Re: Reading from AVRO files

Posted by Timo Walther on
URL: http://deprecated-apache-flink-user-mailing-list-archive.369.s1.nabble.com/Reading-from-AVRO-files-tp35850p35852.html

Hi Lorenzo,

as far as I know we don't support Avro's logical times in Flink's
AvroInputFormat yet. E.g. AvroRowDeserializationSchema [1] supports the
1.8.2 version of logical types but might be incompatible with 1.9.2.

Reg 2) Specific record generated with AVRO 1.9.2 plugin:

Could you send us the full stack trace? I think this should actually
work, because specific records are handled as POJOs and those should be
able to also deal with logical type's classes through Kryo.

Reg 3) Generic record

It would be great if we can make this option possible. We could include
it in the next minor release fix.

Sorry, for the bad user experience. But IMHO logical type are still
experiemental in Avro. Maybe 1.9.2 has finally fixed the biggest
shortcomings such that Flink can properly support them as well.

Regards,
Timo

[1]
https://github.com/apache/flink/blob/master/flink-formats/flink-avro/src/main/java/org/apache/flink/formats/avro/AvroRowDeserializationSchema.java


On 10.06.20 15:08, Lorenzo Nicora wrote:

> Hi,
>
> I need to continuously ingest AVRO files as they arrive.
> Files are written by an S3 Sink Kafka Connect but S3 is not the point
> here. I started trying to ingest a static bunch of files from local fs
> first and I am having weird issues with AVRO deserialization.
>
> I have to say, the records contain logical types, timestamps-ms and decimals
>
> To keep it simple, I extracted the AVRO schema from the data files and
> used avro-maven-plugin to generate POJOs
> I tried multiple combinations, all with no luck
>
> 1) Specific record generated with AVRO 1.8.2 plugin
>
> Path in = new Path(sourceBasePath);
> AvroInputFormat<AccountEntries> inputFormat = new AvroInputFormat<>(in,
> AccountEntries.class);
> DataStream<AccountEntries> accountEntries = env
> .readFile(inputFormat, sourceBasePath,
> FileProcessingMode.PROCESS_CONTINUOUSLY, FILE_SCAN_INTERVAL_MS);
>
> *Result*
> java.lang.ClassCastException: java.lang.Long cannot be cast to
> org.joda.time.DateTime
> (IIRC this is a known AVRO 1.8.2 issue)
>
>
> 2) Specific record generated with AVRO 1.9.2 plugin
> Same code as above but AVRO POJOs are generated with AVRO 1.9.2
>
> *Result*
> org.apache.avro.AvroRuntimeException: Unknown datum type java.time.Instant
>
>
> 3) Generic record
> I am getting the Schema from the generated specific record, for
> convenience, but I am not using the generated POJO as record.
> I also followed the suggestions in this Flink blog post
> <https://flink.apache.org/news/2020/04/15/flink-serialization-tuning-vol-1.html#avro-generic>,
> to explicitly specify the TypeInfo with returns(...)
>
> Path in = new Path(config.sourceFileSystemPath);
> Schema schema = AccountEntries.getClassSchema();
> AvroInputFormat<GenericRecord> inputFormat = new AvroInputFormat<>(in,
> GenericRecord.class);
> DataStream<GenericRecord> accountEntries = env
> .readFile(inputFormat, config.sourceFileSystemPath,
> FileProcessingMode.PROCESS_CONTINUOUSLY, FILE_SCAN_INTERVAL_MS)
> .returns(new GenericRecordAvroTypeInfo(schema));
>
>
> *Result*
> The class 'org.apache.avro.generic.GenericRecord' is not instantiable:
> The class is not a proper class. It is either abstract, an interface, or
> a primitive type.
>
> This looks like a bug.
> I raised the ticket <https://issues.apache.org/jira/browse/FLINK-18223>
> and I will try to submit a fix, but still do not solve my problem as I
> am using a managed Flink I cannot update.
> I cannot believe there is no workaround. I do not think I'm trying to do
> anything bizarre. Am I?
>
> Any ideas?
> Am I missing something obvious?
>
> Cheers
> Lorenzo