CODEWITHSUNDEEP
pythonpandasdaskparquetdask-dataframe
Mike
Mike

Reputation: 155

Read group of rows from Parquet file in Python Pandas / Dask?

I have a Pandas dataframe that looks similar to this:

datetime                 data1  data2
2021-01-23 00:00:31.140     a1     a2
2021-01-23 00:00:31.140     b1     b2       
2021-01-23 00:00:31.140     c1     c2
2021-01-23 00:01:29.021     d1     d2
2021-01-23 00:02:10.540     e1     e2
2021-01-23 00:02:10.540     f1     f2

The real dataframe is very large and for each unique timestamp, there are a few thousand rows.

I want to save this dataframe to a Parquet file so that I can quickly read all the rows that have a specific datetime index, without loading the whole file or looping through it. How do I save it correctly in Python and how do I quickly read only the rows for one specific datetime?

After reading, I would like to have a new dataframe that contains all the rows for that specific datetime. For example, I want to read only the rows for datetime "2021-01-23 00:00:31.140" from the Parquet file and receive this dataframe:

datetime                 data1  data2
2021-01-23 00:00:31.140     a1     a2
2021-01-23 00:00:31.140     b1     b2        
2021-01-23 00:00:31.140     c1     c2

I am wondering it it may be first necessary to convert the data for each timestamp into a column, like this, so it can be accessed by reading a column instead of rows?

2021-01-23 00:00:31.140  2021-01-23 00:01:29.021  2021-01-23 00:02:10.540
           ['a1', 'a2']             ['d1', 'd2']             ['e1', 'e2']
           ['b1', 'b2']                      NaN             ['f1', 'f2']
           ['c1', 'c2']                      NaN                      NaN

I appreciate any help, thank you very much in advance!

Upvotes: 6

Views: 2597

Answers (1)

SultanOrazbayev
SultanOrazbayev

Reputation: 16561

One solution is to index your data by time and use dask, here's an example:

import dask
import dask.dataframe as dd

df = dask.datasets.timeseries(
    start='2000-01-01',
    end='2000-01-2',
    freq='1s',
    partition_freq='1h')

df

print(len(df))
# 86400 rows across 24 files/partitions

%%time
df.loc['2000-01-01 03:40'].compute()
# result returned in about 8 ms

Working with a transposed dataframe like you suggest is not optimal, since you will end up with thousands of columns (if not more) that are unique to each file/partition.

So on your data the workflow would look roughly like this:

import io
data = io.StringIO("""
datetime|data1|data2
2021-01-23 00:00:31.140|a1|a2
2021-01-23 00:00:31.140|b1|b2       
2021-01-23 00:00:31.140|c1|c2
2021-01-23 00:01:29.021|d1|d2
2021-01-23 00:02:10.540|e1|e2
2021-01-23 00:02:10.540|f1|f2""")

import pandas as pd
df = pd.read_csv(data, sep='|', parse_dates=['datetime'])

# make sure the date time column was parsed correctly before
# setting it as an index
df = df.set_index('datetime')


import dask.dataframe as dd
ddf = dd.from_pandas(df, npartitions=3)

ddf.to_parquet('test_parquet')
# note this will create a folder with one file per partition

ddf2 = dd.read_parquet('test_parquet')


ddf2.loc['2021-01-23 00:00:31'].compute()

# if you want to use very precise time, first convert it to datetime format
ts_exact = pd.to_datetime('2021-01-23 00:00:31.140')
ddf2.loc[ts_exact].compute()

Upvotes: 3

Related Questions