Related
I am trying to figure out system design behind Google Trends (or any other such large scale trend feature like Twitter).
Challenges:
Need to process large amount of data to calculate trend.
Filtering support - by time, region, category etc.
Need a way to store for archiving/offline processing. Filtering support might require multi dimension storage.
This is what my assumption is (I have zero practial experience of MapReduce/NoSQL technologies)
Each search item from user will maintain set of attributes that will be stored and eventually processed.
As well as maintaining list of searches by time stamp, region of search, category etc.
Example:
Searching for Kurt Cobain term:
Kurt-> (Time stamp, Region of search origin, category ,etc.)
Cobain-> (Time stamp, Region of search origin, category ,etc.)
Question:
How do they efficiently calculate frequency of search term ?
In other words, given a large data set, how do they find top 10 frequent items in distributed scale-able manner ?
Well... finding out the top K terms is not really a big problem. One of the key ideas in this fields have been the idea of "stream processing", i.e., to perform the operation in a single pass of the data and sacrificing some accuracy to get a probabilistic answer. Thus, assume you get a stream of data like the following:
A B K A C A B B C D F G A B F H I B A C F I U X A C
What you want is the top K items. Naively, one would maintain a counter for each item, and at the end sort by the count of each item. This takes O(U) space and O(max(U*log(U), N)) time, where U is the number of unique items and N is the number of items in the list.
In case U is small, this is not really a big problem. But once you are in the domain of search logs with billions or trillions of unique searches, the space consumption starts to become a problem.
So, people came up with the idea of "count-sketches" (you can read up more here: count min sketch page on wikipedia). Here you maintain a hash table A of length n and create two hashes for each item:
h1(x) = 0 ... n-1 with uniform probability
h2(x) = 0/1 each with probability 0.5
You then do A[h1[x]] += h2[x]. The key observation is that since each value randomly hashes to +/-1, E[ A[h1[x]] * h2[x] ] = count(x), where E is the expected value of the expression, and count is the number of times x appeared in the stream.
Of course, the problem with this approach is that each estimate still has a large variance, but that can be dealt with by maintaining a large set of hash counters and taking the average or the minimum count from each set.
With this sketch data structure, you are able to get an approximate frequency of each item. Now, you simply maintain a list of 10 items with the largest frequency estimates till now, and at the end you will have your list.
How exactly a particular private company does it is likely not publicly available, and how to evaluate the effectiveness of such a system is at the discretion of the designer (be it you or Google or whoever)
But many of the tools and research is out there to get you started. Check out some of the Big Data tools, including many of the top-level Apache projects, like Storm, which allows for the processing of streaming data in real-time
Also check out some of the Big Data and Web Science conferences like KDD or WSDM, as well as papers put out by Google Research
How to design such a system is challenging with no correct answer, but the tools and research are available to get you started
I'm trying to come up with a weighted algorithm for an application. In the application, there is a limited amount of space available for different elements. Once all the space is occupied, the algorithm should choose the best element(s) to remove in order to make space for new elements.
There are different attributes which should affect this decision. For example:
T: Time since last accessed. (It's best to replace something that hasn't been accessed in a while.)
N: Number of times accessed. (It's best to replace something which hasn't been accessed many times.)
R: Number of elements which need to be removed in order to make space for the new element. (It's best to replace the least amount of elements. Ideally this should also take into consideration the T and N attributes of each element being replaced.)
I have 2 problems:
Figuring out how much weight to give each of these attributes.
Figuring out how to calculate the weight for an element.
(1) I realize that coming up with the weight for something like this is very subjective, but I was hoping that there's a standard method or something that can help me in deciding how much weight to give each attribute. For example, I was thinking that one method might be to come up with a set of two sample elements and then manually compare the two and decide which one should ultimately be chosen. Here's an example:
Element A: N = 5, T = 2 hours ago.
Element B: N = 4, T = 10 minutes ago.
In this example, I would probably want A to be the element that is chosen to be replaced since although it was accessed one more time, it hasn't been accessed in a lot of time compared with B. This method seems like it would take a lot of time, and would involve making a lot of tough, subjective decisions. Additionally, it may not be trivial to come up with the resulting weights at the end.
Another method I came up with was to just arbitrarily choose weights for the different attributes and then use the application for a while. If I notice anything obviously wrong with the algorithm, I could then go in and slightly modify the weights. This is basically a "guess and check" method.
Both of these methods don't seem that great and I'm hoping there's a better solution.
(2) Once I do figure out the weight, I'm not sure which way is best to calculate the weight. Should I just add everything? (In these examples, I'm assuming that whichever element has the highest replacementWeight should be the one that's going to be replaced.)
replacementWeight = .4*T - .1*N - 2*R
or multiply everything?
replacementWeight = (T) * (.5*N) * (.1*R)
What about not using constants for the weights? For example, sure "Time" (T) may be important, but once a specific amount of time has passed, it starts not making that much of a difference. Essentially I would lump it all in an "a lot of time has passed" bin. (e.g. even though 8 hours and 7 hours have an hour difference between the two, this difference might not be as significant as the difference between 1 minute and 5 minutes since these two are much more recent.) (Or another example: replacing (R) 1 or 2 elements is fine, but when I start needing to replace 5 or 6, that should be heavily weighted down... therefore it shouldn't be linear.)
replacementWeight = 1/T + sqrt(N) - R*R
Obviously (1) and (2) are closely related, which is why I'm hoping that there's a better way to come up with this sort of algorithm.
What you are describing is the classic problem of choosing a cache replacement policy. Which policy is best for you, depends on your data, but the following usually works well:
First, always store a new object in the cache, evicting the R worst one(s). There is no way to know a priori if an object should be stored or not. If the object is not useful, it will fall out of the cache again soon.
