Many modern performant solutions will use both, but they're not the same thing.
Mass storage is most efficient when doing large sequential reads and writes, so you normally feed your constant-space streaming algorithms from buffers with a large number of input records.
Sometimes you can just tell the OS do efficient chunking prefetch for you.
The hard part for me is the "transpose" or "striding" problem. i.e. when the data is stored in a series of (x, y, z) files for a given hour, day, or month and I need a time series at a point.
If your intended audience are data scientists then why didn’t you mention Dask?
My first inclination would be to have a counter at zero, read in a line, write out the counter then write out the line, then increment the counter.
The space complexity here is O(1), is it not?
On a 64-bit CPU, y=x+1 is O(1) time when Y < 2^64, but becomes O(log N) time for larger numbers. With space it’s even more complex, sometimes for space saving it makes sense to keep less than 64 bits of these integers, i.e. it may become log(N) for much smaller values.
For this reason, I’m not a bug fan of big O. Useful at job interviews. Probably useful for people working on extremely generic stuff like C++ STL which needs to scale both ways by many orders of magnitude, it’s not uncommon to have a vector with 1 element or 100M elements, the standard library must be good at both ends. But I don’t often solve such generic problems, I usually have some ideas how large my N-s are, and can act accordingly.
Sure, for files larger than 18.45 EB at the minimum, computing the new line count will take O(log N) time.
I'm all for nitpicking the issues with Big O notation, but this instance doesn't strike me as one.
The lesson to take away here is that O(log N) is REALLY REALLY small, for all sensible values of N. If an O(log N) algorithm is slow, it’s not because of the O(log N) part, it’s because of the hidden factors in O notation.
No it is not. "Big O" is a way of expressing asymptotic growth of some quantity, most often running time. But you need a machine model to even express running time. And sure, you can define your machine model to require O(log n) time for adding two numbers, but that's not particularly useful. In particular, the word RAM model—which is what is normally used—assumes a word size of θ(log n), and that basic operations on those words take a constant amount of time.
We use models because otherwise even the simplest things would be impossible to analyse. Or do you want to include a three-level cache hierarchy, micro-op decoding, the paging system, branch mispredictions and memory timings in your complexity analysis? Why is it always the "bUt AdDinG tWo nuMbERs tAkeS O(log n) tImE" hill that HN commenters are willing to die on?
First you need to settle on a model. Then you can express time, space, or some other quantity in that model using Big O notation.
Arguably, I've found that HN is one of the worst places to be precise about Big O notation. Everyone here seems to think they know what it is, only to lay bare their misconceptions about it a second later.
First of all, there's absolutely nothing that says that we have to assume that the number of lines is less than 2^64. Big O notation does not care what happens below that number, it only cares what happens when you go to infinity, which is a lot larger than 2^64. At that point you need arbitrarily sized integers, which are not fixed-space. Hence O(log n). Word size or RAM model does not matter one whit.
This is how Big O notation works, it has nothing to do with practicality. As another example of a similar thing, consider Sorting Networks. The AKS sorting networks has the best "asymptotic depth" of any known sorting networks (O(n log n) depth, i think), but the hidden constants are so massive that there are no situations where the AKS networks are actually practical, for any reasonable input they are always beaten by other (asymptotically worse) networks. That doesn't mean that they "aren't O(n log n)" or whatever.
Second: if we were to actually write this algorithm in the way which is most conservative with space, we wouldn't use a 64-bit integer in the first place. If we only care about memory, we'd start the count with a single-byte datatype (i.e. unsigned char) to keep the count. The algorithm now uses 1 byte of space. When that overflows (i.e. reaches 256 lines), we would change to a bigger data type, i.e. unsigned short. The algorithm now takes 2 bytes of data.
When THAT overflows, we again promote the datatype to an unsigned int. Now we use 4 bytes of data. When that overflows, we finally promote it to a uint64_t. Now we're on 8 bytes of data. When that overflows... (etc.)
See how the memory usage is growing? And how it's growing exactly with the logarithm of the number of lines?
