I disagree. A lot of important and large codebases were grown and maintained in Python (Instagram, Dropbox, OpenAI) and it's damn useful to know how to reason your way out of a Python performance problem when you inevitably hit one without dropping out into another language, which is going to be far more complex.
Python is a very useful tool, and knowing these numbers just makes you better at using the tool. The author is a Python Software Foundation Fellow. They're great at using the tool.
In the common case, a performance problem in Python is not the result of hitting the limit of the language but the result of sloppy un-performant code, for example unnecessarily calling a function O(10_000) times in a hot loop.
I wrote up a more focused "Python latency numbers you should know" as a quiz here https://thundergolfer.com/computers-are-fast
Good designs do not happen in a vacuum but informed with knowledge of at least the outlines of the environment.
One can have a breakfast pursuing an idea -- let me spill some sticky milk on the dining table, who cares, I will clean up if it becomes a problem later.
Another is, it's not much of an overbearing constraint not to make a mess with spilt milk in the first place, maybe it will not be a big bother later, but it's not hurting me much now, to be not be sloppy, so let me be a little hygienic.
There's a balance between making a mess and cleaning up and not making a mess in the first place. The other extreme is to be so defensive about the possibility of creating a mess that it paralyses progress.
The sweet spot is somewhere between the extremes and having the ball-park numbers in the back of one's mind helps with that. It informs about the environment.
Python’s issue is that it is incredibly slow in use cases that surprise average developers. It is incredibly slow at very basic stuff, like calling a function or accessing a dictionary.
If Python didn’t have such an enormous number of popular C and C++ based libraries it would not be here. It was saved by Numpy etc etc.
I think these kind of numbers are everywhere and not just specific to Python.
In zig, I sometimes take a brief look to the amount of cpu cycles of various operations to avoid the amount of cache misses. While I need to aware of the alignment and the size of the data type to debloat a data structure. If their logic applies, too bad, I should quit programming since all languages have their own latency on certain operations we should aware of.
There are reasons to not use Python, but that particular reason is not the one.
I think the list itself is super long winded and not very informative. A lot of operations take about the same amount of time. Does it matter that adding two ints is very slightly slower than adding two floats? (If you even believe this is true, which I don’t.) No. A better summary would say “all of these things take about the same amount of time: simple math, function calls, etc. these things are much slower: IO.” And in that form the summary is pretty obvious.
I agree. I have to complement the author for the effort put in. However it misses the point of the original Latency numbers every programmer should know, which is to build an intuition for making good ballpark estimations of the latency of operations and that e.g. A is two orders of magnitude more expensive than B.
For me, it will help with selecting what language is best for a task. I think it won't change my view that python is an excellent language to prototype in though.
O(10_000) is a really weird notation.
How does this happen? Is it just inertia that cause people to write large systems in a essentially type free, interpreted scripting language?
If I told you that we were going to be running a very large payments system, with customers from startups to Amazon, you'd not write it in ruby and put the data in MongoDB, and then using its oplog as a queue... but that's what Stripe looked like. They even hired a compiler team to add type checking to the language, as that made far more sense than porting a giant monorepo to something else.
10 years later "ok it's too slow; our options are a) spend $10m more on servers, b) spend $5m writing a faster Python runtime before giving up later because nobody uses it, c) spend 2 years rewriting it and probably failing, during which time we can make no new features. a) it is then."
When this starts to matter, python stops being the right tool for the job.
But I agree with the spirit of what you wrote - these numbers are interesting but aren’t worth memorizing. Instead, instrument your code in production to see where it’s slow in the real world with real user data (premature optimization is the root of all evil etc), profile your code (with pyspy, it’s the best tool for this if you’re looking for cpu-hogging code), and if you find yourself worrying about how long it takes to add something to a list in Python you really shouldn’t be doing that operation in Python at all.
The interoperability between C and Python makes it great, and you need to know these numbers on Python to know when to actually build something in C. With Zig getting really great interoperability, things are looking better than ever.
Not that you're wrong as such. I wouldn't use Python to run an airplane, but I really don't see why you wouldn't care about the resources just because you're working with an interpreted or GC language.
People usually approach this the other way, use something like pandas or numpy from the beginning if it solves your problem. Do not write matrix multiplications or joins in python at all.
