2.49 Million Requests in Its First Week: How BKK FloodWatch Survived the Launch Surge

2.49 Million Requests in Its First Week: How BKK FloodWatch Survived the Launch Surge

Between September 26 and October 3, 2026—FloodWatch's first week online—Bangkok was still under the 2026 flood, its canals slowly falling from their peak.

Many canals were still over their banks—Khlong Lat Phrao, near Bang Khen, among them. Everyday citizens, motorcycle delivery drivers, parents picking up children from school, and office workers desperately needed to know one simple thing:

"Is my street flooded right now, and will the water reach the road before I get home?"

Commuters turned en masse to BKK FloodWatch (https://flood.autobahn.bot).

During that first week, Cloudflare's dashboard recorded extraordinary numbers: 2.49 Million Total Requests (every page, map tile, script and auto-refresh—not people) 105,800 Visits (visits, not unique people) 34.29 GB Data Transferred Peak Traffic: About 150,000 requests in the busiest hour, around September 30 5xx Error Rate: 1.74%, mostly in two spikes around that peak—when people needed it most Total Cost to Citizens: $0 (No paywalls, no tracking cookies, and zero advertisements)

Here is the engineering breakdown of how a small civic public tool survived a 2.49 million request week without collapsing.

The Trap of Upstream Dependency

That week, BKK FloodWatch aggregated telemetry from about 1,040 gauges on canals and rivers, with views of 62 rivers, from 5 public sources (HII, RID, EGAT, BMA and the Thai Red Cross volunteer network).

Government telemetry feeds are invaluable, but their upstream origin servers were never designed to withstand a whole city refreshing on their phones. If our frontend had made direct client calls to public agency endpoints, two catastrophic failures would have occurred: 1. The public agency servers would have crashed within three minutes. 2. Our users standing in the rain with weak 4G signals would have stared at loading spinners for 45 seconds.

The Defensive Architecture: A Shield Between Phones and Agencies

To protect both our users and the upstream agencies, we put our own server between them:

1. 1Background Polling Daemon: A dedicated background worker polls the 5 public feeds on a strict 5-to-10 minute cadence, normalizes the disparate JSON formats, and validates sensor telemetry.
2. 2Automatic Anomaly Quarantine: In severe storms, water gauges occasionally jump erratically due to electrical shorts or debris. When a sensor exhibits impossible velocity, our pipeline retains its geographic map dot but hides the suspect numeric reading behind an honest note. We never broadcast panic based on faulty sensors.
3. 3Served Through Cloudflare: Every request passes through Cloudflare's network, but in that first week only 11.41% were answered from its cache; the rest reached our single server.

Even if an upstream agency API timed out, the server kept showing the last readings it had collected, with their age.

Designing Calm Tech Under Pressure

When people are anxious about rising floodwaters, sensationalist UI makes the panic worse. We strictly adhered to our Calm Tech design guidelines: No Screaming Red Banners: A gauge with room to spare is a quiet blue; the colour turns yellow, orange and red only as the water nears and crosses the bank—always with the words (over the bank, near the bank, watch, room to spare) and the direction. Tested Ranges, Never False Countdowns: Nature is not a countdown clock. We never show false precision like "Water arriving in 14 minutes." Instead, our backtested models give ranges 24, 48 and 72 hours ahead, each tested on a year of history with the rain for its area and the gauges upstream. Ultralight 4G Optimization:* The mobile interface is built light for a weak signal, so commuters under highway overpasses can still get clear answers.

Public Technology as Civic Duty

Building tools like BKK FloodWatch reminds us why open-source software matters. Technology should not only exist to sell subscriptions; it must protect communities when the rain starts falling.

The first edition of FloodWatch (up to v0.34) is open-source under the MIT license, and our public JSON API remains free and key-less for researchers and developers at https://flood.autobahn.bot/api/docs.

To everyone behind those 105,800 visits who relied on our gauges during the flood: stay dry, stay safe, and know that our team is watching the waters with you.