A cool helper for tracking time and throughput of code blocks, with beautiful human friendly renditions.
Did you ever need to measure the duration of an operation? Yeah, this is easy.
But how to:
- measure the duration of two or more blocks at the same time, including the whole duration?
- instrument a code to cleanly retrieve durations in one line, to log or send to time series databases?
- easily see human friendly durations in s (seconds), ms (milliseconds), Β΅s (microseconds) and even ns (nanoseconds)?
- easily see human friendly counts with SI prefixes like k, M, G, T, etc?
- measure the actual throughput of a block? (this is way harder, since it needs to measure both duration and number of iterations)
- easily see human friendly throughputs in "/second", "/minute", "/hour" or even "/day", including SI prefixes?
Yes, it can get tricky! More interesting details about duration and throughput.
If you'd tried to do it without these magic, it would probably get messy and immensely pollute the code being instrumented.
I have the solution, behold!
from about_time import about_time
def some_func():
import time
time.sleep(85e-3)
return True
def main():
with about_time() as t1: # <-- use it like a context manager!
t2 = about_time(some_func) # <-- use it with any callable!!
t3 = about_time(x * 2 for x in range(56789)) # <-- use it with any iterable or generator!!!
data = [x for x in t3] # then just iterate!
print(f'total: {t1.duration_human}')
print(f' some_func: {t2.duration_human} -> result: {t2.result}')
print(f' generator: {t3.duration_human} -> {t3.count_human} elements, throughput: {t3.throughput_human}')
This main()
function prints:
total: 95.6ms
some_func: 89.7ms -> result: True
generator: 5.79ms -> 56.8k elements, throughput: 9.81M/s
How cool is that? π²π
You can also get the duration in seconds if needed:
In [7]: t1.duration
Out[7]: 0.09556673200064251
But 95.6ms
is way better, isn't it? The same with count
and throughput
!
So, about_time
measures code blocks, both time and throughput, and converts them to beautiful human friendly representations! π
Just install with pip:
β― pip install about-time
There are three modes of operation: context manager, callable and throughput. Let's dive in.
from about_time import about_time
with about_time() as t:
# the code to be measured...
# any lenghty block.
print(f'The whole block took: {t.duration_human}')
This way you can nicely wrap any amount of code.
In this mode, there are the basic fields
duration
andduration_human
.
from about_time import about_time
t = about_time(some_func)
print(f'The whole block took: {t.duration_human}')
print(f'And the result was: {t.result}')
This way you have a nice one liner, and do not need to increase the indent of your code.
In this mode, there is an additional field
result
, with the return of the callable.
If the callable have params, you can use a lambda
or (π new) simply send them:
def add(n, m):
return n + m
t = about_time(add, 1, 41)
# or:
t = about_time(add, n=1, m=41)
# or even:
t = about_time(lambda: add(1, 41))
from about_time import about_time
t = about_time(iterable)
for item in t:
# process item.
print(f'The whole block took: {t.duration_human}')
print(f'It was detected {t.count_human} elements')
print(f'The throughput was: {t.throughput_human}')
This way about_time
also extracts the number of iterations, and with the measured duration it calculates the throughput of the whole loop! It's especially useful with generators, which do not have length.
In this mode, there are the additional fields
count
,count_human
,throughput
andthroughput_human
.
Cool tricks under the hood:
- you can use it even with generator expressions, anything that is iterable to python!
- you can consume it not only in a
for
loop, but also in { list | dict | set } comprehensions,map()
s,filter()
s,sum()
s,max()
s,list()
s, etc, thus any function that expects an iterator! π - the timer only starts when the first element is queried, so you can initialize whatever you need before entering the loop! π
- the
count
/count_human
andthroughput
/throughput_human
fields are updated in real time, so you can use them even inside the loop!
According to the SI standard, there are 1000 bytes in a kilobyte
.
There is another standard called IEC that has 1024 bytes in a kibibyte
, but this is only useful when measuring things that are naturally a power of two, e.g. a stick of RAM.
Be careful to not render IEC quantities with SI scaling, which would be incorrect. But I still support it, if you really want to ;)
By default, this will use SI, 1000
divisor, and no space
between values and scales/units. SI uses prefixes: k
, M
, G
, T
, P
, E
, Z
, and Y
.
These are the optional features:
iec
=> use IEC instead of SI:Ki
,Mi
,Gi
,Ti
,Pi
,Ei
,Zi
,Yi
(implies1024
);1024
=> use1024
divisor β ifiec
is not enabled, use prefixes:K
,M
,G
,T
,P
,E
,Z
, andY
(note the upper 'K');space
=> include a space between values and scales/units everywhere:48 B
instead of48B
,15.6 Β΅s
instead of15.6Β΅s
, and12.4 kB/s
instead of12.4kB/s
.