If there is no library that solves your problem, it's a great indication that you should avoid python. Unless you are willing to spend 5 man-years writing a C or C++ library with good python interop.
time python -I -c 'print("Hello World")'
real 0m0.014s
time bash --noprofile -c 'echo "Hello World"'
real 0m0.001sSome of those number are very important:
- Set membership check is 19.0 ns, list is 3.85 μs. Knowing what data structure to use for the job is paramount.
- Write 1KB file is 35.1 μs but 1MB file is only 207 μs. Knowing the implications of I/O trade off is essential.
- sum() 1,000 integers is only 1,900 ns: Knowing to leverage the stdlib makes all the difference compared to manual loop.
Etc.
A few years ago I did a Python rewrite of a big clients code base. They had a massive calculation process that took 6 servers 2 hours.
We got it down to 1 server, 10 minutes, and it was not even the goal of the mission, just the side effect of using Python correctly.
In the end, quadratic behavior is quadratic behavior.
Your cognition of it is either implicit or explicit.
Even if you didn't know for example that list appends was linear and not quadratic and fairly fast.
Even if you didn't give a shit if simple programs were for some reason 10000x slower than they needed to be because it meets some baseline level of good enough / and or you aren't the one impacted by the problems inefficacy creates.
Library authors beneath you would still know and the APIs you interact with and the pythonic code you see and the code LLMS generate will be affected by that leaky abstraction.
If you think that n^2 naive list appends is a bad example its not btw, python string appends are n^2 and that has and does affect how people do things, f strings for example are lazy.
Similarly a direct consequence of dictionaries being fast in Python is that they are used literally everywhere. The old Pycon 2017 talks from Raymond talk about this.
Ultimately what the author of the blog has provided is this sort of numerical justification for the implicit tacit sort of knowledge performance understanding gives.
Relevant if your problem demands instatiation of a large number of objects. This reminds me of a post where Eric Raymond discusses the problems he faced while trying to use Reposurgeon to migrate GCC. See http://esr.ibiblio.org/?p=8161
From what I've been able to glean, it was basically created in the first few years Jeff worked at Google, on indexing and serving for the original search engine. For example, the comparison of cache, RAM, and disk: determined whether data was stored in RAM (the index, used for retrieval) or disk (the documents, typically not used in retrieval, but used in scoring). Similarly, the comparison of California-Netherlands time- I believe Google's first international data cetner was in NL and they needed to make decisions about copying over the entire index in bulk versus serving backend queries in the US with frontends in the NL.
The numbers were always going out of date; for example, the arrival of flash drives changed disk latency significantly. I remember Jeff came to me one day and said he'd invented a compression algorithm for genomic data "so it can be served from flash" (he thought it would be wasteful to use precious flash space on uncompressed genomic data).
I completely understand why it's frustrating or confusing by itself, though.
This page is a nice reminder of the fact, with numbers. For a while, at least, I will Know, instead of just feel, like I can ignore the low level performance minutiae.
> Strings
>The rule of thumb for strings is the core string object takes 41 bytes. Each additional character is 1 byte.
Collection Access and Iteration
How fast can you get data out of Python’s built-in collections? Here is a dramatic example of how much faster the correct data structure is. item in set or item in dict is 200x faster than item in list for just 1,000 items!
Also the overall title “Python Numbers Every Programmer Should Know” starts with 20 numbers that are merely interesting.
That all said, the formatting is nice and engaging.
I remember refactoring some code to improve readability, then observing something that was previously a few microseconds take tens of seconds.
The original code created a large list of lists. Each child list had 4 fields each field was a different thing, some were ints and one was a string.
I created a new class with the names of each field and helper methods to process the data. The new code created a list of instances of my class. Downstream consumers of the list could look at the class to see what data they were getting. Modern Python developers would use a data class for this.
The new code was very slow. I’d love it if the author measured the time taken to instantiate a class.
Please post your code snippet on StackOverflow ([python] tag) or CodeReview.SE so people can help you fix it.
> created a new class with the names of each field and helper methods to process the data. The new code created a list of instances of my class. Downstream consumers of the list could look at the class to see what data they were getting.
The doctor said, “don’t do that”.
Edit: so yeah a rather snarky reply. Sorry. But it’s worth asking why we want to use classes and objects everywhere. Alan Kay is well known for saying object orientated is about message passing (mostly by Erlang people).
