Guus De Graeve created KAFKA-13505:
--------------------------------------
Summary: Kafka Connect should respect Avro 1.10.X enum defaults
spec
Key: KAFKA-13505
URL: https://issues.apache.org/jira/browse/KAFKA-13505
Project: Kafka
Issue Type: Bug
Components: KafkaConnect
Reporter: Guus De Graeve
We are using Kafka Connect to pipe data from Kafka topics into +parquet files
on+ S3. Our Kafka data is serialised using Avro (schema registry). We use the
Amazon S3 Sink Connector for this.
Up until recently we would set "schema.compatibility" to "NONE" in our
connectors, but this had the pain-full side-effect that during deploys of our
application we got huge file explosions (lots of very small files in HDFS /
S3). This happens because kafka connect will create a new file every time the
schema id of a log changes compared to the previous log. During deploys of our
applications (which can take up to 20 minutes) multiple logs of mixed schema
ids are inevitable and given the huge amounts of logs file explosions of up to
a million files weren't uncommon.
To solve this problem we switched all our connectors "schema.compatibility" to
"BACKWARD", which should only create a new file when a higher schema id is
detected and deserialise all logs with the latest known schema id. Which should
only create one new file during deploys.
An example connector config:
{code:java}
{
"name": "hdfs-Project_Test_Subject",
"config": {
"connector.class": "io.confluent.connect.hdfs.HdfsSinkConnector",
"partition.duration.ms": "86400000",
"topics.dir": "/user/kafka/Project",
"hadoop.conf.dir": "/opt/hadoop/conf",
"flush.size": "1000000",
"schema.compatibility": "BACKWARD",
"topics": "Project_Test_Subject",
"timezone": "UTC",
"hdfs.url": "hdfs://hadoophost:9000",
"value.converter.value.subject.name.strategy":
"io.confluent.kafka.serializers.subject.TopicNameStrategy",
"rotate.interval.ms": "7200000",
"locale": "C",
"hadoop.home": "/opt/hadoop",
"logs.dir": "/user/kafka/_logs",
"format.class": "io.confluent.connect.hdfs.parquet.ParquetFormat",
"partitioner.class":
"io.confluent.connect.storage.partitioner.TimeBasedPartitioner",
"name": "hdfs-Project_Test_Subject",
"errors.tolerance": "all",
"storage.class": "io.confluent.connect.hdfs.storage.HdfsStorage",
"path.format": "YYYY/MM/dd"
}
}{code}
However, we have lots of enum fields in our data records (avro schemas) to
which subjects get added frequently, and this is causing issues with our Kafka
Connect connectors FAILING with these kinds of errors:
{code:java}
Schema parameters not equal. source parameters:
{io.confluent.connect.avro.enum.default.testfield=null,
io.confluent.connect.avro.Enum=Ablo.testfield,
io.confluent.connect.avro.Enum.null=null,
io.confluent.connect.avro.Enum.value1=value1,
io.confluent.connect.avro.Enum.value2=value2} and target parameters:
{io.confluent.connect.avro.enum.default.testfield=null,
io.confluent.connect.avro.Enum=Ablo.testfield,
io.confluent.connect.avro.Enum.null=null,
io.confluent.connect.avro.Enum.value1=value1,
io.confluent.connect.avro.Enum.value2=value2,
io.confluent.connect.avro.Enum.value3=value3}{code}
Since Avro 1.10.X specification, enum values support defaults, which makes
schema evolution possible even when adding subjects (values) to an enum. When
testing our schemas for compatibility using the Schema Registry api we always
get "is_compatible" => true. So schema evolution should in theory not be a
problem.
The error above is thrown in the `SchemaProjector` class which is part of Kafka
Connect, more specifically in the function `checkMaybeCompatible()`. It seems
like this function is not respecting the Avro 1.10.X specification for enum
schema evolution, and I'm not sure if it is meant to respect it? As we
currently don't have any other routes to fix this issue and returning to the
"NONE" schema compatibility is no options considering the file explosions,
we're kinda stuck here.
This issue was discussed more in detail on the Confluent forum in this thread:
[https://forum.confluent.io/t/should-will-kafka-connect-support-schema-evolution-using-avro-1-10-x-enum-defaults/3076/8]
Adem from Confluent is quite confident this is a bug and asked me to file a bug
report here.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
