Skip to content

crflynn/pbspark

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

39 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

pbspark

This package provides a way to convert protobuf messages into pyspark dataframes and vice versa using pyspark udfs.

Installation

To install:

pip install pbspark

Usage

Suppose we have a pyspark DataFrame which contains a column value which has protobuf encoded messages of our SimpleMessage:

syntax = "proto3";

package example;

message SimpleMessage {
  string name = 1;
  int64 quantity = 2;
  float measure = 3;
}

Basic conversion functions

There are two functions for operating on columns, to_protobuf and from_protobuf. These operations convert to/from an encoded protobuf column to a column of a struct representing the inferred message structure. MessageConverter instances (discussed below) can optionally be passed to these functions.

from pyspark.sql.session import SparkSession
from example.example_pb2 import SimpleMessage
from pbspark import from_protobuf
from pbspark import to_protobuf

spark = SparkSession.builder.getOrCreate()

example = SimpleMessage(name="hello", quantity=5, measure=12.3)
data = [{"value": example.SerializeToString()}]
df_encoded = spark.createDataFrame(data)

df_decoded = df_encoded.select(from_protobuf(df_encoded.value, SimpleMessage).alias("value"))
df_expanded = df_decoded.select("value.*")
df_expanded.show()

# +-----+--------+-------+
# | name|quantity|measure|
# +-----+--------+-------+
# |hello|       5|   12.3|
# +-----+--------+-------+

df_reencoded = df_decoded.select(to_protobuf(df_decoded.value, SimpleMessage).alias("value"))

There are two helper functions, df_to_protobuf and df_from_protobuf for use on dataframes. They have a kwarg expanded, which will also take care of expanding/contracting the data between the single value column used in these examples and a dataframe which contains a column for each message field. MessageConverter instances (discussed below) can optionally be passed to these functions.

from pyspark.sql.session import SparkSession
from example.example_pb2 import SimpleMessage
from pbspark import df_from_protobuf
from pbspark import df_to_protobuf

spark = SparkSession.builder.getOrCreate()

example = SimpleMessage(name="hello", quantity=5, measure=12.3)
data = [{"value": example.SerializeToString()}]
df_encoded = spark.createDataFrame(data)

# expanded=True will perform a `.select("value.*")` after converting,
# resulting in each protobuf field having its own column
df_expanded = df_from_protobuf(df_encoded, SimpleMessage, expanded=True)
df_expanded.show()

# +-----+--------+-------+
# | name|quantity|measure|
# +-----+--------+-------+
# |hello|       5|   12.3|
# +-----+--------+-------+

# expanded=True will first pack data using `struct([df[c] for c in df.columns])`,
# use this if the passed dataframe is already expanded
df_reencoded = df_to_protobuf(df_expanded, SimpleMessage, expanded=True)

Column conversion using the MessageConverter

The four helper functions above are also available as methods on the MessageConverter class. Using an instance of MessageConverter we can decode the column of encoded messages into a column of spark StructType and then expand the fields.

from pyspark.sql.session import SparkSession
from pbspark import MessageConverter
from example.example_pb2 import SimpleMessage

spark = SparkSession.builder.getOrCreate()

example = SimpleMessage(name="hello", quantity=5, measure=12.3)
data = [{"value": example.SerializeToString()}]
df_encoded = spark.createDataFrame(data)

mc = MessageConverter()
df_decoded = df_encoded.select(mc.from_protobuf(df_encoded.value, SimpleMessage).alias("value"))
df_expanded = df_decoded.select("value.*")
df_expanded.show()

# +-----+--------+-------+
# | name|quantity|measure|
# +-----+--------+-------+
# |hello|       5|   12.3|
# +-----+--------+-------+

df_expanded.schema
# StructType(List(StructField(name,StringType,true),StructField(quantity,IntegerType,true),StructField(measure,FloatType,true))

We can also re-encode them into protobuf.

df_reencoded = df_decoded.select(mc.to_protobuf(df_decoded.value, SimpleMessage).alias("value"))

For expanded data, we can also encode after packing into a struct column:

from pyspark.sql.functions import struct

df_unexpanded = df_expanded.select(
    struct([df_expanded[c] for c in df_expanded.columns]).alias("value")
)
df_reencoded = df_unexpanded.select(
    mc.to_protobuf(df_unexpanded.value, SimpleMessage).alias("value")
)

Conversion details

Internally, pbspark uses protobuf's MessageToDict, which deserializes everything into JSON compatible objects by default. The exceptions are

  • protobuf's bytes type, which MessageToDict would decode to a base64-encoded string; pbspark will decode any bytes fields directly to a spark BinaryType.
  • protobuf's well known type, Timestamp type, which MessageToDict would decode to a string; pbspark will decode any Timestamp messages directly to a spark TimestampType (via python datetime objects).
  • protobuf's int64 types, which MessageToDict would decode to a string for compatibility reasons; pbspark will decode these to LongType.