The popular squid cache implements the following cache replacement algorithms:
Least Recently Used (LRU):
replacementKey = -T
Least Frequently Used with Dynamic Aging (LFUDA):
replacementKey = N + C
Greedy-Dual-Size-Frequency (GDSF):
replacementKey = (N/R) + C
C refers to a cache age factor here. C is basically the replacementKey of the item that was evicted last (or zero).
NOTE: The replacementKey is calculated when an object is inserted or accessed, and stored alongside the object. The object with the smallest replacementKey is evicted.
LRU is simple and often good enough. The bigger your cache, the better it performs.
LFUDA and GDSF both are tradeoffs. LFUDA prefers to keep large objects even if they are less popular, under the assumption that one hit to a large object makes up lots of hits for smaller objects. GDSF basically makes the opposite tradeoff, keeping many smaller objects over fewer large objects. From what you write, the latter might be a good fit.
If none of these meet your needs, you can calculate optimal values for T, N and R (and compare different formulas for combining them) by minimizing regret, the difference in performance between your formula and the optimal algorithm, using, for example, Linear regression.
This is a completely subjective issue -- as you yourself point out. And a distinct possibility is that if your test cases consist of pairs (A,B) where you prefer A to B, then you might find that you prefer A to B , B to C but also C over A -- i.e. its not an ordering.
If you are not careful, your function might not exist !
If you can define a scalar function of your input variables, with various parameters for coefficients and exponents, you might be able to estimate said parameters by using regression, but you will need an awful lot of data if you have many parameters.
This is the classical statistician's approach of first reviewing the data to IDENTIFY a model, and then using that model to ESTIMATE a particular realisation of the model. There are large books on this subject.
I'm attempting to estimate the total amount of results for app engine queries that will return large amounts of results.
In order to do this, I assigned a random floating point number between 0 and 1 to every entity. Then I executed the query for which I wanted to estimate the total results with the following 3 settings:
* I ordered by the random numbers that I had assigned in ascending order
* I set the offset to 1000
* I fetched only one entity
I then plugged the entities's random value that I had assigned for this purpose into the following equation to estimate the total results (since I used 1000 as the offset above, the value of OFFSET would be 1000 in this case):
1 / RANDOM * OFFSET
The idea is that since each entity has a random number assigned to it, and I am sorting by that random number, the entity's random number assignment should be proportionate to the beginning and end of the results with respect to its offset (in this case, 1000).
The problem I am having is that the results I am getting are giving me low estimates. And the estimates are lower, the lower the offset. I had anticipated that the lower the offset that I used, the less accurate the estimate should be, but I thought that the margin of error would be both above and below the actual number of results.
Below is a chart demonstrating what I am talking about. As you can see, the predictions get more consistent (accurate) as the offset increases from 1000 to 5000. But then the predictions predictably follow a 4 part polynomial. (y = -5E-15x4 + 7E-10x3 - 3E-05x2 + 0.3781x + 51608).
Am I making a mistake here, or does the standard python random number generator not distribute numbers evenly enough for this purpose?
Thanks!
Edit:
It turns out that this problem is due to my mistake. In another part of the program, I was grabbing entities from the beginning of the series, doing an operation, then re-assigning the random number. This resulted in a denser distribution of random numbers towards the end.
I did a little more digging into this concept, fixed the problem, and tried it again on a different query (so the number of results are different from above). I found that this idea can be used to estimate the total results for a query. One thing of note is that the "error" is very similar for offsets that are close by. When I did a scatter chart in excel, I expected the accuracy of the predictions at each offset to "cloud". Meaning that offsets at the very begging would produce a larger, less dense cloud that would converge to a very tiny, dense could around the actual value as the offsets got larger. This is not what happened as you can see below in the cart of how far off the predictions were at each offset. Where I thought there would be a cloud of dots, there is a line instead.
This is a chart of the maximum after each offset. For example the maximum error for any offset after 10000 was less than 1%:
When using GAE it makes a lot more sense not to try to do large amounts work on reads - it's built and optimized for very fast requests turnarounds. In this case it's actually more efficent to maintain a count of your results as and when you create the entities.
If you have a standard query, this is fairly easy - just use a sharded counter when creating the entities. You can seed this using a map reduce job to get the initial count.
If you have queries that might be dynamic, this is more difficult. If you know the range of possible queries that you might perform, you'd want to create a counter for each query that might run.
If the range of possible queries is infinite, you might want to think of aggregating counters or using them in more creative ways.
If you tell us the query you're trying to run, there might be someone who has a better idea.
Some quick thought:
Have you tried Datastore Statistics API? It may provide a fast and accurate results if you won't update your entities set very frequently.
http://code.google.com/appengine/docs/python/datastore/stats.html
[EDIT1.]
I did some math things, I think the estimate method you purposed here, could be rephrased as an "Order statistic" problem.
http://en.wikipedia.org/wiki/Order_statistic#The_order_statistics_of_the_uniform_distribution
For example:
If the actual entities number is 60000, the question equals to "what's the probability that your 1000th [2000th, 3000th, .... ] sample falling in the interval [l,u]; therefore, the estimated total entities number based on this sample, will have an acceptable error to 60000."
If the acceptable error is 5%, the interval [l, u] will be [0.015873015873015872, 0.017543859649122806]
I think the probability won't be very large.
This doesn't directly deal with the calculations aspect of your question, but would using the count attribute of a query object work for you? Or have you tried that out and it's not suitable? As per the docs, it's only slightly faster than retrieving all of the data, but on the plus side it would give you the actual number of results.
http://code.google.com/appengine/docs/python/datastore/queryclass.html#Query_count
I'm currently preparing for an interview, and it reminded me of a question I was once asked in a previous interview that went something like this:
"You have been asked to design some software to continuously display the top 10 search terms on Google. You are given access to a feed that provides an endless real-time stream of search terms currently being searched on Google. Describe what algorithm and data structures you would use to implement this. You are to design two variations:
(i) Display the top 10 search terms of all time (i.e. since you started reading the feed).