You're just wrong about this. Yes, you can absolutely say "assuming storage of all numbers is O(1), then the algorithm is O(1)" (essentially what OP said) but that is not a default assumption and it's not accurate to reality (since actually storing numbers requires O(log n) of space). And yes, branch prediction misses and cache misses are far more important factors here, but that doesn't change the fact that it uses O(log n) space. That's just simply true.
Also: we're not dying on a hill. OP made an entirely accurate and true statement, which was met with an avalanche of people criticizing him using incorrect arguments and misunderstandings of Big O notation. We're just saying that he was correct all along.
Arbitrarily large numbers, represented in binary, certainly take time that is linear in the length of their representation to add. Nobody's arguing with that. But using a model that assumes bitwise operations would be needlessly complex, so outside of some really specialized contexts, nobody does it.
> You're just wrong about this. Yes, you can absolutely say "assuming storage of all numbers is O(1), then the algorithm is O(1)" (essentially what OP said) but that is not a default assumption
I don't know the basis on which you argue that the word RAM model isn't the default assumption, but I feel confident in claiming that outside of some niche communities, it most certainly is.
> and it's not accurate to reality (since actually storing numbers requires O(log n) of space).
That's why it's called a model. We use models because reality is too messy to work with, and we usually don't care about the exact details. Just ask a physicist, they make simplifying assumptions all the time because without them, they'd be getting absolutely nowhere. It's pretty much the same in algorithmics.
The issue was about space complexity, not time complexity. Very large numbers take more space than smaller numbers.
Does it make sense when expressed compactly like that?
No. Incrementing a 128-bit counter is also O(1) time on amd64/ARM8/... Only when your numbers are arbitrarily large would that become true.
But that's why we use models that make clear what our assumptions are. In particular, the word RAM model (which is normally used to talk about sequential algorithms) assumes a word size of θ(log n), and that some basic operations on words (such as arithmetic) are possible in constant time.
Models obviously have flaws (otherwise they wouldn't be models), but practitioners often make the mistake of overthinking them. How big, exactly, is your file that a constant number of 64-bit integers doesn't suffice? You can easily increment a 512-bit counter in constant time on an 8-bit Arduino. It doesn't matter whether I have to check one hypothetical 512-bit counter, 8 64-bit counters or 64 8-bit counters, it's a constant number and therefore O(1).
Whenever this "Big O is useful for job interviews but not much else" sentiment comes up on HN, it's usually driven by a misunderstanding of what Big O and asymptotic analysis actually mean.
Technically you’re right, I should have put “arbitrarily large” there, but assumed it’s obvious from the context.
> but practitioners often make the mistake of overthinking them
As an experienced practitioner, I happen to know how extremely inaccurate are these models.
When I was inexperienced, the models helped more then they do now, after I have leared about NUMA, cache hierarchy, micro-ops fusion, MMU, prefetcher, DMA IO, interrupts, and more of these hairy implementation details. Before I knew all these things, heuristics based on asymptotic analysis were generally helpful. Now they’re less so, I sometimes deliberately pick an asymptotically slower algorithm or data structure because it’s faster in reality.
I guess this is where I pitch what we call "Algorithm Engineering", which considers both theoretical guarantees and practical performance, including properties of real-world hardware such as the ones you mentioned :) https://en.wikipedia.org/wiki/Algorithm_engineering
Are you sure about that? The most recent source of that article is from 2000. Half of the stuff I have mentioned didn’t exist back then.
As far as I’m aware, there’re very few people in academia who study modern hardware while focused on software performance. I only know about Agner Fog, from technical university of Denmark.
Agner Fog does some cool stuff (and I've used his RNG libraries many times) but as you said he studies modern hardware with a focus on software performance. Algorithm engineering is about finding algorithms that have nice theoretical guarantees (e.g., proving correctness or that there aren't any inputs which would result in substantially worse running time) while keeping in mind that the result would be quite useless if it couldn't be implemented efficiently on the machines we have, so the goal is to come up with algorithms that are also really fast in practice.
