutes it asks the service for the newest reading and acts if it shows rain. The learner runs that agent over the whole year at different intervals and counts three things. How many requests did it make? How many of those found nothing? And how many wet spells began and ended between two checks, so the agent never saw them at all?
An agent that checks only the newest reading, 2025, our Python run| Checks every | Requests | Found it dry | Spells never seen |
|---|
| 30 minutes | 17,520 | 16,358 | 188 |
| 1 hour | 8,760 | 8,181 | 357 |
| 2 hours | 4,380 | 4,089 | 496 |
| 4 hours | 2,190 | 2,042 | 590 |
| 6 hours | 1,460 | 1,361 | 629 |
Checking hourly, the agent asked 8,760 times, found dry weather 8,181 times, and still missed 357 of the 724 spells, which between them held 100.3 mm of rain. There is a sneakier problem too. Its average delay, measured only over the spells it did catch, comes out at a flattering 14.7 minutes. That number looks good because the misses are left out: short showers are the ones it skips, and they are exactly the ones that would have dragged the average up. An average that silently drops its failures is one of the most useful things a young programmer can learn to distrust.
The second version still polls, but on each check it asks for every reading since the last one instead of just the newest. It never misses a spell. It is simply late: an average of 23.4 minutes behind at hourly checks, 52.9 at two-hourly and 170.9 at six-hourly. The third version is event-driven. Instead of asking, the agent subscribes and is sent a message when a spell starts and another when it stops. Over 2025 that would have been 1,448 messages in total, fewer than a sixth of the hourly poller's requests, and each one arrives as soon as the reading exists.
Ages 8 to 11
A classroom game: one child peeks at a hidden card every so often, another is told when it changes. Who notices first?
Ages 11 to 15
Load the readings in Python, find the wet spells and count what an hourly checker misses.
Ages 15 and up
Write all three agents, compare requests and delays, and explain why one average flatters itself.
Where the numbers come from
Rainfall is from the Environment Agency hydrology service, station E9350, 15-minute totals for 2025, used under the Open Government Licence. The spell definition, the three agents and every figure in the table come from our own Python run. We make no claim about flooding or about any official warning service.
