Streamdf is a lightweight data frame library built on top of the dictionary of numpy array, developed for Kaggle's time-series code competition.

• Fast and robust insertion
  • The insertion of row can be performed with amortized constant time (much faster than np.append)
  • Automatically falls back to the default value when an abnormal value is inserted
• Time-travel
  • Get the past state of the data as a slice of the original dataframe without copying
• Null/empty-safe aggregations
  • Provides a set of aggregation methods that can be safely called when an element has nan or is empty.
• Columnar layout
  • Internal data is stored in a simple columnar format, which is easier to use for analysis than numpy's structured array

import pandas as pd
from streamdf import StreamDf

df = pd.read_csv('test.csv')
sdf = StreamDf.from_pandas(df)

# extend
sdf.extend({
    'x': 1,
    'y': 2
})

assert len(sdf) == len(df) + 1

# access
print(sdf['x'])

# aggregate
sdf.last_value('x')

import numpy as np
from streamdf import StreamDf

sdf = StreamDf.empty({'x': np.int32, 'time': 'datetime64[D]'}, 'time')

sdf.extend({'x': 1, 'time': np.datetime64('2018-01-01')})
sdf.extend({'x': 5, 'time': np.datetime64('2018-02-01')})
sdf.extend({'x': 3, 'time': np.datetime64('2018-02-03')})

assert len(sdf) == 3

# Time travel (zero copy)
sliced = sdf.slice_until(np.datetime64('2018-02-02'))

assert len(sliced) == 2