使用 DataFrame / BigQuery 加速 Python 循环

Question

This loop is currently taking almost 3 hours on my desktop running at 5ghz (OC). How would I go about speeding it up?

df = pd.DataFrame(columns=['clientId', 'url', 'count'])

idx = 0
for row in rows:
    df.loc[idx] = pd.Series({'clientId': row.clientId, 'url': row.pagePath, 'count': row.count})
    idx += 1

Rows is JSON data stored in (BigQuery) RowIterator.

<google.cloud.bigquery.table.RowIterator object at 0x000001ADD93E7B50>
<class 'google.cloud.bigquery.table.RowIterator'>

JSON data looks like:

Row(('xxxxxxxxxx.xxxxxxxxxx', '/en-us/index.html', 45), {'clientId': 0, 'pagePath': 1, 'count': 2})
Row(('xxxxxxxxxx.xxxxxxxxxx', '/en-us/contact.html', 65), {'clientId': 0, 'pagePath': 1, 'count': 2})
Row(('xxxxxxxxxx.xxxxxxxxxx', '/en-au/index.html', 64), {'clientId': 0, 'pagePath': 1, 'count': 2})
Row(('xxxxxxxxxx.xxxxxxxxxx', '/en-au/products.html', 56), {'clientId': 0, 'pagePath': 1, 'count': 2})
Row(('xxxxxxxxxx.xxxxxxxxxx', '/en-us/employees.html', 54), {'clientId': 0, 'pagePath': 1, 'count': 2})
Row(('xxxxxxxxxx.xxxxxxxxxx', '/en-us/contact/cookies.html', 44), {'clientId': 0, 'pagePath': 1, 'count': 2})
Row(('xxxxxxxxxx.xxxxxxxxxx', '/en-au/careers.html', 91), {'clientId': 0, 'pagePath': 1, 'count': 2})
Row(('xxxxxxxxxx.xxxxxxxxxx', '/en-ca/careers.html', 42), {'clientId': 0, 'pagePath': 1, 'count': 2})
Row(('xxxxxxxxxx.xxxxxxxxxx', '/en-us/contact.html', 44), {'clientId': 0, 'pagePath': 1, 'count': 2})
Row(('xxxxxxxxxx.xxxxxxxxxx', '/', 115), {'clientId': 0, 'pagePath': 1, 'count': 2})
Row(('xxxxxxxxxx.xxxxxxxxxx', '/suppliers', 51), {'clientId': 0, 'pagePath': 1, 'count': 2})
Row(('xxxxxxxxxx.xxxxxxxxxx', '/en-us/search.html', 60), {'clientId': 0, 'pagePath': 1, 'count': 2})
Row(('xxxxxxxxxx.xxxxxxxxxx', '/en-au/careers.html', 50), {'clientId': 0, 'pagePath': 1, 'count': 2})

Answer 1

This is not how you use the pandas dataframe. The dataframe represents data vertically, meaning each column is a Series under the hood, which uses a fixed-sized numpy array (although columns of same data type have their arrays contiguous to others).

Everytime you append a new row to the dataframe, every column's array is resized (ie, reallocation) and that itself is expensive. You are doing this for every row meaning you have n iterations of array reallocations for each column of a unique datatype and this is extremely inefficient. Furthermore, you are also creating a pd.Series for each row, which incurs more allocations that is not useful when the dataframe represents data vertically.

You can verify this by looking at the id of the columns

>>> import pandas as pd
>>> df = pd.DataFrame(columns=['clientId', 'url', 'count'])

# Look at the ID of the DataFrame and the columns
>>> id(df)
1494628715776

# These are the IDs of the empty Series for each column
>>> id(df['clientId']), id(df['url']), id(df['count'])
(1494628789264, 1494630670400, 1494630670640)

# Assigning a series at an index that didn't exist before
>>> df.loc[0] = pd.Series({'clientId': 123, 'url': 123, 'count': 100})

# ID of the dataframe remains the same
>>> id(df)
1494628715776

# However, the underlying Series objects are different (newly allocated)
>>> id(df['clientId']), id(df['url']), id(df['count'])
(1494630712656, 1494630712176, 1494630712272)

By iteratively adding a new row, you are re-creating new Series objects every iteration, hence why it is slow. This is also warned in the pandas documentation under the .append() method (the argument holds although it is deprecated): https://pandas.pydata.org/docs/reference/api/pandas.DataFrame.append.html#pandas.DataFrame.append

Iteratively appending rows to a DataFrame can be more computationally intensive than a single concatenate. A better solution is to append those rows to a list and then concatenate the list with the original DataFrame all at once.

You'd be better off doing the iterations and appending into a data structure more suited for dynamic-sized operations like the native Python list before calling pd.DataFrame on it. For your simple case, however, you can just pass a generator into the pd.DataFrame call:

# No need to specify columns since you provided the dictionary with the keys
df = pd.DataFrame({'clientId': row.clientId, 'url': row.pagePath, 'count': row.count} for row in rows)

To demonstrate the difference in jupyter notebook:

def reallocating_way(rows):
    df = pd.DataFrame(columns=['clientId', 'url', 'count'])
    for idx, row in enumerate(rows):
        df.loc[idx] = pd.Series({'clientId': row.clientId, 'url': row.pagePath, 'count': row.count})
    return df

def better_way(rows):
    return pd.DataFrame({'clientId': row.clientId, 'url': row.pagePath, 'count': row.count} for row in rows)

# Making an arbitrary list of 1000 rows
rows = [Row() for _ in range(1000)]

%timeit reallocating_way(rows)
%timeit better_way(rows)

2.45 s ± 118 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
1.8 ms ± 112 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

# Making an arbitrary list of 10000 rows
rows = [Row() for _ in range(10000)]

%timeit reallocating_way(rows)
%timeit better_way(rows)

27.3 s ± 1.88 s per loop (mean ± std. dev. of 7 runs, 1 loop each)
12.4 ms ± 142 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

more than 1000x faster for 1000 rows and more than 2000x faster for 10000 rows

Answer 2

I ran across the to_dataframe() method in BigQuery. Extremely fast. Took 3 hours down to 3 seconds.

df = query_job.result().to_dataframe()

google.cloud.bigquery.table.RowIterator

Downloading BigQuery data to pandas using the BigQuery Storage API