(ii) Display only the top 10 search terms for the past month, updated hourly.
You can use an approximation to obtain the top 10 list, but you must justify your choices."
I bombed in this interview and still have really no idea how to implement this.
The first part asks for the 10 most frequent items in a continuously growing sub-sequence of an infinite list. I looked into selection algorithms, but couldn't find any online versions to solve this problem.
The second part uses a finite list, but due to the large amount of data being processed, you can't really store the whole month of search terms in memory and calculate a histogram every hour.
The problem is made more difficult by the fact that the top 10 list is being continuously updated, so somehow you need to be calculating your top 10 over a sliding window.
Any ideas?
Frequency Estimation Overview
There are some well-known algorithms that can provide frequency estimates for such a stream using a fixed amount of storage. One is Frequent, by Misra and Gries (1982). From a list of n items, it find all items that occur more than n / k times, using k - 1 counters. This is a generalization of Boyer and Moore's Majority algorithm (Fischer-Salzberg, 1982), where k is 2. Manku and Motwani's LossyCounting (2002) and Metwally's SpaceSaving (2005) algorithms have similar space requirements, but can provide more accurate estimates under certain conditions.
The important thing to remember is that these algorithms can only provide frequency estimates. Specifically, the Misra-Gries estimate can under-count the actual frequency by (n / k) items.
Suppose that you had an algorithm that could positively identify an item only if it occurs more than 50% of the time. Feed this algorithm a stream of N distinct items, and then add another N - 1 copies of one item, x, for a total of 2N - 1 items. If the algorithm tells you that x exceeds 50% of the total, it must have been in the first stream; if it doesn't, x wasn't in the initial stream. In order for the algorithm to make this determination, it must store the initial stream (or some summary proportional to its length)! So, we can prove to ourselves that the space required by such an "exact" algorithm would be Ω(N).
Instead, these frequency algorithms described here provide an estimate, identifying any item that exceeds the threshold, along with some items that fall below it by a certain margin. For example the Majority algorithm, using a single counter, will always give a result; if any item exceeds 50% of the stream, it will be found. But it might also give you an item that occurs only once. You wouldn't know without making a second pass over the data (using, again, a single counter, but looking only for that item).
The Frequent Algorithm
Here's a simple description of Misra-Gries' Frequent algorithm. Demaine (2002) and others have optimized the algorithm, but this gives you the gist.
Specify the threshold fraction, 1 / k; any item that occurs more than n / k times will be found. Create an an empty map (like a red-black tree); the keys will be search terms, and the values will be a counter for that term.
Look at each item in the stream.
If the term exists in the map, increment the associated counter.
Otherwise, if the map less than k - 1 entries, add the term to the map with a counter of one.
However, if the map has k - 1 entries already, decrement the counter in every entry. If any counter reaches zero during this process, remove it from the map.
Note that you can process an infinite amount of data with a fixed amount of storage (just the fixed-size map). The amount of storage required depends only on the threshold of interest, and the size of the stream does not matter.
Counting Searches
In this context, perhaps you buffer one hour of searches, and perform this process on that hour's data. If you can take a second pass over this hour's search log, you can get an exact count of occurrences of the top "candidates" identified in the first pass. Or, maybe its okay to to make a single pass, and report all the candidates, knowing that any item that should be there is included, and any extras are just noise that will disappear in the next hour.
Any candidates that really do exceed the threshold of interest get stored as a summary. Keep a month's worth of these summaries, throwing away the oldest each hour, and you would have a good approximation of the most common search terms.
Well, looks like an awful lot of data, with a perhaps prohibitive cost to store all frequencies. When the amount of data is so large that we cannot hope to store it all, we enter the domain of data stream algorithms.
Useful book in this area:
Muthukrishnan - "Data Streams: Algorithms and Applications"
Closely related reference to the problem at hand which I picked from the above:
Manku, Motwani - "Approximate Frequency Counts over Data Streams" [pdf]
By the way, Motwani, of Stanford, (edit) was an author of the very important "Randomized Algorithms" book. The 11th chapter of this book deals with this problem. Edit: Sorry, bad reference, that particular chapter is on a different problem. After checking, I instead recommend section 5.1.2 of Muthukrishnan's book, available online.
Heh, nice interview question.
This is one of the research project that I am current going through. The requirement is almost exactly as yours, and we have developed nice algorithms to solve the problem.
The Input
The input is an endless stream of English words or phrases (we refer them as tokens).
The Output
Output top N tokens we have seen so
far (from all the tokens we have
seen!)
Output top N tokens in a
historical window, say, last day or
last week.
An application of this research is to find the hot topic or trends of topic in Twitter or Facebook. We have a crawler that crawls on the website, which generates a stream of words, which will feed into the system. The system then will output the words or phrases of top frequency either at overall or historically. Imagine in last couple of weeks the phrase "World Cup" would appears many times in Twitter. So does "Paul the octopus". :)
String into Integers
The system has an integer ID for each word. Though there is almost infinite possible words on the Internet, but after accumulating a large set of words, the possibility of finding new words becomes lower and lower. We have already found 4 million different words, and assigned a unique ID for each. This whole set of data can be loaded into memory as a hash table, consuming roughly 300MB memory. (We have implemented our own hash table. The Java's implementation takes huge memory overhead)
Each phrase then can be identified as an array of integers.
This is important, because sorting and comparisons on integers is much much faster than on strings.
Archive Data
The system keeps archive data for every token. Basically it's pairs of (Token, Frequency). However, the table that stores the data would be so huge such that we have to partition the table physically. Once partition scheme is based on ngrams of the token. If the token is a single word, it is 1gram. If the token is two-word phrase, it is 2gram. And this goes on. Roughly at 4gram we have 1 billion records, with table sized at around 60GB.
