Fly.io
On Fly.io, Timelake runs inside an app on Fly Machines and keeps its tables in a Tigris bucket that fly storage create makes for the app. Unlike a function, a Machine keeps running, so it can collect readings for a few seconds and append them in one commit. One Python file has a role per process group: web takes HTTP requests and Tigris notifications, worker subscribes to MQTT, and jobs runs scheduled work.
| Design | Devices send | Visible to queries after | Fits |
|---|---|---|---|
| HTTP intake, buffered | JSON over HTTPS | the flush interval (10 s) | many small senders; the readings in memory at a crash are lost |
| HTTP intake to landing objects | JSON over HTTPS | the next jobs run (1 minute) | a reading must be stored before the device gets its answer |
| MQTT worker | MQTT messages to your broker | the flush interval (10 s) | devices that already speak MQTT |
| Tigris notifications | files uploaded to Tigris | each upload | gateways that buffer readings and upload files |
Recommendation. For devices that send readings over HTTPS, start with HTTP intake to landing objects: each reading is stored in Tigris before the device gets a response, and one commit a minute keeps requests low. Choose HTTP intake, buffered when 10-second freshness matters more than losing the readings in memory if a Machine crashes. Use MQTT worker for MQTT devices and Tigris notifications for gateways that upload files. The scheduled checks and maintenance run in the jobs process for every design.
Other designs
- Scheduled Machines.
fly machine run --schedulestarts a Machine hourly, daily, weekly, or monthly, counted from when it was created rather than at fixed clock times. That fits a dailymaintainor backfill, not a 5-minute check, which is why the checks here run in the long-runningjobsprocess. For cron syntax, Fly's task scheduling guide covers Supercronic in a process group and its Cron Manager app. - Machines that stop when idle. For low traffic with HTTP intake to landing objects, set
auto_stop_machines = "stop"on thewebgroup: the Machine starts on the next request, andjobskeeps appending. Thejobsandworkergroups have no HTTP service, so they run until you scale them down. - Readers on Fly. A dashboard or API in another Fly app reads the same table from the bucket with the Python package or the TypeScript client, with no load on the writers.
Cost
Each commit sends about 7 writes and 2 reads to Tigris. Tigris charges $0.005 per 1,000 writes and LISTs and $0.0005 per 1,000 reads, and nothing for data transfer out. At one commit every 10 seconds (HTTP intake, buffered and MQTT worker), that is about 30 cents a day; at one a minute, about 5 cents. HTTP intake to landing objects adds a PUT per request, and Tigris notifications a PUT and a GET per file. Storage & configuration breaks the requests down. Fly bills each Machine per second while it runs, by its CPU and memory size.