Automatically exported from
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.
input.csv renamed some functions Jun 20, 2010


Automatically exported from

== What is it? ==

streaming-pickle allows you to save/load a sequence of Python data structures to/from disk in a streaming (incremental) manner, thus using far less memory than [ regular pickle].

== When is it useful? ==

streaming-pickle is useful for any ad-hoc data processing task involving a linear sequence of records.

For example, let's say you write a script A to perform some analysis and then dump 1 million records to disk in some textual format, with each record taking up one line. Then you write another script B that reads in those records one line at a time and performs some more analysis. This strategy is memory-efficient (you only need to store one record at a time in RAM) and provides the benefits of incremental stream processing. However, you need to write the parsing and unparsing code to convert between your plaintext format and Python data structures, which is tedious and error-prone.

An alternative is to have script A create a Python list and pickle it to disk. Then script B simply unpickles the list and iterates over it to perform its analysis. Now you don't need to write any parsing and unparsing code, but unfortunately your scripts might consume far too much memory (they need to store the entire 1 million element list in RAM) and cannot be streamed in a pipeline.

streaming-pickle combines the best of both strategies: memory-efficient stream processing of persistent data without requiring any extra parsing/unparsing code.

== How do I use it? ==

streaming-pickle has a very similar interface to [ regular pickle]. To get start it, simply download the single source file and import it into your project.

=== Basic usage ===

To save a list of Python data to disk, use s_dump:

{{{ lst = ... big list of data you want to save to disk ... sPickle.s_dump(lst, open('lst.spkl', 'w')) }}}

To load data from disk in a streaming manner, use s_load and iterate over its result:

{{{ for element in sPickle.s_load(open('lst.spkl')): ... process element ... }}}

As you iterate, only one element will be loaded into memory at a time (so you can process huge lists without running out of memory).

=== Advanced usage ===

s_dump can save any iterable object to disk, not just lists. In the example below, I create a [ generator expression] that reads input.csv, extracts the first field in each line, processes it with process_data(), and saves the sequence of results to disk in the file lst.spkl.

{{{ sPickle.s_dump((process_data(line.split(',')[0]) for line in open('input.csv')), open('lst.spkl', 'w')) }}}

Since we're using a generator expression, only one line from input.csv will be loaded into memory at a time.

Alternatively, if you don't feel comfortable pickling iterables, you can use s_dump_elt to save one element at a time to disk. This code does the same thing as the above example:

{{{ f = open('lst.spkl', 'w')

for line in open('input.csv'): sPickle.s_dump_elt(process_data(line.split(',')[0]), f)

f.close() }}}

Lastly, remember that you can use any file-like object with streaming-pickle, so you can also stream Python data over pipes or network sockets.

=== Gotchas ===

Remember that s_dump saves an iterable object to disk, so if you try to s_dump a dict, it will actually save only its keys because the default iterator for a dict generates its keys. If you want to save key/value pairs, use:

{{{ d = ... dict you want to save to disk ... sPickle.s_dump(d.iteritems(), open('dict.spkl', 'w')) }}}

(However, s_dump_elt can save any picklable object to disk.)

== Contact information ==

Questions? Complaints? Bug reports? Feature requests? Please email Philip Guo (