Processing Incoming Streams
The system will absorbs incoming sentences until memory becomes fully utilized (Ya, we need a MemoryManager). After taking N sentences and storing in memory, the system pauses, and starts tokenize each sentence into words and phrases. Each token (word or phrase) is counted.
For highly frequent tokens, they are always kept in memory. For less frequent tokens, they are sorted based on IDs (remember we translate the String into an array of integers), and serialized into a disk file.
(However, for your problem, since you are counting only words, then you can put all word-frequency map in memory only. A carefully designed data structure would consume only 300MB memory for 4 million different words. Some hint: use ASCII char to represent Strings), and this is much acceptable.
Meanwhile, there will be another process that is activated once it finds any disk file generated by the system, then start merging it. Since the disk file is sorted, merging would take a similar process like merge sort. Some design need to be taken care at here as well, since we want to avoid too many random disk seeks. The idea is to avoid read (merge process)/write (system output) at the same time, and let the merge process read form one disk while writing into a different disk. This is similar like to implementing a locking.
End of Day
At end of day, the system will have many frequent tokens with frequency stored in memory, and many other less frequent tokens stored in several disk files (and each file is sorted).
The system flush the in-memory map into a disk file (sort it). Now, the problem becomes merging a set of sorted disk file. Using similar process, we would get one sorted disk file at the end.
Then, the final task is to merge the sorted disk file into archive database.
Depends on the size of archive database, the algorithm works like below if it is big enough:
for each record in sorted disk file
update archive database by increasing frequency
if rowcount == 0 then put the record into a list
end for
for each record in the list of having rowcount == 0
insert into archive database
end for
The intuition is that after sometime, the number of inserting will become smaller and smaller. More and more operation will be on updating only. And this updating will not be penalized by index.
Hope this entire explanation would help. :)
You could use a hash table combined with a binary search tree. Implement a <search term, count> dictionary which tells you how many times each search term has been searched for.
Obviously iterating the entire hash table every hour to get the top 10 is very bad. But this is google we're talking about, so you can assume that the top ten will all get, say over 10 000 hits (it's probably a much larger number though). So every time a search term's count exceeds 10 000, insert it in the BST. Then every hour, you only have to get the first 10 from the BST, which should contain relatively few entries.
This solves the problem of top-10-of-all-time.
The really tricky part is dealing with one term taking another's place in the monthly report (for example, "stack overflow" might have 50 000 hits for the past two months, but only 10 000 the past month, while "amazon" might have 40 000 for the past two months but 30 000 for the past month. You want "amazon" to come before "stack overflow" in your monthly report). To do this, I would store, for all major (above 10 000 all-time searches) search terms, a 30-day list that tells you how many times that term was searched for on each day. The list would work like a FIFO queue: you remove the first day and insert a new one each day (or each hour, but then you might need to store more information, which means more memory / space. If memory is not a problem do it, otherwise go for that "approximation" they're talking about).
This looks like a good start. You can then worry about pruning the terms that have > 10 000 hits but haven't had many in a long while and stuff like that.
case i)
Maintain a hashtable for all the searchterms, as well as a sorted top-ten list separate from the hashtable. Whenever a search occurs, increment the appropriate item in the hashtable and check to see if that item should now be switched with the 10th item in the top-ten list.
O(1) lookup for the top-ten list, and max O(log(n)) insertion into the hashtable (assuming collisions managed by a self-balancing binary tree).
case ii)
Instead of maintaining a huge hashtable and a small list, we maintain a hashtable and a sorted list of all items. Whenever a search is made, that term is incremented in the hashtable, and in the sorted list the term can be checked to see if it should switch with the term after it. A self-balancing binary tree could work well for this, as we also need to be able to query it quickly (more on this later).
In addition we also maintain a list of 'hours' in the form of a FIFO list (queue). Each 'hour' element would contain a list of all searches done within that particular hour. So for example, our list of hours might look like this:
Time: 0 hours
-Search Terms:
-free stuff: 56
-funny pics: 321
-stackoverflow: 1234
Time: 1 hour
-Search Terms:
-ebay: 12
-funny pics: 1
-stackoverflow: 522
-BP sucks: 92
Then, every hour: If the list has at least 720 hours long (that's the number of hours in 30 days), look at the first element in the list, and for each search term, decrement that element in the hashtable by the appropriate amount. Afterwards, delete that first hour element from the list.
So let's say we're at hour 721, and we're ready to look at the first hour in our list (above). We'd decrement free stuff by 56 in the hashtable, funny pics by 321, etc., and would then remove hour 0 from the list completely since we will never need to look at it again.
The reason we maintain a sorted list of all terms that allows for fast queries is because every hour after as we go through the search terms from 720 hours ago, we need to ensure the top-ten list remains sorted. So as we decrement 'free stuff' by 56 in the hashtable for example, we'd check to see where it now belongs in the list. Because it's a self-balancing binary tree, all of that can be accomplished nicely in O(log(n)) time.
Edit: Sacrificing accuracy for space...
It might be useful to also implement a big list in the first one, as in the second one. Then we could apply the following space optimization on both cases: Run a cron job to remove all but the top x items in the list. This would keep the space requirement down (and as a result make queries on the list faster). Of course, it would result in an approximate result, but this is allowed. x could be calculated before deploying the application based on available memory, and adjusted dynamically if more memory becomes available.
Rough thinking...
For top 10 all time
Using a hash collection where a count for each term is stored (sanitize terms, etc.)