A list of lists (where each list is four different types repeated) seems a fine data structure, which can be operated on by external functions, and serialised pretty easily. Turning it into classes and objects might not be a useful refactoring, I would certainly want to learn more before giving the go ahead.
When you’re in a project with a few million lines of code and 10 years of history it can get confusing.
Your data will have been handled by many different functions before it gets to you. If you do this with raw lists then the code gets very confusing. In one data structure customer name might be [4] and another structure might have it in [9]. Worse someone adds a new field in [5] then when two lists get concatenated name moves to [10] in downstream code which consumes the concatenated lists.
customers[3][4]
is a lot less readable than
customers[3].balance
It's -5 to 256, and these have very tricky behavior for programmers that confuse identity and equality.
>>> a = -5
>>> b = -5
>>> a is b
True
>>> a = -6
>>> b = -6
>>> a is b
FalseAfter skimming over all of them, it seems like most "simple" operations take on the order of 20ns. I will leave with that rule of thumb in mind.
I usually prefer classic %-formatting for readability when the arguments are longer and f-strings when the arguments are shorter. Knowing there is a material performance difference at scale, might shift the balance in favour of f-strings for some situations.
Last time I benchmarked a VPS it was about the performance of an Ivy Bridge generation laptop.
I have a number of Intel N95 systems around the house for various things. I've found them to be a pretty accurate analog for small instances VPSes. The N95 are Intel E-cores which are effectively Sandy Bridge/Ivy Bridge cores.
Stuff can fly on my MacBook but than drag on a small VPS instance but validating against an N95 (I already have) is helpful. YMMV.
Thanks for the feedback everyone. I appreciate your posting it @woodenchair and @aurornis for pointing out the intent of the article.
The idea of the article is NOT to suggest you should shave 0.5ns off by choosing some dramatically different algorithm or that you really need to optimize the heck out of everything.
In fact, I think a lot of what the numbers show is that over thinking the optimizations often isn't worth it (e.g. caching len(coll) into a variable rather than calling it over and over is less useful that it might seem conceptually).
Just write clean Python code. So much of it is way faster than you might have thought.
My goal was only to create a reference to what various operations cost to have a mental model.
For example, from the post "Maybe we don’t have to optimize it out of the test condition on a while loop looping 100 times after all."
Though IMHO it suffices just to know that "Python is 40-50x slower than C and is bad at using multiple CPUs" is not just some sort of anti-Python propaganda from haters, but a fairly reasonable engineering estimate. If you know that you don't really need that chart. If your task can tolerate that sort of performance, you're fine; if not, figure out early how you are going to solve that problem, be it through the several ways of binding faster code to Python, using PyPy, or by not using Python in the first place, whatever is appropriate for your use case.
They're both about a full second on my old laptop.
I ended up writing my own simple LLM library just so I wouldn't have to import OpenAI anymore for my interactive scripts.
(It's just some wrapper functions around the equivalent of a curl request, which is honestly basically everything I used the OpenAI library for anyway.)
Additionally, regardless of the code you can profile the system to determine where the "hot spots" are and refactor or call-out to more performant (Rust, Go, C) run-times for those workflows where necessary.
More contentiously: don't fret too much over performance in Python. It's a slow language (except for some external libraries, but that's not the point of the OP).
Surely the 100-char string information of 141 bytes is not correct as it would only apply to ASCII 100-char strings.
It would be more useful to know the overhead for unicode strings presumably utf-8 encoded. And again I would presume 100-Emoji string would take 441 bytes (just a hypothesis) and 100-umlaut chars string would take 241bytes.
Care about orders of magnitude instead, in combination with the speed of hardware https://gist.github.com/jboner/2841832 you'll have a good understanding of how much overhead is due to the language and the constructs to favor for speed improvements.
Just reading the code should give you a sense of its speed and where it will spend most time. Combined with general timing metrics you can also have a sense of the overhead of 3rd party libraries (pydantic I'm looking at you).
So yeah, I find that list quite useful during the code design, likely reduce time profiling slow code in prod.
Firstly, I want to start with the fact that the base system is a macOS/M4Pro, hence;
- Memory related access is possibly much faster than a x86 server. - Disk access is possibly much slower than a x86 server.
