[python] Convert spark DataFrame column to python list

I work on a dataframe with two column, mvv and count.

+---+-----+
|mvv|count|
+---+-----+
| 1 |  5  |
| 2 |  9  |
| 3 |  3  |
| 4 |  1  |

i would like to obtain two list containing mvv values and count value. Something like

mvv = [1,2,3,4]
count = [5,9,3,1]

So, I tried the following code: The first line should return a python list of row. I wanted to see the first value:

mvv_list = mvv_count_df.select('mvv').collect()
firstvalue = mvv_list[0].getInt(0)

But I get an error message with the second line:

AttributeError: getInt

Let's create the dataframe in question

df_test = spark.createDataFrame(
    [
        (1, 5),
        (2, 9),
        (3, 3),
        (4, 1),
    ],
    ['mvv', 'count']
)
df_test.show()

Which gives

+---+-----+
|mvv|count|
+---+-----+
|  1|    5|
|  2|    9|
|  3|    3|
|  4|    1|
+---+-----+

and then apply rdd.flatMap(f).collect() to get the list

test_list = df_test.select("mvv").rdd.flatMap(list).collect()
print(type(test_list))
print(test_list)

which gives

<type 'list'>
[1, 2, 3, 4]

If you get the error below :

AttributeError: 'list' object has no attribute 'collect'

This code will solve your issues :

mvv_list = mvv_count_df.select('mvv').collect()

mvv_array = [int(i.mvv) for i in mvv_list]

Following one liner gives the list you want.

mvv = mvv_count_df.select("mvv").rdd.flatMap(lambda x: x).collect()

I ran a benchmarking analysis and list(mvv_count_df.select('mvv').toPandas()['mvv']) is the fastest method. I'm very surprised.

I ran the different approaches on 100 thousand / 100 million row datasets using a 5 node i3.xlarge cluster (each node has 30.5 GBs of RAM and 4 cores) with Spark 2.4.5. Data was evenly distributed on 20 snappy compressed Parquet files with a single column.

Here's the benchmarking results (runtimes in seconds):

+-------------------------------------------------------------+---------+-------------+
|                          Code                               | 100,000 | 100,000,000 |
+-------------------------------------------------------------+---------+-------------+
| df.select("col_name").rdd.flatMap(lambda x: x).collect()    |     0.4 | 55.3        |
| list(df.select('col_name').toPandas()['col_name'])          |     0.4 | 17.5        |
| df.select('col_name').rdd.map(lambda row : row[0]).collect()|     0.9 | 69          |
| [row[0] for row in df.select('col_name').collect()]         |     1.0 | OOM         |
| [r[0] for r in mid_df.select('col_name').toLocalIterator()] |     1.2 | *           |
+-------------------------------------------------------------+---------+-------------+

* cancelled after 800 seconds

Golden rules to follow when collecting data on the driver node:

• Try to solve the problem with other approaches. Collecting data to the driver node is expensive, doesn't harness the power of the Spark cluster, and should be avoided whenever possible.
• Collect as few rows as possible. Aggregate, deduplicate, filter, and prune columns before collecting the data. Send as little data to the driver node as you can.

toPandas was significantly improved in Spark 2.3. It's probably not the best approach if you're using a Spark version earlier than 2.3.

See here for more details / benchmarking results.

The following code will help you

mvv_count_df.select('mvv').rdd.map(lambda row : row[0]).collect()

This will give you all the elements as a list.

mvv_list = list(
    mvv_count_df.select('mvv').toPandas()['mvv']
)

On my data I got these benchmarks:

>>> data.select(col).rdd.flatMap(lambda x: x).collect()

0.52 sec

>>> [row[col] for row in data.collect()]

0.271 sec

>>> list(data.select(col).toPandas()[col])

0.427 sec

The result is the same

A possible solution is using the collect_list() function from pyspark.sql.functions. This will aggregate all column values into a pyspark array that is converted into a python list when collected:

mvv_list   = df.select(collect_list("mvv")).collect()[0][0]
count_list = df.select(collect_list("count")).collect()[0][0]