An sorted array which contains the ongoing top 10, a term/count in added to this array whenever the count of a term becomes equal or greater than the smallest count in the array
For monthly top 10 updated hourly:
Using an array indexed on number of hours elapsed since start modulo 744 (the number of hours during a month), which array entries consist of hash collection where a count for each term encountered during this hour-slot is stored. An entry is reset whenever the hour-slot counter changes
the stats in the array indexed on hour-slot needs to be collected whenever the current hour-slot counter changes (once an hour at most), by copying and flattening the content of this array indexed on hour-slots
Errr... make sense? I didn't think this through as I would in real life
Ah yes, forgot to mention, the hourly "copying/flattening" required for the monthly stats can actually reuse the same code used for the top 10 of all time, a nice side effect.
Exact solution
First, a solution that guarantees correct results, but requires a lot of memory (a big map).
"All-time" variant
Maintain a hash map with queries as keys and their counts as values. Additionally, keep a list f 10 most frequent queries so far and the count of the 10th most frequent count (a threshold).
Constantly update the map as the stream of queries is read. Every time a count exceeds the current threshold, do the following: remove the 10th query from the "Top 10" list, replace it with the query you've just updated, and update the threshold as well.
"Past month" variant
Keep the same "Top 10" list and update it the same way as above. Also, keep a similar map, but this time store vectors of 30*24 = 720 count (one for each hour) as values. Every hour do the following for every key: remove the oldest counter from the vector add a new one (initialized to 0) at the end. Remove the key from the map if the vector is all-zero. Also, every hour you have to calculate the "Top 10" list from scratch.
Note: Yes, this time we're storing 720 integers instead of one, but there are much less keys (the all-time variant has a really long tail).
Approximations
These approximations do not guarantee the correct solution, but are less memory-consuming.
Process every N-th query, skipping the rest.
(For all-time variant only) Keep at most M key-value pairs in the map (M should be as big as you can afford). It's a kind of an LRU cache: every time you read a query that is not in the map, remove the least recently used query with count 1 and replace it with the currently processed query.
Top 10 search terms for the past month
Using memory efficient indexing/data structure, such as tightly packed tries (from wikipedia entries on tries) approximately defines some relation between memory requirements and n - number of terms.
In case that required memory is available (assumption 1), you can keep exact monthly statistic and aggregate it every month into all time statistic.
There is, also, an assumption here that interprets the 'last month' as fixed window.
But even if the monthly window is sliding the above procedure shows the principle (sliding can be approximated with fixed windows of given size).
This reminds me of round-robin database with the exception that some stats are calculated on 'all time' (in a sense that not all data is retained; rrd consolidates time periods disregarding details by averaging, summing up or choosing max/min values, in given task the detail that is lost is information on low frequency items, which can introduce errors).
Assumption 1
If we can not hold perfect stats for the whole month, then we should be able to find a certain period P for which we should be able to hold perfect stats.
For example, assuming we have perfect statistics on some time period P, which goes into month n times.
Perfect stats define function f(search_term) -> search_term_occurance.
If we can keep all n perfect stat tables in memory then sliding monthly stats can be calculated like this:
add stats for the newest period
remove stats for the oldest period (so we have to keep n perfect stat tables)
However, if we keep only top 10 on the aggregated level (monthly) then we will be able to discard a lot of data from the full stats of the fixed period. This gives already a working procedure which has fixed (assuming upper bound on perfect stat table for period P) memory requirements.
The problem with the above procedure is that if we keep info on only top 10 terms for a sliding window (similarly for all time), then the stats are going to be correct for search terms that peak in a period, but might not see the stats for search terms that trickle in constantly over time.
This can be offset by keeping info on more than top 10 terms, for example top 100 terms, hoping that top 10 will be correct.
I think that further analysis could relate the minimum number of occurrences required for an entry to become a part of the stats (which is related to maximum error).
(In deciding which entries should become part of the stats one could also monitor and track the trends; for example if a linear extrapolation of the occurrences in each period P for each term tells you that the term will become significant in a month or two you might already start tracking it. Similar principle applies for removing the search term from the tracked pool.)
Worst case for the above is when you have a lot of almost equally frequent terms and they change all the time (for example if tracking only 100 terms, then if top 150 terms occur equally frequently, but top 50 are more often in first month and lest often some time later then the statistics would not be kept correctly).
Also there could be another approach which is not fixed in memory size (well strictly speaking neither is the above), which would define minimum significance in terms of occurrences/period (day, month, year, all-time) for which to keep the stats. This could guarantee max error in each of the stats during aggregation (see round robin again).
What about an adaption of the "clock page replacement algorithm" (also known as "second-chance")? I can imagine it to work very well if the search requests are distributed evenly (that means most searched terms appear regularly rather than 5mio times in a row and then never again).
Here's a visual representation of the algorithm:
The problem is not universally solvable when you have a fixed amount of memory and an 'infinite' (think very very large) stream of tokens.
A rough explanation...
To see why, consider a token stream that has a particular token (i.e., word) T every N tokens in the input stream.
Also, assume that the memory can hold references (word id and counts) to at most M tokens.
With these conditions, it is possible to construct an input stream where the token T will never be detected if the N is large enough so that the stream contains different M tokens between T's.
This is independent of the top-N algorithm details. It only depends on the limit M.
To see why this is true, consider the incoming stream made up of groups of two identical tokens:
T a1 a2 a3 ... a-M T b1 b2 b3 ... b-M ...
where the a's, and b's are all valid tokens not equal to T.
Notice that in this stream, the T appears twice for each a-i and b-i. Yet it appears rarely enough to be flushed from the system.
Starting with an empty memory, the first token (T) will take up a slot in the memory (bounded by M). Then a1 will consume a slot, all the way to a-(M-1) when the M is exhausted.
When a-M arrives the algorithm has to drop one symbol so let it be the T.
The next symbol will be b-1 which will cause a-1 to be flushed, etc.
So, the T will not stay memory-resident long enough to build up a real count. In short, any algorithm will miss a token of low enough local frequency but high global frequency (over the length of the stream).