*) I took x86 server as the basis as most of the applications run on x86 Linux boxes nowadays, although a good amount of fingerprint is also on other ARM CPUs.
Although it probably does not change the memory footprint much, the libraries loaded and their architecture (ie. being Rosetta or not) will change the overall footprint of the process.
As it was mentioned on one of the sibling comments -> Always inspect/trace your own workflow/performance before making assumptions. It all depends on specific use-cases for higher-level performance optimizations.
The list of floats is larger, despite also being simply an array of 1000 8-byte pointers. I assume that it's because the int array is constructed from a range(), which has a __len__(), and therefore the list is allocated to exactly the required size; but the float array is constructed from a generator expression and is presumably dynamically grown as the generator runs and has a bit of free space at the end.
For example, my M4 Max running Python 3.14.2 from Homebrew (built, not poured) takes 19.73MB of RAM to launch the REPL (running `python3` at a prompt).
The same Python version launched on the same system with a single invocation for `time.sleep()`[1] takes 11.70MB.
My Intel Mac running Python 3.14.2 from Homebrew (poured) takes 37.22MB of RAM to launch the REPL and 9.48MB for `time.sleep`.
My number for "how much memory it's using" comes from running `ps auxw | grep python`, taking the value of the resident set size (RSS column), and dividing by 1,024.
1: python3 -c 'from time import sleep; sleep(100)'
I guess you could find yourself in a situation where a 2X speedup is make or break and you're not a week away from needing 4X, etc. But not very often.
[1] https://hex.pm/packages/snakepit [2] https://hex.pm/packages/snakebridge
String operations in Python are fast as well. f-strings are the fastest formatting style, while even the slowest style is still measured in just nano-seconds.
Concatenation (+) 39.1 ns (25.6M ops/sec)
f-string 64.9 ns (15.4M ops/sec)
"literal1 " + str(expression) + " literal2"
f"literal1 {expression} literal2"
If I have only plain Python installed and a .py file that I want to test, then what's the easiest way to get a visualization of the call tree (or something similar) and the computational cost of each item?
Benchmark Iteration Process
Core Approach:
- Warmup Phase: 100 iterations to prepare the operation (default)
- Timing Runs: 5 repeated runs (default), each executing the operation a specified number of times
- Result: Median time per operation across the 5 runs
Iteration Counts by Operation Speed: - Very fast ops (arithmetic): 100,000 iterations per run
- Fast ops (dict/list access): 10,000 iterations per run
- Medium ops (list membership): 1,000 iterations per run
- Slower ops (database, file I/O): 1,000-5,000 iterations per run
Quality Controls:
- Garbage collection is disabled during timing to prevent interference
- Warmup runs prevent cold-start bias
- Median of 5 runs reduces noise from outliers
- Results are captured to prevent compiler optimization elimination
Total Executions: For a typical benchmark with 1,000 iterations and 5 repeats, each operation runs 5,100 times (100 warmup + 5×1,000 timed) before reporting the median result.
Nobody is going to remember any of the numbers on this new list.
To use a trivial example, using a set instead of a list to check membership is a very basic replacement, and can dramatically improve your running time in Python. Just because you use Python doesn't mean anything goes regarding performance.
Looking at performance numbers is important regardless if it's python, assembly or HDL. If you don't understand why your code is slow you can always look at how many cycles things take and learn to understand how code works at a deeper level, as you mature as a programmer things will become obvious, but going through the learning process and having references like these will help you to get there sooner, seeing the performance numbers and asking why some things take much longer—or sometimes why they take the exact same time—is the perfect opportunity to learn.
Early in my python career I had a python script that found duplicate files across my disks, the first iteration of the script was extremely slow, optimizing the script went through several iterations as I learned how to optimize at various levels. None of them required me to use C. I just used caching, learned to enumerate all files on disk fast, and used sets instead of lists. The end result was that doing subsequent runs made my script run in 10 seconds instead of 15 minutes. Maybe implementing in C would make it run in 1 second, but if I had just assumed my script was slow because of python then I would've spent hours doing it in C only to go from 15 minutes to 14 minutes and 51 seconds.
There's an argument to be made that it would be useful to see C numbers next to the python ones, but for the same reason people don't just tell you to just use an FPGA instead of using C, it's also rude to say python is the wrong tool when often it isn't.
