How do I choose a poll interval?

Start from the product's staleness tolerance, not from the server. If users are fine seeing changes 10 seconds late, poll every 10 seconds. Then check the math: clients divided by interval is your request rate, and if that number is uncomfortable, either lengthen the interval, cache the response, or switch to long polling or SSE. Adaptive intervals that speed up during active windows and slow down after repeated empty responses get you most of the benefit of both settings.

Is short polling ever cheaper than a push connection?

Often, yes. A poll is a stateless request that any CDN, load balancer, or serverless platform handles natively, while push requires held connections, heartbeats, reconnect logic, and servers that drain gracefully on deploys. At low client counts or with bounded freshness windows, the polling requests cost less than the engineering and operations of a connection fleet.

What is the actual difference between short and long polling?

Timing of the response. Short polling answers immediately, even with nothing, so freshness is capped by the interval and most responses are empty. Long polling parks the request until data arrives or a timeout fires, so responses almost always carry data and delivery latency drops to near real time while keeping ordinary HTTP semantics.

Real-Timechapter 1 of 4 · 3 lessons

Short Polling

Client asks 'anything new?' on a fixed timer. Simple but wasteful.

← All chapters View in cheatsheet

Your dashboard asks the server every 5 seconds whether anything changed. Forty support agents have it open.

That is 8 requests a second, and roughly 99 of every 100 come back empty. It is also about fifteen lines of code, it works through every proxy and firewall on earth, and nobody has ever been paged for it.

Now the same design with 100,000 phones instead of forty laptops: 20,000 requests a second, still 99 percent empty, and a fleet of servers whose main job is saying nothing yet.

Same technique, same code, and one of those is obviously right while the other is obviously wrong. Everything else in this course exists because of where that line falls, and plenty of teams have built a socket fleet to replace a timer that was doing fine.

Lessons

3 in this chapter

How It WorksA timer, a fetch, and a freshness bound equal to the interval.2 min
The Cost ModelLoad scales with audience size, never with how much actually happens.2 min
When It's Honestly FineSmall numbers, bounded windows, and staleness tolerance make polling the correct tool.2 min