Communication platform

How 40,000 messages leave the building without getting you banned.

Add integrations, set their throttle windows, and watch the scheduler spread the load. Every job is one minute apart inside a window; when the window closes, the lane goes quiet for the gap and picks up on the next cycle. Flip to no throttling to see what the vendor does when you hand it everything at once.

Total messages to send

Dispatch mode

Integrations

Meta WhatsApp80 TPS cap

batchwindow (min)gap (min)

2,000 ÷ 20 = 100 per job · 20 jobs, one a minute · 10m pause · sustained 67/min

Why lanes can share a scheduler

Every integration above runs its own methods internally — its own auth, endpoints and parsing, whatever the vendor demands. What it returns is not its own invention: each one hands back a defined Pydantic model, and that typed object is processed onward into the single standard format the rest of the application understands.

A voice integration reporting call_completed becomes ANSWERED at the boundary; nothing downstream has ever heard of call_completed. That is what lets one scheduler throttle, batch and roll up status across providers that agree on nothing — the abstraction was determined up front, so the differences never leak.

On the real thing: the scheduler resolves template, tenant and recipient context on every job. Roughly 80% of those reads never reach Postgres — they come out of Redis, with the cache keyed per tenant and invalidated on write. Throttling controls what you send; caching is what makes sending it affordable.

total40,000

lanes1

sustained67/min

completes in9h 50m

delivered100%

T+0m / 9h 50m

dispatched0

in flight this minute100

remaining40,000

Meta WhatsApp40,000 allocated100/job · 20 cycles9h 50m