I have a calculator that iterates a couple of hundred object and produces Nx1 arrays for each of those objects. N here being 1-10m depending on configurations. Right now I am summing over these by using a generator expression, so memory consumption is low. However, I would like to store the Nx1 arrays to file, so I can do other computations.(Compute quantiles, partial sums etc. pandas style) Preferably I would like to use pa.memory_map on a single file (in order to have dataframes not loaded into memory), but I can not see how I can produce such a file without generating the entire result first. (Monte Carlo results on 200-500*10m floats).
If I understand correctly RecordBatchStreamWriter needs a part of the entire table, and I can not produce only a part of it. The parts the calculator produces is the columns, not parts of all columns. Is there any way of writing "columns" one by one? Either by appending, or create an empty arrow file which can be filled? (schema known).
As I see it, my alternative is to write several files and use "dataset" /tabular data to "join" them together. My "other computations" would then have to filter or pull parts into memory as I can`t see in the docs that "dataset()" work with memory_map.The result set is to big to fit in memory. (At least on the server it is running on)
I`m on day 2 of digging the docs and trying to understand how it all works, so apologies if the "lingo" is not all correct.
On further inspection, it looks like all files used in datasets() must have same schema, so I can not split "columns" in separate files either, can I..
EDIT
After wrestling with this library, I now produce single column files which I later combine in a single file. However, in following the suggested solution visible memory consumption (task manager) skyrockets in the step of combining the files. I would expect peaks for every "rowgroup" or combined recordbatch, but instead steadily increase to use all memory. A snip of this step:
readers = [pa.ipc.open_stream(file) for file in self.tempfiles]
combined_schema = pa.unify_schemas([r.schema for r in readers])
with pa.ipc.new_stream(
os.path.join(self.filepath, self.outfile_name + ".arrow"),
schema=combined_schema,
) as writer:
for group in zip(*readers):
combined_batch = pa.RecordBatch.from_arrays(
[g.column(0) for g in group], names=combined_schema.names
)
writer.write_batch(combined_batch)
From this link I would expect that running memory consumption to be that of combined_batch and some.
You could do the write in two passes.
First, write each column to its own file. Make sure to set a row group size small enough that a table consisting of one row group from each file comfortably fits into memory.
Second, create a streaming reader for each file you created and one writer. Read a single row group from each one. Create a table by combining all of the partial columns and write the table to your writer. Repeat until you've exhausted all your readers.
I'm not sure that memory mapping is going to help you much here.
Related
I want to be able to do two things:
Store a hash of a datasets contents (so I can decide whether it has updated). To date, I have done this via a second output dataset with a single row that stores the hash and row count. In my Transform I can read that output and compare it to the current build's hash and row count to decide if data has updated. This works fine, but I'd like to avoid having a second dataset if possible.
Pass through timestamps from upstream dependencies so that in downstream workflows I can answer "when did dependency X last update?"
It seems like both of these could be solved by some sort of key-value metadata store on the dataset.
You're correct that one of the most straightforward ways to do this is to decorate the rows with a timestamp value, and in fact with Foundry's Parquet storage system, this will be encoded using Dictionary Encoding, a highly efficient mechanism to store repeated values.
The problem with this approach is you'll have to stack a new column for each phase of updating you want to keep track of. This might prove annoying to maintain in practice.
However, if you don't want to add this data to your rows and instead simply want to store your metadata, you have two options, one of which you've already found:
Store metadata in a separate dataset
Write an 'unused' file (probably .csv or .txt) to your output keeping track of this information
Foundry won't consider your .csv or .txt extra file on the output if you're writing a standard DataFrame to it since your schema by default will only read Parquet files. This means you can store this little snippet of information without affecting your output. If you check platform documentation, you can confirm that it's possible to write both a DataFrame to an output and a file of your own.
It may be simpler to interact with a second output however since the mechanisms of Incremental Transforms and schema handling will be taken care of for you, so I'd recommend proceeding with 1. as you are right now.
I have a Python program that is controlling some machines and stores some data. The data is produced at a rate of about 20 rows per second (and about 10 columns or so). The whole run of this program can be as long as one week, as a result there is a large dataframe.
