Why not to use setInterval

How about making an async version of setInterval? Something like...

const setIntervalAsync = (fn, ms) => {
  fn().then(() => {
    setTimeout(() => setIntervalAsync(fn, ms), ms);
  });
};

So we can just do

setIntervalAsync(() => fetch(/* blah */), 3000);

Which would call fetch every 3000ms properly.

P.S. Haven't really tested it, just an idea...

Another approach is to check that the previous process isn't still running inside the setInterval callback and choose whether to skip that iteration or to kill and restart the timer.

Better would be to trigger actions on user events rather than timed events if that's possible, or to use timers only for pure cosmetics. I don't think setInterval/setTimeout are evil as such, as long as you're careful since suddenly everything's in global scope.

There's a temptation to use them to get out of a fix, though, and I think you're definitely right to treat them as a bit of a smell.

There are perfectly sane ways of using setInterval without requests backing up, e.g. (warning: untested RxJS code ahead):

interval(10000)
    .switchMap(() => Observable.create(observer => {
        const request = superagent.get(
            'https://example.com/api/v1/foo',
            (err, res) => {
                if (!err) {
                    observer.next(res.body)
                }

                observer.complete()
            }
        )

        return () => request.abort()
    }))
    .subscribe(console.log)

As you can see from printed console.log statement that whole load of setInterval events arrive immediately after each other without any delay.

Didn't get a point, as I see from printed console, results are fully predicted. Every single "insideSetInterval" prints with delay of two seconds (like described in setInterval), and also - "returning from server" appears after 4 seconds since first setInterval, which was called with timeout 4000ms, where is "without any delay"?

If you speak about no delay between "returning from server" and "insideSetInterval", of course, because setTimeout with 4sec will wait until interval will be executed twice), and dont have to be in sync with interval prints. With heavy calculation it can load stack, but for simple refreshes - like a timer display update or some other light calculations - working properly, if not forget to handle instance with "clearInterval" when it's no needed anymore.

Another reason is because setIntervals stop running if your browser window loses focus, your callbacks will queue up and run once the page has focus again. Or even worse, they will be dropped. See jsfiddle.net/mendesjuan/y3h1n6mk/

In order to understand why setInterval is evil

Thanks for the detailed explanation of setInterval in sync and async contexts, but why is it this "evil"? With a proper understanding of how the js event loop works, it's easy to understand the reason why a "new rhythm" gets established in the sync case, and why it doesn't in the async case.

You don't need to create the variable, but it's a good practice as you can use that variable with clearInterval to stop the currently running interval.

var int = setInterval("doSomething()", 5000 ); /* 5 seconds /
var int = setInterval(doSomething, 5000 ); / same thing, no quotes, no parens */

If you need to pass parameters to the doSomething function, you can pass them as additional parameters beyond the first two to setInterval.

Without overlapping

setInterval, as above, will run every 5 seconds (or whatever you set it to) no matter what. Even if the function doSomething() takes long than 5 seconds to run. That can create issues. If you just want to make sure there is that pause in between runnings of doSomething, you can do this:
(function(){
doSomething();
setTimeout(arguments.callee, 5000);
})()

I get what you are trying to say here. A small mistake here though. The timeout in fakeCallToServer should have been a random from 200ms to 15 sec to mimic an actual api response scenario.