Store the count of search terms in a giant hash table, where each new search causes a particular element to be incremented by one. Keep track of the top 20 or so search terms; when the element in 11th place is incremented, check if it needs to swap positions with #10* (it's not necessary to keep the top 10 sorted; all you care about is drawing the distinction between 10th and 11th).
*Similar checks need to be made to see if a new search term is in 11th place, so this algorithm bubbles down to other search terms too -- so I'm simplifying a bit.
sometimes the best answer is "I don't know".
Ill take a deeper stab. My first instinct would be to feed the results into a Q. A process would continually process items coming into the Q. The process would maintain a map of
term -> count
each time a Q item is processed, you simply look up the search term and increment the count.
At the same time, I would maintain a list of references to the top 10 entries in the map.
For the entry that was currently implemented, see if its count is greater than the count of the count of the smallest entry in the top 10.(if not in the list already). If it is, replace the smallest with the entry.
I think that would work. No operation is time intensive. You would have to find a way to manage the size of the count map. but that should good enough for an interview answer.
They are not expecting a solution, that want to see if you can think. You dont have to write the solution then and there....
One way is that for every search, you store that search term and its time stamp. That way, finding the top ten for any period of time is simply a matter of comparing all search terms within the given time period.
The algorithm is simple, but the drawback would be greater memory and time consumption.
What about using a Splay Tree with 10 nodes? Each time you try to access a value (search term) that is not contained in the tree, throw out any leaf, insert the value instead and access it.
The idea behind this is the same as in my other answer. Under the assumption that the search terms are accessed evenly/regularly this solution should perform very well.
edit
One could also store some more search terms in the tree (the same goes for the solution I suggest in my other answer) in order to not delete a node that might be accessed very soon again. The more values one stores in it, the better the results.
Dunno if I understand it right or not.
My solution is using heap.
Because of top 10 search items, I build a heap with size 10.
Then update this heap with new search. If a new search's frequency is greater than heap(Max Heap) top, update it. Abandon the one with smallest frequency.
But, how to calculate the frequency of the specific search will be counted on something else.
Maybe as everyone stated, the data stream algorithm....
Use cm-sketch to store count of all searches since beginning, keep a min-heap of size 10 with it for top 10.
For monthly result, keep 30 cm-sketch/hash-table and min-heap with it, each one start counting and updating from last 30, 29 .., 1 day. As a day pass, clear the last and use it as day 1.
Same for hourly result, keep 60 hash-table and min-heap and start counting for last 60, 59, ...1 minute. As a minute pass, clear the last and use it as minute 1.
Montly result is accurate in range of 1 day, hourly result is accurate in range of 1 min
I am looking for an algorithm that determines percentiles for live data capture.
For example, consider the development of a server application.
The server might have response times as follows:
17 ms
33 ms
52 ms
60 ms
55 ms
etc.
It is useful to report the 90th percentile response time, 80th percentile response time, etc.
The naive algorithm is to insert each response time into a list. When statistics are requested, sort the list and get the values at the proper positions.
Memory usages scales linearly with the number of requests.
Is there an algorithm that yields "approximate" percentile statistics given limited memory usage? For example, let's say I want to solve this problem in a way that I process millions of requests but only want to use say one kilobyte of memory for percentile tracking (discarding the tracking for old requests is not an option since the percentiles are supposed to be for all requests).
Also require that there is no a priori knowledge of the distribution. For example, I do not want to specify any ranges of buckets ahead of time.
If you want to keep the memory usage constant as you get more and more data, then you're going to have to resample that data somehow. That implies that you must apply some sort of rebinning scheme. You can wait until you acquire a certain amount of raw inputs before beginning the rebinning, but you cannot avoid it entirely.
So your question is really asking "what's the best way of dynamically binning my data"? There are lots of approaches, but if you want to minimise your assumptions about the range or distribution of values you may receive, then a simple approach is to average over buckets of fixed size k, with logarithmically distributed widths. For example, lets say you want to hold 1000 values in memory at any one time. Pick a size for k, say 100. Pick your minimum resolution, say 1ms. Then
The first bucket deals with values between 0-1ms (width=1ms)
Second bucket: 1-3ms (w=2ms)
Third bucket: 3-7ms (w=4ms)
Fourth bucket: 7-15ms (w=8ms)
...
Tenth bucket: 511-1023ms (w=512ms)
This type of log-scaled approach is similar to the chunking systems used in hash table algorithms, used by some filesystems and memory allocation algorithms. It works well when your data has a large dynamic range.
As new values come in, you can choose how you want to resample, depending on your requirements. For example, you could track a moving average, use a first-in-first-out, or some other more sophisticated method. See the Kademlia algorithm for one approach (used by Bittorrent).
Ultimately, rebinning must lose you some information. Your choices regarding the binning will determine the specifics of what information is lost. Another way of saying this is that the constant size memory store implies a trade-off between dynamic range and the sampling fidelity; how you make that trade-off is up to you, but like any sampling problem, there's no getting around this basic fact.
If you're really interested in the pros and cons, then no answer on this forum can hope to be sufficient. You should look into sampling theory. There's a huge amount of research on this topic available.
For what it's worth, I suspect that your server times will have a relatively small dynamic range, so a more relaxed scaling to allow higher sampling of common values may provide more accurate results.
Edit: To answer your comment, here's an example of a simple binning algorithm.
You store 1000 values, in 10 bins. Each bin therefore holds 100 values. Assume each bin is implemented as a dynamic array (a 'list', in Perl or Python terms).
When a new value comes in:
Determine which bin it should be stored in, based on the bin limits you've chosen.
If the bin is not full, append the value to the bin list.
If the bin is full, remove the value at the top of the bin list, and append the new value to the bottom of the bin list. This means old values are thrown away over time.