To change them, just use the properties:
from about_time import FEATURES
FEATURES.feature_1024
FEATURES.feature_iec
FEATURES.feature_space
For example, to enable spaces between scales/units:
from about_time import FEATURES
FEATURES.feature_space = True
I've used just one key concept in designing the human duration features: cleanliness.
3.44s
is more meaningful than3.43584783784s
, and14.1us
is much nicer than.0000141233333s
.
So what I do is: round values to at most two decimal places (three significant digits), and find the best scale unit to represent them, minimizing resulting values smaller than 1
. The search for the best unit considers even the rounding been applied!
0.000999999
does not end up as999.99us
(truncate) nor1000.0us
(bad unit), but is auto-upgraded to the next unit1.0ms
!
The duration_human
units change seamlessly from nanoseconds to hours.
- values smaller than 60 seconds are always rendered as "num.D[D]unit", with one or two decimals;
- from 1 minute onward it changes to "H:MM:SS".
It feels much more humane, humm? ;)
Some examples:
duration (float seconds) | duration_human |
---|---|
.00000000185 | '1.85ns' |
.000000999996 | '1.00Β΅s' |
.00001 | '10.0Β΅s' |
.0000156 | '15.6Β΅s' |
.01 | '10.0ms' |
.0141233333333 | '14.1ms' |
.1099999 | '110ms' |
.1599999 | '160ms' |
.8015 | '802ms' |
3.434999 | '3.43s' |
59.999 | '0:01:00' |
68.5 | '0:01:08' |
125.825 | '0:02:05' |
4488.395 | '1:14:48' |
I've made the throughput_human
with a similar logic. It is funny how much trickier "throughput" is to the human brain!
If something took
1165263 seconds
to handle123 items
, how fast did it go? It's not obvious...
It doesn't help even if we divide the duration by the number of items, 9473 seconds/item
, which still does not mean much. How fast was that? We can't say.
How many items did we do per time unit?
Oh, we just need to invert it, so
0,000105555569858 items/second
, there it is! π
To make some sense of it we need to multiply that by 3600 (seconds in an hour) to get 0.38/h
, which is much better, and again by 24 (hours in a day) to finally get 9.12/d
!! Now we know how fast that process was! \o/ As you see, it's not easy at all.
The throughput_human
unit changes seamlessly from per-second, per-minute, per-hour, and per-day.
It also automatically inserts SI-prefixes, like k, M, and G. π
duration (float seconds) | number of elements | throughput_human |
---|---|---|
1. | 10 | '10.0/s' |
1. | 2500 | '2.50k/s' |
1. | 1825000 | '1.82M/s' |
2. | 1 | '30.0/m' |
2. | 10 | '5.00/s' |
1.981981981981982 | 11 | '5.55/s' |
100. | 10 | '6.00/m' |
1600. | 3 | '6.75/h' |
.99 | 1 | '1.01/s' |
1165263. | 123 | '9.12/d' |
about_time
supports all versions of python, but in pythons >= 3.3
it performs even better, with much higher resolution and smaller propagation of errors, thanks to the new time.perf_counter
. In older versions, it uses time.time
as usual.
- 4.2.1: makes fixed precision actually gain more resolution, when going from a default 1 to 2 decimals
- 4.2.0: support for fixed precision, useful when one needs output without varying lengths; official Python 3.11 support
- 4.1.0: enable to cache features within closures, to improve performance for https://github.com/rsalmei/alive-progress
- 4.0.0: new version, modeled after my Rust implementation in https://crates.io/crates/human-repr; includes new global features, new objects for each operation, and especially, new simpler human friendly representations; supports Python 3.7+
- 3.3.0: new interfaces for count_human and throughput_human; support more common Kbyte for base 2 (1024), leaving IEC one as an alternate
- 3.2.2: support IEC kibibyte standard for base 2 (1024)
- 3.2.1: support divisor in throughput_human
- 3.2.0: both durations and throughputs now use 3 significant digits; throughputs now include SI-prefixes
- 3.1.1: make
duration_human()
andthroughput_human()
available for external use - 3.1.0: include support for parameters in callable mode; official support for python 3.8, 3.9 and 3.10
- 3.0.0: greatly improved the counter/throughput mode, with a single argument and working in real time
- 2.0.0: feature complete, addition of callable and throughput modes
- 1.0.0: first public release, context manager mode
This software is licensed under the MIT License. See the LICENSE file in the top distribution directory for the full license text.
Maintaining an open source project is hard and time-consuming, and I've put much β€οΈ and effort into this.
If you've appreciated my work, you can back me up with a donation! Thank you π