Makes me wonder if the cpython devs have ever considered v8-like NaN-boxing or pointer stuffing.
Then again, if you're worried about any of the numbers in this article maybe you shouldn't be using Python at all. I joke, but please do at least use Numba or Numpy so you aren't paying huge overheads for making an object of every little datum.
note that protobuf attributes are 20-50x worse than this
- If slotted attribute reads and regular attribute reads are the same latency, I suspect that either the regular class may not have enough "bells on" (inheritance/metaprogramming/dunder overriding/etc) to defeat simple optimizations that cache away attribute access, thus making it equivalent in speed to slotted classes. I know that over time slotting will become less of a performance boost, but--and this is just my intuition and I may well be wrong--I don't get the impression that we're there yet.
- Similarly "read from @property" seems suspiciously fast to me. Even with descriptor-protocol awareness in the class lookup cache, the overhead of calling a method seems surprisingly similar to the overhead of accessing a field. That might be explained away by the fact that property descriptors' "get" methods are guaranteed to be the simplest and easiest to optimize of all call forms (bound method, guaranteed to never be any parameters), and so the overhead of setting up the stack/frame/args may be substantially minimized...but that would only be true if the property's method body was "return 1" or something very fast. The properties tested for these benchmarks, though, are looking up other fields on the class, so I'd expect them to be a lot slower than field access, not just a little slower (https://github.com/mikeckennedy/python-numbers-everyone-shou...).
- On the topic of "access fields of objects" (properties/dataclasses/slots/MRO/etc.), benchmarks are really hard to interpret--not just these benchmarks, all of them I've seen. That's because there are fundamentally two operations involved: resolving a field to something that produces data for it, and then accessing the data. For example, a @property is in a class's method cache, so resolving "instance.propname" is done at the speed of the methcache. That might be faster than accessing "instance.attribute" (a field, not a @property or other descriptor), depending on the inheritance geometry in play, slots, __getattr[ibute]__ overrides, and so on. On the other hand, accessing the data at "instance.propname" is going to be a lot more expensive for most @properties (because they need to call a function, use an argument stack, and usually perform other attribute lookups/call other functions/manipulate locals, etc); accessing data at "instance.attribute" is going to be fast and constant-time--one or two pointer-chases away at most.
- Nitty: why's pickling under file I/O? Those benchmarks aren't timing pickle functions that perform IO, they're benchmarking the ser/de functionality and thus should be grouped with json/pydantic/friends above.
- Asyncio's no spring chicken, but I think a lot of the benchmarks listed tell a worse story than necessary, because they don't distinguish between coroutines, Tasks, and Futures. Coroutines are cheap to have and call, but Tasks and Futures have a little more overhead when they're used (even fast CFutures) and a lot more overhead to construct since they need a lot more data resources than just a generator function (which is kinda what a raw coroutine desugars to, but that's not as true as most people think it is...another story for another time). Now, "run_until_complete{}" and "gather()" initially take their arguments and coerce them into Tasks/Futures--that detection, coercion, and construction takes time and consumes a lot of overhead. That's good to know (since many people are paying that coercion tax unknowingly), but it muddies the boundary between "overhead of waiting for an asyncio operation to complete" and "overhead of starting an asyncio operation". Either calling the lower-level functions that run_until_complete()/gather() use internally, or else separating out benchmarks into ones that pass Futures/Tasks/regular coroutines might be appropriate.
- Benchmarking "asyncio.sleep(0)" as a means of determining the bare-minimum await time of a Python event loop is a bad idea. sleep(0) is very special (more details here: https://news.ycombinator.com/item?id=46056895) and not representative. To benchmark "time it takes for the event loop to spin once and produce a result"/the python equivalent of process.nextTick, it'd be better to use low-level loop methods like "call_soon" or defer completion to a Task and await that.
1) Measurements are faulty. List of 1,000 ints can be 4x smaller. Most time measurements depend on circumstances that are not mentioned, therefore can't be reproduced.
2) Brainrot AI style. Hashmap is not "200x faster than list!", that's not how complexity works.
3) orjson/ujson are faulty, which is one of the reasons they don't replace stdlib implementation. Expect crashes, broken jsons, anything from them
4) What actually will be used in number-crunching applications - numpy or similar libraries - is not even mentioned.
Your computer is slow. /s