eachSeries performance regression from 1.2.x to 1.3.x

[goldfire]

I haven't had a chance to look through and figure out what has changed with eachSeries between 1.2.x and 1.3.x, but I have noticed there is a severe performance regression (to the point of it being unusable). It is easily reproducible with the following fiddles:

1.2.1 - http://jsfiddle.net/4vg8p84w/
1.3.2 - http://jsfiddle.net/j3r5az56/

[suguru03]

@goldfire
I'm sorry to reply late.
Your benchmark is very right.
On v1.3.x, series functions (eachSeries, eachLimit, series…) were changed to call async.nextTick at the tail of each iterator when the iterator was called synchronously (nextTick is not called in the whole of the iterator.) to avoid stack overflow even if developer does not care about it.
async v1.x is already changed it.

[xpiwo]

looks like they plan to remove it. caolan/async#883

If the whole purpose of nextTick is just to prevent stack overflows, i think the detection should be several 'sync' in a row.

eg i have waterfall:

[async, async, async, sync_verify_result_is_valid, async, async, async...]

so should use nextTick only after X sync calls.

[suguru03]

I want to remove nextTick too.
But there are some cases that had occurred stack overflow by Async on production environment, I do not want to cause it.

On v2.x, I’m going to enable to select safe or fast. (default is safe)