print() is 68 times faster than ic()

[offchan42]

Try the following code to test speed of print() vs ic():

from icecream import ic
from time import time

start = time()
for i in range(1000):
    print("i:", i)
stop = time()
elapsed = stop - start

start = time()
for i in range(1000):
    ic(i)
stop = time()
elapsed2 = stop - start

print(elapsed)
print(elapsed2)
print(f"print() is {elapsed2 / elapsed} times faster than ic()")

I found that print is 68 times faster!

How to make it faster? Why is it so slow? Is it because of the coloring?

[alexmojaki]

Finding where ic was called from is a very complicated operation, Python doesn't offer it for free. I'm surprised the difference isn't bigger. In fact you call print 100 times but ic 1000 times, which suggests ic is just 6.8 times faster. If you replace print with sys.stdout.write you might get a fairer comparison.

#33 should make repeated calls faster.

[offchan42]

Sorry about the 100 times vs 1000 times. The elapsed time is actually correct, I just put in wrong code.
I've changed it now to 1000 vs 1000. So it's still 68 times faster using print.
Note for future reader: You won't see 100 because I already modified the code to 1000. For you, the code is correct, so skip this comment.

[gruns]

Is ic()'s performance presenting a problem? If so, how and where?

IceCream is a debugging library and thus should not be performance-critical (eg used in a tight loop in production). As @alexmojaki mentioned, ic() does a lot of work to determine output (open source files, parse Python code, etc).

If you want to dive in to improve IceCream's performance, I'd ebulliently welcome and merge such PRs. But performance is not my focus with IceCream until it presents itself as a problem.

As for tight loops and/or production, the forthcoming #33 (also a huge thanks to @alexmojaki there) will help.

[offchan42]

When I was interacting with UI elements and the UI will report some number in the command line. E.g. dragging a slider and then it computes something. The ic() will make the slider look unresponsive if you move the slider passes through many callback points quickly (like from 0 to 100).
But it is fine in that case I change to use print(). It would be better though if I can use ic there.

[gruns]

Aha. That would do it. Great example of an ic() use case where performance matters.

A threaded implementation would solve this. Do all ic() work in a non-UI thread, eg via the threading library. File I/O is requisite for ic(), and there's no easy way to speed that up.

Is this something you could tackle?

[offchan42]

No. I don't know threading in python yet.

[alexmojaki]

Try using the pp function in snoop.

[offchan42]

I see that snoop seems to be more intense in its number of features, providing more options than ic. Though ic is simple and easy to get started.
I tried using snoop and it looks good.
Do you suggest me to try pp function because you think it's faster?

[alexmojaki]

Yes, pp uses executing which is more accurate than ic and has caching which makes it much faster over many calls. #33 is to bring the same functionality to icecream.

Using pp should be the same experience as using ic. The rest of snoop is also great but serves different purposes.

[offchan42]

@alexmojaki Yes. I tried experimenting with this code

from time import time

from icecream import ic
from snoop import pp


def process(print_func, n_iter):
    start = time()
    for i in range(n_iter):
        print_func(i)
    stop = time()
    elapsed = stop - start
    return elapsed


n_iter = 10_000
e1 = process(lambda i: print("i:", i), n_iter)
e2 = process(ic, n_iter)
e3 = process(pp, n_iter)

print("print elapsed:", e1)
print("ic elapsed:", e2)
print("pp elapsed:", e3)
print(f"print() is {e2 / e1} times faster than ic()")
print(f"print() is {e3 / e1} times faster than pp()")

And here is the result:

print elapsed: 0.5972144603729248
ic elapsed: 43.248204708099365
pp elapsed: 3.786238431930542
print() is 72.41653974870843 times faster than ic()
print() is 6.339830468214487 times faster than pp()

So yeah, pp is definitely faster than ic! I don't know why I would not use it. You need a more catchy naming of the library and that's it. Also it doesn't have the issue I mentioned here: #35
It looks to me that it's better in every way, as of now.

It should print within one line though if the print can fit in one line.

[offchan42]

After testing both pp and ic, I've found that pp works better in cmd, but ic is better in jupyter lab.

See that pp does not report the argument name. It only reports the argument name if it's inside a function. And I like the one line style more.

[alexmojaki]

Thanks for pointing this out to me. IPython does some extra magic which makes it harder for executing to know that everything is working OK. If you pip install -U executing you should find that pp works again.

[offchan42]

It seems you know what you doing!

[offchan42]

I will close this issue for now as it seems like I can get around with slowness issue by using pp.
I think improving speed of ic will need some significant core upgrades, so better leave it as is now.