What are safe and correct ways to store this data? With safe I mean that if something fails in the day 6, I will still have all the data from days 1→6. With correct I mean not re-writing the whole dataframe to a file in each loop.
My current solution is a CSV file, I just print each row manually. This solution is both safe and correct, but the problem is that CSV does not preserve data types and also occupies more memory. So I would like to know if there is a binary solution. I like the feather format as it is really fast, but it does not allow to append rows.
I can think of two easy options:
store chunks of data (e.g. every 30 seconds or whatever suits your use case) into separate files; you can then postprocess them back into a single dataframe.
store each row into an SQL database as it comes in. Sqlite will likely be a good start, but I'd maybe really go for PostgreSQL. That's what databases are meant for, after all.
I am currently using Python Record Linkage Toolkit to perform deduplication on data sets at work. In an ideal world, I would just use blocking or sortedneighborhood to trim down the size of the index of record pairs, but sometimes I need to do a full index on a data set with over 75k records, which results in a couple billion records pairs.
The issue I'm running into is that the workstation I'm able to use is running out of memory, so it can't store the full 2.5-3 billion pair multi-index. I know the documentation has ideas for doing record linkage with two large data sets using numpy split, which is simple enough for my usage, but doesn't provide anything for deduplication within a single dataframe. I actually incorporated this subset suggestion into a method for splitting the multiindex into subsets and running those, but it doesn't get around the issue of the .index() call seemingly loading the entire multiindex into memory and causing an out of memory error.
Is there a way to split a dataframe and compute the matched pairs iteratively so I don't have to load the whole kit and kaboodle into memory at once? I was looking at dask, but I'm still pretty green on the whole python thing, so I don't know how to incorporate the dask dataframes into the record linkage toolkit.
While I was able to solve this, sort of, I am going to leave it open because I suspect given my inexperience with python, my process could be improved.
Basically, I had to ditch the index function from record linkage toolkit. I pulled out the Index of the dataframe I was using, and then converted it to a list, and passed it through the itertools combinations function.
candidates = fl
candidates = candidates.index
candidates = candidates.tolist()
candidates = combinations(candidates,2)
This then gave me an iteration object full of tuples, without having to load everything in to memory. I then passed it into an islice grouper as a for loop.
for x in iter(lambda: list(islice(candidates,1000000)),[]):
I then proceeded to perform all of the necessary comparisons in the for loop, and added the resultant dataframe to a dictionary, which I then concatenate at the end for the full list. Python's memory usage hasn't risen above 3GB the entire time.
I would still love some information on how to incorporate dask into this, so I will accept any answer that can provide that (unless the mods think I should open a new question).
After the recommendation from Jeff's Answer to check out this Google Forum, I still didn't feel satisfied on what the conclusion was regarding the appendCSV method. Below, you can see my implementation of reading many XLS files. Is there a way to significantly increase the speed of this? It currently takes over 10 minutes for around 900,000 rows.
listOfFiles = glob.glob(file_location)
frame = pd.DataFrame()
for idx, a_file in enumerate(listOfFiles):
data = pd.read_excel(a_file, sheetname=0, skiprows=range(1,2), header=1)
data.rename(columns={'Alphabeta':'AlphaBeta'}, inplace=True)
frame = frame.append(data)
# Save to CSV..
frame.to_csv(output_dir, index=False, encoding='utf-8', date_format="%Y-%m-%d")
The very first important point
Optimize only code that is required to be optimized.
If you need to convert all you files just once then you have already made a great job, congrats! If you, however, need to reuse it really often (and by really I mean that there is a source that produce your Excel files with a speed at least of 900K rows per 10 minutes and you need to parse them in real-time) then what you need to do is to analyze your profiling results.
Profiling analysis
Sorting your profile in descending order by 'cumtime', which is cumulative execution time of function including its subcalls, you will discover that out of ~2000 seconds of runtime ~800 seconds are taken by 'read_excel' method and ~1200 seconds are taken by 'to_csv' method.
If then you will sort profile by 'tottime' which is total execution time of functions themselves you will find out that top time consumers are populated with functions that are connected with reading and writing lines and conversion between formats. So, the real problem is that either libraries you use are slow, or the amount of data you are parsing is really huge.
