[HUDI-1906] Deltastreamer/SparkDatasource ingestion breaks when changing target schema provider options

• Type: Improvement
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Component/s: deltastreamer
• Labels: schema-evolution sev:normal
• Assignee: sivabalan narayanan
• Reporter: Nishith Agarwal
• Created: 15/May/21 05:07
• Updated: 13/Dec/21 14:14

Description

Currently, there are few different options to the user to provide target schemas such as file based, schema registry. At a high level, there are 2 main flows 

1. Target Schema is provided by the user
2. Target schema is not provided by the user (which is then inferred from the incoming data)

 

Schema post processor enabled	Transformers	User provided target schema	Cur behavior
yes	No	Yes	table schema's has no namespace. matches user provided schema
yes	yes	No	had to make minor fix in post processor for NPE. with the fix, table schema has namespace in it.
yes	yes	yes	table schema has namespace
no	no	yes	table schema's has no namespace. matches user provided schema
no	yes	yes	table schema's has no namespace. matches user provided schema
no	yes	no	table's schema has namespace.

 

Source -> https://github.com/apache/hudi/pull/2937

As you can see above, if one switches from a non-user-provided schema flow to a user-provided-schema flow, we switch from namespace in schema to no namespace in schema. 

Parquet does not store the namespace, so when moving across avro schemas with and without namespace, the parquet-avro writer or reader does not complain since parquet itself does not store namespace. 

However, for MergeOnRead tables, we serialize data and schema in the log blocks. The GenericDatumReader that takes a reader & writer schema to translate breaks when one schema has namespace while the other doesn't. 

 

The following exception is thrown 

51511 [Executor task launch worker for task 502] ERROR org.apache.hudi.common.table.log.AbstractHoodieLogRecordScanner  - Got exception when reading log file
org.apache.avro.AvroTypeException: Found hoodie.source.hoodie_source.height.fixed, expecting fixed
	at org.apache.avro.io.ResolvingDecoder.doAction(ResolvingDecoder.java:292)
	at org.apache.avro.io.parsing.Parser.advance(Parser.java:88)
	at org.apache.avro.io.ValidatingDecoder.checkFixed(ValidatingDecoder.java:135)
	at org.apache.avro.io.ValidatingDecoder.readFixed(ValidatingDecoder.java:146)
	at org.apache.avro.generic.GenericDatumReader.readFixed(GenericDatumReader.java:342)
	at org.apache.avro.generic.GenericDatumReader.readWithoutConversion(GenericDatumReader.java:180)
	at org.apache.avro.generic.GenericDatumReader.read(GenericDatumReader.java:153)
	at org.apache.avro.generic.GenericDatumReader.readField(GenericDatumReader.java:232)
	at org.apache.avro.generic.GenericDatumReader.readRecord(GenericDatumReader.java:222)
	at org.apache.avro.generic.GenericDatumReader.readWithoutConversion(GenericDatumReader.java:175)
	at org.apache.avro.generic.GenericDatumReader.read(GenericDatumReader.java:153)
	at org.apache.avro.generic.GenericDatumReader.read(GenericDatumReader.java:145)
	at org.apache.hudi.common.table.log.block.HoodieAvroDataBlock.deserializeRecords(HoodieAvroDataBlock.java:157)
	at org.apache.hudi.common.table.log.block.HoodieDataBlock.createRecordsFromContentBytes(HoodieDataBlock.java:128)
	at org.apache.hudi.common.table.log.block.HoodieDataBlock.getRecords(HoodieDataBlock.java:106)
	at org.apache.hudi.common.table.log.AbstractHoodieLogRecordScanner.processDataBlock(AbstractHoodieLogRecordScanner.java:275)
	at org.apache.hudi.common.table.log.AbstractHoodieLogRecordScanner.processQueuedBlocksForInstant(AbstractHoodieLogRecordScanner.java:308)
	at org.apache.hudi.common.table.log.AbstractHoodieLogRecordScanner.scan(AbstractHoodieLogRecordScanner.java:241)
	at org.apache.hudi.common.table.log.HoodieMergedLogRecordScanner.<init>(HoodieMergedLogRecordScanner.java:81)
	at org.apache.hudi.HoodieMergeOnReadRDD$.scanLog(HoodieMergeOnReadRDD.scala:259)
	at org.apache.hudi.HoodieMergeOnReadRDD$$anon$2.<init>(HoodieMergeOnReadRDD.scala:164)
	at org.apache.hudi.HoodieMergeOnReadRDD.payloadCombineFileIterator(HoodieMergeOnReadRDD.scala:154)
	at org.apache.hudi.HoodieMergeOnReadRDD.compute(HoodieMergeOnReadRDD.scala:67)
	at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:324)
	at org.apache.spark.rdd.RDD.iterator(RDD.scala:288)
	at org.apache.spark.rdd.MapPartitionsRDD.compute(MapPartitionsRDD.scala:52)
	at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:324)
	at org.apache.spark.rdd.RDD.iterator(RDD.scala:288)
	at org.apache.spark.rdd.MapPartitionsRDD.compute(MapPartitionsRDD.scala:52)
	at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:324)
	at org.apache.spark.rdd.RDD.iterator(RDD.scala:288)
	at org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:99)
	at org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:55)
	at org.apache.spark.scheduler.Task.run(Task.scala:123)
	at org.apache.spark.executor.Executor$TaskRunner$$anonfun$10.apply(Executor.scala:408)
	at org.apache.spark.util.Utils$.tryWithSafeFinally(Utils.scala:1360)
	at org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:414)
	at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149)
	at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
	at java.lang.Thread.run(Thread.java:748

 This seems like an AVRO shortcoming. We need a way to avoid breaking the decoding of avro data in log files if the user moved around provider options. One way is to implement a custom GenericDatumReader.