To find the 90th percentile, sort bin 10. The 90th percentile is the first value in the sorted list (element 900/1000).
If you don't like throwing away old values, then you can implement some alternative scheme to use instead. For example, when a bin becomes full (reaches 100 values, in my example), you could take the average of the oldest 50 elements (i.e. the first 50 in the list), discard those elements, and then append the new average element to the bin, leaving you with a bin of 51 elements that now has space to hold 49 new values. This is a simple example of rebinning.
Another example of rebinning is downsampling; throwing away every 5th value in a sorted list, for example.
I hope this concrete example helps. The key point to take away is that there are lots of ways of achieving a constant memory aging algorithm; only you can decide what is satisfactory given your requirements.
I've once published a blog post on this topic. The blog is now defunct but the article is included in full below.
The basic idea is to reduce the requirement for an exact calculation in favor of "95% percent of responses take 500ms-600ms or less" (for all exact percentiles of 500ms-600ms).
As we’ve recently started feeling that response times of one of our webapps got worse, we decided to spend some time tweaking the apps’ performance. As a first step, we wanted to get a thorough understanding of current response times. For performance evaluations, using minimum, maximum or average response times is a bad idea: “The ‘average’ is the evil of performance optimization and often as helpful as ‘average patient temperature in the hospital'” (MySQL Performance Blog). Instead, performance tuners should be looking at the percentile: “A percentile is the value of a variable below which a certain percent of observations fall” (Wikipedia). In other words: the 95th percentile is the time in which 95% of requests finished. Therefore, a performance goals related to the percentile could be similar to “The 95th percentile should be lower than 800 ms”. Setting such performance goals is one thing, but efficiently tracking them for a live system is another one.
I’ve spent quite some time looking for existing implementations of percentile calculations (e.g. here or here). All of them required storing response times for each and every request and calculate the percentile on demand or adding new response times in order. This was not what I wanted. I was hoping for a solution that would allow memory and CPU efficient live statistics for hundreds of thousands of requests. Storing response times for hundreds of thousands of requests and calculating the percentile on demand does neither sound CPU nor memory efficient.
Such a solution as I was hoping for simply seems not to exist. On second thought, I came up with another idea: For the type of performance evaluation I was looking for, it’s not necessary to get the exact percentile. An approximate answer like “the 95th percentile is between 850ms and 900ms” would totally suffice. Lowering the requirements this way makes an implementation extremely easy, especially if upper and lower borders for the possible results are known. For example, I’m not interested in response times higher than several seconds – they are extremely bad anyway, regardless of being 10 seconds or 15 seconds.
So here is the idea behind the implementation:
Define any random number of response time buckets (e.g. 0-100ms, 100-200ms, 200-400ms, 400-800ms, 800-1200ms, …)
Count number of responses and number of response each bucket (For a response time of 360ms, increment the counter for the 200ms – 400ms bucket)
Estimate the n-th percentile by summing counter for buckets until the sum exceeds n percent of the total
It’s that simple. And here is the code.
Some highlights:
public void increment(final int millis) {
final int i = index(millis);
if (i < _limits.length) {
_counts[i]++;
}
_total++;
}
public int estimatePercentile(final double percentile) {
if (percentile < 0.0 || percentile > 100.0) {
throw new IllegalArgumentException("percentile must be between 0.0 and 100.0, was " + percentile);
}
for (final Percentile p : this) {
if (percentile - p.getPercentage() <= 0.0001) {
return p.getLimit();
}
}
return Integer.MAX_VALUE;
}
This approach only requires two int values (= 8 byte) per bucket, allowing to track 128 buckets with 1K of memory. More than sufficient for analysing response times of a web application using a granularity of 50ms). Additionally, for the sake of performance, I’ve intentionally implemented this without any synchronization(e.g. using AtomicIntegers), knowing that some increments might get lost.
By the way, using Google Charts and 60 percentile counters, I was able to create a nice graph out of one hour of collected response times:
I believe there are many good approximate algorithms for this problem. A good first-cut approach is to simply use a fixed-size array (say 1K worth of data). Fix some probability p. For each request, with probability p, write its response time into the array (replacing the oldest time in there). Since the array is a subsampling of the live stream and since subsampling preserves the distribution, doing the statistics on that array will give you an approximation of the statistics of the full, live stream.
This approach has several advantages: it requires no a-priori information, and it's easy to code. You can build it quickly and experimentally determine, for your particular server, at what point growing the buffer has only a negligible effect on the answer. That is the point where the approximation is sufficiently precise.
If you find that you need too much memory to give you statistics that are precise enough, then you'll have to dig further. Good keywords are: "stream computing", "stream statistics", and of course "percentiles". You can also try "ire and curses"'s approach.
(It's been quite some time since this question was asked, but I'd like to point out a few related research papers)
There has been a significant amount of research on approximate percentiles of data streams in the past few years. A few interesting papers with full algorithm definitions:
A fast algorithm for approximate quantiles in high speed data streams
Space-and time-efficient deterministic algorithms for biased quantiles over data streams
Effective computation of biased quantiles over data streams
All of these papers propose algorithms with sub-linear space complexity for the computation of approximate percentiles over a data stream.
Try the simple algorithm defined in the paper “Sequential Procedure for Simultaneous Estimation of Several Percentiles” (Raatikainen). It’s fast, requires 2*m+3 markers (for m percentiles) and tends to an accurate approximation quickly.
Use a dynamic array T[] of large integers or something where T[n] counts the numer of times the response time was n milliseconds. If you really are doing statistics on a server application then possibly 250 ms response times are your absolute limit anyway. So your 1 KB holds one 32 bits integer for every ms between 0 and 250, and you have some room to spare for an overflow bin.