Possible solutions
For the first reason, please keep in mind that parsing Excel lines and converting them could be a really complex task. It is hard to advice you without having an example of your input data. But there could be a real time loss just because the library you are using is for everything and it does hard work parsing rows several times when you actually do not need it, because your rows have very simple structure. In this case you may try to switch to different libraries, that does not perform complex parsing of input data, for example use xlrd for reading data from Excel. But in title you mentioned that input files are also CSVs so if this is applicable in your case then load lines with just:
line.strip().split(sep)
instead of complex Excel format parsing. And of course if your rows are simple than you can always use
','.join(list_of_rows)
to write CSV instead of using complex DataFrames at all. However, if your files contain Unicode symbols, complex fields and so on then these libraries are probably the best choice.
For the second reason - 900K rows could contain from 900K to infinite bytes, so it is really hard to understand whether your data input is really so big, without an example again. If you have really a lot of data then probably there is not too much you could do and you just have to wait. And remember that disk is actually a very slow device. Usual disks could provide you with ~100Mb/s at its best so if you are copying (because ultimately that is what you are doing) 10Gb of data then you can see that at least 3-4 minutes will be required for just physically reading raw data and writing the result. But in case if you are not using your disk bandwidth for 100% (for example if parsing one row with library that you are using takes comparable time with just reading this row from disk) you might also try to increase speed of your code by asynchronous data reading with multiprocessing map_async instead of cycle.
If you are using pandas, you could do this:
dfs = [pd.read_excel(path.join(dir, name), sep='\t', encoding='cp1252', error_bad_lines=False ) for name in os.listdir(dir) if name.endswith(suffix)]
df = pd.concat(dfs, axis=0, ignore_index=True)
This is screaming fast compared to other methods of getting data into pandas. Other tips:
You can also speed this up by specifying dtype for all columns.
If you are doing read_csv, use the engine='c' to speed up the import.
Skip rows on error
I'm currently rewriting some python code to make it more efficient and I have a question about saving python arrays so that they can be re-used / manipulated later.
I have a large number of data, saved in CSV files. Each file contains time-stamped values of the data that I am interested in and I have reached the point where I have to deal with tens of millions of data points. The data has got so large now that the processing time is excessive and inefficient---the way the current code is written the entire data set has to be reprocessed every time some new data is added.
What I want to do is this:
Read in all of the existing data to python arrays
Save the variable arrays to some kind of database/file
Then, the next time more data is added I load my database, append the new data, and resave it. This way only a small number of data need to be processed at any one time.
I would like the saved data to be accessible to further python scripts but also to be fairly "human readable" so that it can be handled in programs like OriginPro or perhaps even Excel.
My question is: whats the best format to save the data in? HDF5 seems like it might have all the features I need---but would something like SQLite make more sense?
EDIT: My data is single dimensional. I essentially have 30 arrays which are (millions, 1) in size. If it wasn't for the fact that there are so many points then CSV would be an ideal format! I am unlikely to want to do lookups of single entries---more likely is that I might want to plot small subsets of data (eg the last 100 hours, or the last 1000 hours, etc).
HDF5 is an excellent choice! It has a nice interface, is widely used (in the scientific community at least), many programs have support for it (matlab for example), there are libraries for C,C++,fortran,python,... It has a complete toolset to display the contents of a HDF5 file. If you later want to do complex MPI calculation on your data, HDF5 has support for concurrently read/writes. It's very well suited to handle very large datasets.
Maybe you could use some kind of key-value database like Redis, Berkeley DB, MongoDB... But it would be nice some more info about the schema you would be using.
EDITED
If you choose Redis for example, you can index very long lists:
The max length of a list is 232 - 1 elements (4294967295, more than 4
billion of elements per list). The main features of Redis Lists from
the point of view of time complexity are the support for constant time
insertion and deletion of elements near the head and tail, even with
many millions of inserted items. Accessing elements is very fast near
the extremes of the list but is slow if you try accessing the middle
of a very big list, as it is an O(N) operation.
I would use a single file with fixed record length for this usecase. No specialised DB solution (seems overkill to me in that case), just plain old struct (see the documentation for struct.py) and read()/write() on a file. If you have just millions of entries, everything should be working nicely in a single file of some dozens or hundreds of MB size (which is hardly too large for any file system). You also have random access to subsets in case you will need that later.