How to save a huge pandas dataframe to hdfs?

Question

Im working with pandas and with spark dataframes. The dataframes are always very big (> 20 GB) and the standard spark functions are not sufficient for those sizes. Currently im converting my pandas dataframe to a spark dataframe like this:

dataframe = spark.createDataFrame(pandas_dataframe)  

I do that transformation because with spark writing dataframes to hdfs is very easy:

dataframe.write.parquet(output_uri, mode="overwrite", compression="snappy")

But the transformation is failing for dataframes which are bigger than 2 GB. If I transform a spark dataframe to pandas I can use pyarrow:

// temporary write spark dataframe to hdfs
dataframe.write.parquet(path, mode="overwrite", compression="snappy")

// open hdfs connection using pyarrow (pa)
hdfs = pa.hdfs.connect("default", 0)
// read parquet (pyarrow.parquet (pq))
parquet = pq.ParquetDataset(path_hdfs, filesystem=hdfs)
table = parquet.read(nthreads=4)
// transform table to pandas
pandas = table.to_pandas(nthreads=4)

// delete temp files
hdfs.delete(path, recursive=True)

This is a fast converstion from spark to pandas and it also works for dataframes bigger than 2 GB. I yet could not find a way to do it the other way around. Meaning having a pandas dataframe which I transform to spark with the help of pyarrow. The problem is that I really cant find how to write a pandas dataframe to hdfs.

My pandas version: 0.19.0

Answer 1

From https://issues.apache.org/jira/browse/SPARK-6235

Support for parallelizing R data.frame larger than 2GB

is resolved.

From https://pandas.pydata.org/pandas-docs/stable/r_interface.html

Converting DataFrames into R objects

you can convert a pandas dataframe to an R data.frame

So perhaps the transformation pandas -> R -> Spark -> hdfs?

Answer 2

An hack could be to create N pandas dataframes (each less than 2 GB) (horizontal partitioning) from the big one and create N different spark dataframes, then merging (Union) them to create a final one to write into HDFS. I am assuming that your master machine is powerful but you also have available a cluster in which you are running Spark.

Answer 3

Meaning having a pandas dataframe which I transform to spark with the help of pyarrow.

pyarrow.Table.fromPandas is the function your looking for:

Table.from_pandas(type cls, df, bool timestamps_to_ms=False, Schema schema=None, bool preserve_index=True)

Convert pandas.DataFrame to an Arrow Table

import pyarrow as pa

pdf = ...  # type: pandas.core.frame.DataFrame
adf = pa.Table.from_pandas(pdf)  # type: pyarrow.lib.Table

The result can be written directly to Parquet / HDFS without passing data via Spark:

import pyarrow.parquet as pq

fs  = pa.hdfs.connect()

with fs.open(path, "wb") as fw
    pq.write_table(adf, fw)

See also

• @WesMcKinney answer to read a parquet files from HDFS using PyArrow.
• Reading and Writing the Apache Parquet Format in the pyarrow documentation.
• Native Hadoop file system (HDFS) connectivity in Python

Spark notes:

Furthermore since Spark 2.3 (current master) Arrow is supported directly in createDataFrame (SPARK-20791 - Use Apache Arrow to Improve Spark createDataFrame from Pandas.DataFrame). It uses SparkContext.defaultParallelism to compute number of chunks so you can easily control the size of individual batches.

Finally defaultParallelism can be used to control number of partitions generated using standard _convert_from_pandas, effectively reducing size of the slices to something more manageable.

Unfortunately these are unlikely to resolve your current memory problems. Both depend on parallelize, therefore store all data in memory of the driver node. Switching to Arrow or adjusting configuration can only speedup the process or address block size limitations.

In practice I don't see any reason to switch to Spark here, as long as you use local Pandas DataFrame as the input. The most severe bottleneck in this scenario is driver's network I/O and distributing data won't address that.