ClimateWatch

The problem

People in Indonesia hear about climate change all the time, but almost nobody can point to their own city's numbers. Is my city actually getting hotter, wetter or drier? The data to answer that exists. It sits in global reanalysis datasets as raw daily rows, which a student, a journalist or a farmer can't read without writing code.

I built ClimateWatch for those readers: the curious public, students, journalists who need quick context, and farmers who care when the rain starts. It turns decades of daily temperature and rainfall into charts you can read in seconds. It's free and needs no login.

What I built

One pinned setting stops a fake warming trend

Open-Meteo's archive quietly switches weather models partway through the record, and I found the switch alone added about 1.5°C to Jakarta's daily highs in a single year, enough to fake a warming trend that isn't in the data. I pin every request to one model family so the whole 1950-to-present record stays comparable. Indonesia has few weather stations with long, unbroken records anyway. That's why this reanalysis dataset was the right fit from the start: gap-free daily data for any point on the map, station or no station, for all 94 cities the same way.

Every chart reads from a precomputed summary

Ninety-four cities with daily data since 1950 add up to millions of rows. I roll them up into monthly and yearly summary tables ahead of time, in scheduled background jobs. A page load reads a small table and never adds up decades of rows on request. The daily data lives in a time-series database, split by date and indexed the way the pages ask for it: by city, then by date.

Missing measurements stay missing

When the source has no value for a day, I store it as empty. I never turn it into zero. Zero millimetres of rain and "we don't know" are different facts, and gaps are common in the older years of the record. A warning badge tells a reader when a city or month's coverage is too thin to trust.

A hot day means something different by region

The standard 35°C threshold for a "hot day" is a temperate-climate convention, dead across most of maritime Indonesia. A quarter of the loaded cities never once crossed it in 77 years, so their hot-day chart just rendered empty. I added a second, region-relative threshold, each city's own hottest 5% of days from its 1951-1980 baseline, and kept the fixed 35°C figure only for comparing across cities.

The same frontend runs two ways, live or static

Open-Meteo has been known to stall or rate-limit requests coming from GitHub Actions, so importing data during a cloud build wasn't reliable, and a backend costs money to keep running continuously. The frontend now switches on one environment variable between calling the live Django API and reading a pre-baked JSON export. I run the import from a normal connection myself and commit the result, so the public deploy needs no server and no live call to Open-Meteo at all.

Result

ClimateWatch is live on GitHub Pages, and the code is public. It covers 94 Indonesian cities with 77 years of daily data, from 1950 to the present. Each city page counts heat days, heavy rain days and the longest heatwave per year with a trend line, and estimates when the wet season started. A "Looking ahead" panel now projects each city's future from several CMIP6 models, beyond the original plan. You can compare two cities, see how 45 tracked El Niño and La Niña episodes changed a city's weather, and rank every city by heat, rain or warming speed.

The live site is a static snapshot: Open-Meteo throttles cloud build servers, so I run the import myself and publish the resulting data files, refreshing the numbers only when I re-run that export. I built it between July and September 2026.