Skip to main content

Data Subjugation library for row-major organization of tabular data and control over the Excel Application

Project description

For example usage, see:

https://github.com/michael-ross-ven/vengeance_example/blob/main/vengeance_example/flux_example.py

https://github.com/michael-ross-ven/vengeance_example/blob/main/vengeance_example/excel_example.py

Managing data stored as rows and columns shouldn't be complicated.

When given a list of lists in Python, your first instinct is to loop over rows and modify column values, in-place. It's the most natural way to think about the data, because conceptually, each row is some entity, and each column is a property of that row, much like a list of objects.

A headache when dealing with list of lists however, is having to keep track of columns by integer elements; it would be nice to replace the indices on each row with named attributes, and have these applied even when the columns are not known ahead of time, such as when pulling data from a sql table or csv file.

for row in matrix:
    row[17]            # what's in that 18th column again?

for row in matrix:
    row.customer_id    # oh, duh

Doesn't the pandas DataFrame already already solve this?

In a DataFrame, data is taken out of its default nested list format and is organized in column-major order, which comes with some advantages as well as drawbacks.

Row-Major Order:
    [['attribute_a', 'attribute_b', 'attribute_c'],
     ['a',           'b',           3.0],
     ['a',           'b',           3.0],
     ['a',           'b',           3.0]]
Column-Major Order:
    {'attribute_a': array(['a', 'a', 'a'], dtype='<U1'),
     'attribute_b': array(['b', 'b', 'b'], dtype='<U1'),
     'attribute_c': array([3.,   3.,  3.], dtype=float64)}

In column-major order, values in a single column are usually all of the same datatype, which means they can be packed into consecutive addresses in memory as an actual array and iterated extremely quickly. But this speed comes at a cost: re-organizing the data as it's intuitively understood by humans, where each row is some entity, and each column is a property of that row, is agonizingly slow. (DataFrame.iterrows() and DataFrame.apply() incur a huge performance penalty, and can be 1_000x times slower to iterate than Python's built-in list.)

DataFrames are also intended to make heavy use of 'vectorization', where operations can be broadcast and applied to an entire set of values in parallel, performed as SIMD instructions at the microprocessor level. But again, this performance optimization comes at a cost: the restricted use of explicit loops over a DataFrame requires pandas to provide specialized API methods for almost every operation and modification, which all must be memorized by the developer. This often leads to convoluted syntax that's counter-inituitive to write, and effortful to read, especially when method-chaining is overused.

# wait, what exactly does this do again?
df['column'] = np.sign(df.column.diff().fillna(0)).shift(-1).fillna(0) \
                 .apply(lambda x: (x['column'].head(1),
                                   x.shape[0],
                                   x['start'].iloc[-1] - x['start'].iloc[0]))
(see also 'So You Wanna Be a Pandas Expert? - James Powell' for how impenetrable this syntax can really get, espcially by less experienced developers)

DataFrame Advantages:
  • vectorized operations on contiguous arrays are memory-efficient and very fast
DataFrame Disadvantages:
  • syntax doesnt always drive intuition or conceptual understanding
  • iteration by rows is effectively out of the question
    (and makes working with JSON format notoriously difficult)
  • vectorized operations are harder to debug / inspect when they encounter an error
  • unexpected loss of precision on numerical data
But I mean, why are we working in Python to begin with?
  • emphasis on code readability
  • datatypes are abstracted away
  • hyper-optimized execution times are less of a priority
So does the DataFrame really reinforce what makes Python so great?

"Explicit is better than implicit"
"Sparse is better than dense"
"Readability counts"
"There should be one– and preferably only one –obvious way to do it"


vengeance.flux_cls

  • similar idea behind a pandas DataFrame, but is more closely aligned with Python's design philosophy
  • when you're willing to trade for a little bit of speed for a lot simplicity
  • a pure-python, row-major wrapper class for list of list data
  • applies named attributes to rows -- attribute values are mutable during iteration
  • provides convenience aggregate operations (sort, filter, groupby, etc)
  • excellent for extremely fast prototyping and data subjugation
row-major iteration
# organized like csv data, attribute names are provided in first row
matrix = [['attribute_a', 'attribute_b', 'attribute_c'],
          ['a',           'b',           3.0],
          ['a',           'b',           3.0],
          ['a',           'b',           3.0]]
flux = vengeance.flux_cls(matrix)

# row attributes can be accessed by name or by sequential index
for row in flux:
    a = row.attribute_a
    a = row['attribute_a']
    a = row[-1]
    a = row.values[:-2]

    row.attribute_a    = None
    row['attribute_a'] = None
    row[-1]            = None
    row.values[:2]     = [None, None]

# transformations are compositional and self-documenting
for row in flux:
    row.hypotenuse = math.sqrt(row.side_a**2 +,
                               row.side_b**2)

matrix = list(flux.values())
columns
column = flux['attribute_a']

flux.rename_columns({'attribute_a': 'renamed_a',
                     'attribute_b': 'renamed_b'})
flux.insert_columns((0, 'inserted_a'),
                    (2, 'inserted_b'))
flux.delete_columns('inserted_a',
                    'inserted_b')
rows
rows = [['c', 'd', 4.0],
        ['c', 'd', 4.0],
        ['c', 'd', 4.0]]

flux.append_rows(rows)
flux.insert_rows(5, rows)

flux_c = flux_a + flux_b
sort / filter / apply
flux.sort('attribute_c')
flux.filter(lambda row: row.attribute_b != 'c')
u = flux.unique('attribute_a', 'attribute_b')

# apply functions like you'd normally do in Python: with comprehensions
flux['attribute_new'] = [some_function(v) for v in flux['attribute_a']]
group / map rows
matrix = [['year', 'month', 'random_float'],
          ['2000', '01',     random.uniform(0, 9)],
          ['2000', '02',     random.uniform(0, 9)],
          ['2001', '01',     random.uniform(0, 9)],
          ['2001', '01',     random.uniform(0, 9)],
          ['2001', '01',     random.uniform(0, 9)],
          ['2002', '01',     random.uniform(0, 9)]]
flux = vengeance.flux_cls(matrix)

dict_1   = flux.map_rows_append('year', 'month')
countifs = {k: len(rows) for k, rows in dict_1.items()}
sumifs   = {k: sum(row.random_float for row in rows)
                                    for k, rows in dict_1.items()}

dict_2 = flux.map_rows_nested('year', 'month')
rows_1 = dict_1[('2001', '01')]
rows_2 = dict_2['2001']['01']
read / write files
flux.to_csv('file.csv')
flux = flux_cls.from_csv('file.csv')

flux.to_json('file.json')
flux = flux_cls.from_json('file.json')

flux.to_file('file.pickle')
flux = flux_cls.from_file('file.pickle')

vengeance.lev_cls

  • (description coming soon...)

Project details


Release history Release notifications | RSS feed

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

vengeance-1.1.38.tar.gz (72.7 kB view hashes)

Uploaded Source

Built Distribution

vengeance-1.1.38-py3-none-any.whl (80.1 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page