Encoding / formatting issues with python kafka library

Question

I've been trying to use the python kafka library for a bit now and can't get a producer to work.

After a bit of research I've found out that kafka sends (and I'm guessing expects as well) an additional 5 byte header (one 0 byte, one long containing a schema id for schema-registry) to consumers. I've managed to get a consumer working by simply stripping this first bytes.

Am I supposed to prepend a similar header when writing a producer?

Below the exception that comes out:

    [2016-09-14 13:32:48,684] ERROR Task hdfs-sink-0 threw an uncaught and unrecoverable exception (org.apache.kafka.connect.runtime.WorkerTask:142)
org.apache.kafka.connect.errors.DataException: Failed to deserialize data to Avro: 
    at io.confluent.connect.avro.AvroConverter.toConnectData(AvroConverter.java:109)
    at org.apache.kafka.connect.runtime.WorkerSinkTask.convertMessages(WorkerSinkTask.java:357)
    at org.apache.kafka.connect.runtime.WorkerSinkTask.poll(WorkerSinkTask.java:226)
    at org.apache.kafka.connect.runtime.WorkerSinkTask.iteration(WorkerSinkTask.java:170)
    at org.apache.kafka.connect.runtime.WorkerSinkTask.execute(WorkerSinkTask.java:142)
    at org.apache.kafka.connect.runtime.WorkerTask.doRun(WorkerTask.java:140)
    at org.apache.kafka.connect.runtime.WorkerTask.run(WorkerTask.java:175)
    at java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:471)
    at java.util.concurrent.FutureTask.run(FutureTask.java:262)
    at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1145)
    at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:615)
    at java.lang.Thread.run(Thread.java:745)
    Caused by: org.apache.kafka.common.errors.SerializationException: Error deserializing Avro message for id -1
    Caused by: org.apache.kafka.common.errors.SerializationException: Unknown magic byte!

I'm using the latest stable releases of both kafka and python-kafka.

EDIT

Consumer

from kafka import KafkaConsumer
import avro.io
import avro.schema
import io
import requests
import struct

# To consume messages
consumer = KafkaConsumer('hadoop_00',
                         group_id='my_group',
                         bootstrap_servers=['hadoop-master:9092'])

schema_path = "resources/f1.avsc"
for msg in consumer:
    value = bytearray(msg.value)
    schema_id = struct.unpack(">L", value[1:5])[0]
    response = requests.get("http://hadoop-master:8081/schemas/ids/" + str(schema_id))
    schema = response.json()["schema"]
    schema = avro.schema.parse(schema)
    bytes_reader = io.BytesIO(value[5:])
    # bytes_reader = io.BytesIO(msg.value)
    decoder = avro.io.BinaryDecoder(bytes_reader)
    reader = avro.io.DatumReader(schema)
    temp = reader.read(decoder)
    print(temp)

Producer

from kafka import KafkaProducer
import avro.schema
import io
from avro.io import DatumWriter

producer = KafkaProducer(bootstrap_servers="hadoop-master")

# Kafka topic
topic = "hadoop_00"

# Path to user.avsc avro schema
schema_path = "resources/f1.avsc"
schema = avro.schema.parse(open(schema_path).read())
range = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
for i in range:
    producer.send(topic, b'{"f1":"value_' + str(i))

Answer 1

I'm able to have my python producer sending messages to Kafka-Connect with Schema-Registry:

...
import avro.datafile
import avro.io
import avro.schema
from kafka import KafkaProducer

producer = KafkaProducer(bootstrap_servers='kafka:9092')
with open('schema.avsc') as f:
    schema = avro.schema.Parse(f.read())

def post_message():
    bytes_writer = io.BytesIO()
    # Write the Confluent "Magic Byte"
    bytes_writer.write(bytes([0]))
    # Should get or create the schema version with Schema-Registry
    ...
    schema_version = 1
    bytes_writer.write(
        int.to_bytes(schema_version, 4, byteorder='big'))

    # and then the standard Avro bytes serialization
    writer = avro.io.DatumWriter(schema)
    encoder = avro.io.BinaryEncoder(bytes_writer)
    writer.write({'key': 'value'}, encoder)
    producer.send('topic', value=bytes_writer.getvalue())

Documentation about the "Magic Byte": https://github.com/confluentinc/schema-registry/blob/4.0.x/docs/serializer-formatter.rst

Answer 2

Since you are reading with BinaryDecoder and DatumReader, if you send the data in the reverse(using DatumWriter with the BinaryEncoder as encoder), your messages will be fine, I suppose.

Something like this:

Producer

from kafka import KafkaProducer
import avro.schema
import io
from avro.io import DatumWriter, BinaryEncoder
producer = KafkaProducer(bootstrap_servers="hadoop-master")

# Kafka topic
topic = "hadoop_00"

# Path to user.avsc avro schema
schema_path = "resources/f1.avsc"
schema = avro.schema.parse(open(schema_path).read())
# range is a bad variable name. I changed it here
value_range = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
for i in value_range:
    datum_writer = DatumWriter(schema)
    byte_writer = io.BytesIO()
    datum_encoder = BinaryEncoder(byte_writer)
    datum_writer.write({"f1" : "value_%d" % (i)}, datum_encoder)
    producer.send(topic, byte_writer.getvalue())

The few changes I made are:

• use the DatumWriter and BinaryEncoder
• Instead of a json, I am sending a dictionary in the byte stream(You might have to check your code with a normal dictionary and it might work too; but I am not sure)
• Sending the message to the kafka topic using the byte stream(For me, sometimes it failed and in those cases, I assigned the .getvalue method to a variable and use the variable in producer.send. I don't know the reason for failure but assigning to a variable always worked)

I could not test the code I added. But that's the piece of code I wrote while using avro previously. If it's not working for you, please let me know in the comments. It might be because of my rusty memory. I will update this answer with a working one once I reach my home where I can test the code.