I thought that big-O does not actually consider the underlying hardware. How can you specify the CPU model / instruction conditions and still use the semantics of big-O analysis? The mathematical definition of O(_) does not involve hardware, it implies that there exists some coefficients and a constant. I get the point you are making but is big-O an appropriate way of measuring in this context?
When say that comparison sort is O(n * lg n), we're assuming a computational model where comparing two integers is a constant time operation. That's true in most physical computers when we use the CPU's comparison instructions. It's also true in any theoretical model where we define a (finite) largest possible integer. So it works well enough.
I’m pointing out that, for example, comparison sort is only O(n * lg n) in theory if the theoretical model of computation we’re dealing with can compare any two items in constant time. Comparing arbitrarily large integers can not be done in constant time in theory, so if that’s the problem we’re discussing then the complexity will not be O(n * lg n). But again, we generally are implicitly concerned with models of computation where comparison is a constant time operation.
Yes, Big O notation measures asymptotic growth of functions. But what do those functions express? For them to express running times, you need to define a machine model. It's a model because it's not a real machine – we don't stroll into a computer shop and pick up some computer and use that as our reference. Instead, we define an abstract machine model. Often, that's the word RAM model which assumes that for an input size n, the machine can process words of size θ(log n) in constant time. This is reasonable close to how our actual computers work: 64-bit words, and we can do stuff with them in constant time. But it's not perfect, because the RAM model doesn't have any caches (let alone multiple levels of cache hierarchy), doesn't have disks, doesn't model the paging system, etc. If any of these things are of significant concern to you, pick a different model (e.g. an external-memory model to consider memory hierarchy, or look into cache-oblivious algorithms, ...).
But talking about asymptotic running time without specifying which model you're using is absolutely pointless.
I don't have formal CS education. I'm less interested in math and more oriented towards practical aspects. Infinity doesn't exist in practice (except these IEEE float magic values), but we can reason about shape of functions, e.g. "count of executed machine instructions as a function of N".
Not the same thing as CS-defined big O, but close.
Can be useful to reason about complexity of algorithms. More precisely, about complexity of their implementations.
Usually useless to reason about time: a machine instruction takes variable time to run on a modern CPU, the same instruction can take less than 1 cycle or much more than 1M cycles.
If you want to talk about something like supporting arbitrarily large integers or defining the memory of the physical computer we're using, we need to be explicit about that, because it changes a lot of things. After all, a finite-memory computer is not Turing complete, and everything it can compute can be computed in constant time. :)
- Read byte from input
- Write byte to output
- When byte is "\n": write the counter, write a space, increase the counter by one
Most likely, whatever logic you use to turn the counter into a string will already use more memory than your actual loop.
Technically it's still not O(1), as you will need one more bit for your counter variable every time you double the number of lines; realistically, you will most likely be using an integer anyway, so you can consider it "constant".
We have very good asymptotic complexity for multiplication, but don't use the ones that scale the closest to linear because the time constant is enormous.
Also relevant here is the RAM machine model that is often employed, where you can have essentially unlimited memory that can be randomly accessed at fixed time. This again is roughly in line with real practical machines running algorithms that use, and fit within RAM. In reality if you wanted a more physically consistent model eventually your access times must differentiate and increase for increasing memory -- data occupies physical space and must be fetched at increasingly long distances, limited by the speed of light. In principle this means m bits of memory at best can be accessed in O(m^(1/3)) time [1]. Of course, this realism isn't always practically relevant (maybe for datacenter-scale problems, perhaps even larger) and complicates the analysis of algorithms.
[1]: Just for fun, if you want to get really physically accurate, this isn't quite right either I suspect -- that's because with enough data (physical bits) occupying a volume at constant density it will eventually collapse into a black hole, which kinda destroys your computer :). So for planetary-scale computers you eventually need to spread your data across a disk, like a small galaxy (or Discworld, if you prefer :)), so it won't collapse, giving it O(sqrt(m)) access time. Surprise, very large worlds must be flat.
