Compile performance metrics on direct vs. queued execution about proposal-observable

[zenparsing]

Here's some performance data I've gathered, based on the script located under "perf/subscription.js", executed on my rather old 2009 macbook:

[Microtask subscription]
Subscriptions: 100000
Sequence Size: 10
Deliveries: 1000000
Time: 8423ms

[Synchronous subscription]
Subscriptions: 100000
Sequence Size: 10
Deliveries: 1000000
Time: 2774ms

Delta: 5649ms
Delta/subscription: 0.05649ms
Delta/delivery: 0.005649ms

Obviously it's slower to go through the microtask queue than to perform a direct function call, but it's hard for me to tell what to make of this.

Since we pay the queuing cost only once per sequence, the averaged cost per delivery is inversely proportional to the size of the sequence. Whether we view the cost as significant or not seems to depend on what we choose to be a typical sequence size.

@jhusain what do you think?

[jhusain]

Here we pay the queuing cost only once per subscription because there is one observable. All it takes is one flatMap on this observable that retrieves data from a data store with a cache, and now we have a schedule per notification.

[domenic]

That's a common misconception, but not true. Once you're in the microtask queue, you've already paid the cost, and it won't be paid again. In other words the microtask queue is implemented as

while (queue.length > 0) {
  task = queue.pop();
  task();
}

If task throws an exception, you might pay a hit, as you have to back out and go back through the queue again. But the cost for a single microtask subscription is the same as for multiple---namely, it's the cost of letting the rest of the synchronous code that's planning to run this turn, run.

[jhusain]

The cost I'm referring to here is the cost of unwinding the stack and winding it up again vs. just building up the stack. If the overhead here is indeed small, I'm open to asynchrony all the time. The acid test would be whether the scheduling was visible to the naked eye in a mouse drag or other gesture for example. If the scheduling introduces any latency it would be a serious problem as event composition is one of the key scenarios. Should be easy to test.

[zenparsing]

We can't really do asynchronous delivery (i.e. per-iteration) because the data flow is two-way. Data for the iteration flows from the observable to the observer, and a completion value flows back down to the observable. That may be a thrown exception ("throw") or it may be { done: true} ("return"), or it may just carry some value (a backpressure signal, for instance). But we need to maintain the current stack in order to receive that data from the observer.