If you want something with more bins, go with 8 bit numbers for 1000 bins, and the moment a counter would overflow (i.e. 256th request at that response time) you shift the bits in all bins down by 1. (effectively halving the value in all bins). This means you disregard all bins that capture less than 1/127th of the delays that the most visited bin catches.
If you really, really need a set of specific bins I'd suggest using the first day of requests to come up with a reasonable fixed set of bins. Anything dynamic would be quite dangerous in a live, performance sensitive application. If you choose that path you'd better know what your doing, or one day you're going to get called out of bed to explain why your statistics tracker is suddenly eating 90% CPU and 75% memory on the production server.
As for additional statistics: For mean and variance there are some nice recursive algorithms that take up very little memory. These two statistics can be usefull enough in themselves for a lot of distributions because the central limit theorem states that distributions that that arise from a sufficiently large number of independent variables approach the normal distribution (which is fully defined by mean and variance) you can use one of the normality tests on the last N (where N sufficiently large but constrained by your memory requirements) to monitor wether the assumption of normality still holds.
#thkala started off with some literature citations. Let me extend that.
Implementations
T-digest from the 2019 Dunning paper has reference implementation in Java, and ports on that page to Python, Go, Javascript, C++, Scala, C, Clojure, C#, Kotlin, and C++ port by facebook, and a further rust port of that C++ port
Spark implements "GK01" from the 2001 Greenwald/Khanna paper for approximate quantiles
Beam: org.apache.beam.sdk.transforms.ApproximateQuantiles has approximate quantiles
Java: Guava:com.google.common.math.Quantiles implements exact quantiles, thus taking more memory
Rust: quantiles crate has implementations for the 2001 GK algorithm "GK01", and the 2005 CKMS algorithm. (caution: I found the CKMS implementation slow - issue)
C++: boost quantiles has some code, but I didn't understand it.
I did some profiling of the options in Rust [link] for up to 100M items, and found GK01 the best, T-digest the second, and "keep 1% top values in priority queue" the third.
Literature
2001: Space-efficient online computation of quantile summaries (by Greenwald, Khanna). Implemented in Rust: quantiles::greenwald_khanna.
2004: Medians and beyond: new aggregation techniques for sensor networks (by Shrivastava, Buragohain, Agrawal, Suri). Introduces "q-digests", used for fixed-universe data.
2005: Effective computation of biased quantiles over data streams (by Cormode, Korn, Muthukrishnan, Srivastava)... Implemented in Rust: quantiles::ckms which notes that the IEEE presentation is correct but the self-published one has flaws. With carefully crafted data, space can grow linearly with input size. "Biased" means it focuses on P90/P95/P99 rather than all the percentiles).
2006: Space-and time-efficient deterministic algorithms for biased quantiles over data streams (by Cormode, Korn, Muthukrishnan, Srivastava)... improved space bound over 2005 paper
2007: A fast algorithm for approximate quantiles in high speed data streams (by Zhang, Wang). Claims 60-300x speedup over GK. The 2020 literature review below says this has state-of-the-art space upper bound.
2019 Computing extremely accurate quantiles using t-digests (by Dunning, Ertl). Introduces t-digests, O(log n) space, O(1) updates, O(1) final calculation. It's neat feature is you can build partial digests (e.g. one per day) and merge them into months, then merge months into years. This is what the big query engines use.
2020 A survey of approximate quantile computation on large-scale data (technical report) (by Chen, Zhang).
2021 The t-digest: Efficient estimates of distributions - an approachable wrap-up paper about t-digests.
Cheap hack for P99 of <10M values: just store top 1% in a priority queue!
This'll sound stupid, but if I want to calculate the P99 of 10M float64s, I just created a priority queue with 100k float32s (takes 400kB). This takes only 4x as much space as "GK01" and is much faster. For 5M or fewer items, it takes less space than GK01!!
struct TopValues {
values: std::collections::BinaryHeap<std::cmp::Reverse<ordered_float::NotNan<f32>>>,
}
impl TopValues {
fn new(count: usize) -> Self {
let capacity = std::cmp::max(count / 100, 1);
let values = std::collections::BinaryHeap::with_capacity(capacity);
TopValues { values }
}
fn render(&mut self) -> String {
let p99 = self.values.peek().unwrap().0;
let max = self.values.drain().min().unwrap().0;
format!("TopValues, p99={:.4}, max={:.4}", p99, max)
}
fn insert(&mut self, value: f64) {
let value = value as f32;
let value = std::cmp::Reverse(unsafe { ordered_float::NotNan::new_unchecked(value) });
if self.values.len() < self.values.capacity() {
self.values.push(value);
} else if self.values.peek().unwrap().0 < value.0 {
self.values.pop();
self.values.push(value);
} else {
}
}
}
You can try the following structure:
Take on input n, ie. n = 100.
We'll keep an array of ranges [min, max] sorted by min with count.
Insertion of value x – binary search for min range for x. If not found take preceeding range (where min < x). If value belongs to range (x <= max) increment count. Otherwise insert new range with [min = x, max = x, count = 1].
If number of ranges hits 2*n – collapse/merge array into n (half) by taking min from odd and max from even entries, summing their count.
To get ie. p95 walk from the end summing the count until next addition would hit threshold sum >= 95%, take p95 = min + (max - min) * partial.
It will settle on dynamic ranges of measurements. n can be modified to trade accuracy for memory (to lesser extent cpu). If you make values more discrete, ie. by rounding to 0.01 before insertion – it'll stabilise on ranges sooner.
You could improve accuracy by not assuming that each range holds uniformly distributed entries, ie. something cheap like sum of values which will give you avg = sum / count, it would help to read closer p95 value from range where it sits.
You can also rotate them, ie. after m = 1 000 000 entries start filling new array and take p95 as weighted sum on count in array (if array B has 10% of count of A, then it contributes 10% to p95 value).