This is related to the BH entropy formula, and the so called Holographic principle, I guess -- the entropy and hence information content of a volume is surprisingly bounded by its area, since a Black Hole's area is proportional to its mass. The weird thing of course is how the universe itself perhaps should collapse to a black hole since it has a lot of stuff in any sufficiently large volume, being young and dynamic it hasn't happened yet. It's alsogiven by its fractal scale, but I guess this digression has grown large enough already :)
(do tell if you want to learn more)
So integers count as constant memory AND can be considered to be arbitrarily large for the purpose of the current problem. Because having more than 2^64 lines in a file will never happen in practice, and even if it does then 2 integers would do the job for the rest of the observable universe.
Be pragmatic, you are on Hacker News.
In my experience, practitioners just have a lot of misconceptions about "Big O".
Or "database"
That's on point, some of the worst and most inefficient code I had to maintain was written by data scientists. They are for sure leagues above any of my data science knowledge in their respective field but the average engineering knowledge and best practices seems pretty low.
"The simplest and most common way to implement indexing is by naming files in a directory:"
mydata/
2019-Jan.csv
2019-Feb.csv
2019-Mar.csv
2019-Apr.csv
mydata/
2019/
11/
12/
2019111200.csv
2019111201.csv
Usually you want to process it into a columnar format like Parquet from there, though
I guess I'm misinterpreting the task, because I'm having trouble seeing how it's possible to access each line in a file without it being O(n) at a minimum. I understand O(log n) to mean that not every data point is "accessed" due to the way the data structure is organized, lines of a file in this case, but to me it appears that not every line would be prepended with a number. My instinct is that there's a lot going on with new lines/line breaks in files that I'm not aware of, so I'd really appreciate any reading material on the subject to help me better understand this. Thanks!
(My mistake obviously, the OP was very clear.)
In term of RAM this can be done an arbitrarily small buffer in the sense that this can be done by reading the file byte by byte. I.e. O(1).
Edit: For all practical purposes the line counter is also constant space.
printf(“0: “);
ln = 1;
while ((ch = getch()) != EOF) {
putch(ch);
if (ch == EOL) {
printf(“%d: “, ln);
ln++;
}
}
Can you elaborate on the O(1) solution?
Ah, simple :)
cat file | wc -l wc -l file
wc -l < fileAnyway, since he said number we probably want `nl` not `wc`.
Separating pipe-setup from processing is sound engineering to me and makes things much easier. For instance as I iterate, I will have cat piped to head piped to filter but eventually I'll take the head out and run the whole thing. That's a trivial ^W versus moving around arguments and worrying about arg order, etc.
However using the input redirect is fine, especially if you format it that way (which is exactly the same thing):
<file wc -l
<file grep meh |wc -l <file grep -c mehI know `< file wc -l | ... | ...` is identical because the redirection is done before the command is executed, but what does putting it in front help with?
The first iteration looks like this:
<file head -n10 | grep foo | tr baz qux
The second iteration looks like this:
<file gzcat | head -n10 | grep foo | tr baz qux
It makes the change much easier, since the first thing written is the first thing that happens. In `gzcat < file`, the logical first step - reading and streaming file - is now the physical second step. Like a German sentence, whenever you want to prepend the file, you have to maintain an item in second position.
And not change anything else in the pipeline.
head filename | ...
In this particular case, people (like me) would wonder, "what's the point" and then go searching for the details of how `cat` and `wc` work, assuming that there's might be a reason the original developer wrote this code, as opposed to the simpler `wc -l` (i.e. Chesterton's Fence).
Also, context is really important. Not that people on the internet ever stop to think about what the context is that the person is writing the code in. I don't write bash scripts that run mission critical workloads. I just ad-hoc type stuff into my shell to produce output I need for stuff at the time.
cat -n file nl file sed = 1.txt
A semi-common beginning programmer's exercise is to write a program that numbers the lines in a text file. The naive solution will use O(n) space, while a bit more thought reveals that this can be done in constant (to be really precise, O(log n) where n is the number of lines in the file) space.