Custom conversion of message types

Custom serde is also supported. Suppose we use our NestedMessage from the repository's example and we want to serialize the key and value together into a single string.

message NestedMessage {
  string key = 1;
  string value = 2;
}

We can create and register a custom serializer with the MessageConverter.

from pbspark import MessageConverter
from example.example_pb2 import ExampleMessage
from example.example_pb2 import NestedMessage
from pyspark.sql.types import StringType

mc = MessageConverter()

# register a custom serializer
# this will serialize the NestedMessages into a string rather than a
# struct with `key` and `value` fields
encode_nested = lambda message:  message.key + ":" + message.value

mc.register_serializer(NestedMessage, encode_nested, StringType())

# ...

from pyspark.sql.session import SparkSession
from pyspark import SparkContext
from pyspark.serializers import CloudPickleSerializer

sc = SparkContext(serializer=CloudPickleSerializer())
spark = SparkSession(sc).builder.getOrCreate()

message = ExampleMessage(nested=NestedMessage(key="hello", value="world"))
data = [{"value": message.SerializeToString()}]
df_encoded = spark.createDataFrame(data)

df_decoded = df_encoded.select(mc.from_protobuf(df_encoded.value, ExampleMessage).alias("value"))
# rather than a struct the value of `nested` is a string
df_decoded.select("value.nested").show()

# +-----------+
# |     nested|
# +-----------+
# |hello:world|
# +-----------+

How to write conversion functions

More generally, custom serde functions should be written in the following format.

# Encoding takes a message instance and returns the result
# of the custom transformation.
def encode_nested(message: NestedMessage) -> str:
    return message.key + ":" + message.value

# Decoding takes the encoded value, a message instance, and path string
# and populates the fields of the message instance. It returns `None`.
# The path str is used in the protobuf parser to log parse error info.
# Note that the first argument type should match the return type of the
# encoder if using both.
def decode_nested(s: str, message: NestedMessage, path: str):
    key, value = s.split(":")
    message.key = key
    message.value = value

Avoiding PicklingErrors

A seemingly common issue with protobuf and distributed processing is when a PicklingError is encountered when transmitting (pickling) protobuf message types from a main process to a fork. To avoid this, you need to ensure that the fully qualified module name in your protoc-generated python file is the same as the module path from which the message type is imported. In other words, for the example here, the descriptor module passed to the builder is example.example_pb2

# from example/example_pb2.py
_builder.BuildTopDescriptorsAndMessages(DESCRIPTOR, "example.example_pb2", globals())
                                                     ^^^^^^^^^^^^^^^^^^^

And to import the message type we would call the same module path:

from example.example_pb2 import ExampleMessage
     ^^^^^^^^^^^^^^^^^^^

Note that the import module is the same as the one passed to the builder from the protoc-generated python. If these do not match, then you will encounter a PicklingError. From the pickle documentation: pickle can save and restore class instances transparently, however the class definition must be importable and live in the same module as when the object was stored.

To ensure that the module path is correct, you should run protoc from the relative root path of your proto files. For example, in this project, in the Makefile under the gen command, we call protoc from the project root rather than from within the example directory.

export PROTO_PATH=.

gen:
	poetry run protoc -I $$PROTO_PATH --python_out=$$PROTO_PATH --mypy_out=$$PROTO_PATH --proto_path=$$PROTO_PATH $$PROTO_PATH/example/*.proto

Known issues

RecursionError when using self-referencing protobuf messages. Spark schemas do not allow for arbitrary depth, so protobuf messages which are circular- or self-referencing will result in infinite recursion errors when inferring the schema. If you have message structures like this you should resort to creating custom conversion functions, which forcibly limit the structural depth when converting these messages.

Development

Ensure that asdf is installed, then run make setup.

  • To format code make fmt
  • To test code make test
  • To run protoc make gen