[jira] [Created] (KAFKA-7662) Avro schema upgrade not supported on globalTable

["Frederic Tardif (JIRA)" <ji...@apache.org>]

Frederic Tardif created KAFKA-7662:
--------------------------------------

             Summary: Avro schema upgrade not supported on globalTable 
                 Key: KAFKA-7662
                 URL: https://issues.apache.org/jira/browse/KAFKA-7662
             Project: Kafka
          Issue Type: Improvement
          Components: streams
    Affects Versions: 1.0.0
            Reporter: Frederic Tardif


I did quite a bit of testing around the avro upgrades, and it did not behave as I would have expected when the avro is used as a Key for a global table with a rocksDB store

setup:
 * local confluent suite 4.0.2
 * test with stream app and producer (v 1.0.0)
 * schemas (key) :

schemas :
{code:java}
schema version @1
{
"namespace": "com.bell.cts.livecms.livemedia.topic",
"type" : "record",
"name" : "EventKey",
"fields" : [
{"name" : "keyAttribute1", "type" : "string"}
]
}
schema version @2
{
"namespace": "com.bell.cts.livecms.livemedia.topic",
"type" : "record",
"name" : "EventKey",
"fields" : [
{"name" : "keyAttribute1", "type" : "string"},
{"name" : "keyAttribute2", "type" : ["null", "string"], "default" : null}
]
}{code}
 
 * TEST1 (PASS)
 ** using schema version @1 
 ** produce record1=[k@1, v@1] 
 ** stream apps loads record1 in global table and store locally in rocksdb 
 ** asyncAssert that store.get(k@1)=v@1 : PASS
 * TEST2 (PASS)
 ** using schema version @1
 ** delete local store (and checkpoint)
 ** stream apps loads record1 in global table and store locally in rocksdb
 ** asyncAssert that store.get(k@1)=v@1 : PASS
 * TEST3 (FAIL)
 ** using schema version @2 
 ** keep local store
 ** stream apps does not reload record1 from topic because of local offset
 ** asyncAssert that store.get(k@1)=v@1 : FAIL
 ** however store.all().next().key.equals(k@2) , as built using schema version 2
 ** this would be explained by the fact that the rocksdb store has some magic byte persisted of the record based on schema version 1
 ** Not ideal, but I could consider accceptable to delete the local store in this cases.
 * TEST4 (FAIL)
 ** using schema version @2
 ** delete local store (and checkpoint)
 ** stream apps loads record1 (produced from schema @1) in global table and store locally in rocksdb
 ** asyncAssert that store.get(k@2)=v@2 : FAIL
 ** however store.all().next().key.equals(k@2) , as built using schema version 2
 ** I can't quite understand this one. I would have expected that the rockdb store should now be provisioned with a serialized version of the record based on the schema v2 (as it went though the stream app underpinning the store materialization)
 * TEST5 (FAIL)
 ** using schema version @2 
 ** produce record2=[k@2, v@2] (meant to be backward compatible and logically equals to record1) 
 ** stream apps does the processing of record1(produced from schema @1) and record2 (produced from schema @2) and materialize the global table stored locally in rocksdb
 ** asyncAssert that store.get(k@2)=v@2 : PASS but the store now has 2 entries !!!
 ** it looks as if the stream.groupBy(key) of the topic underpinning the globaltable materialization did not group the 2 record keys together, although record1.key.equals(record2.key) is true in Java (by looping in the store)

reading from the upstream raw topic throughout the testing :
{code:java}
/tmp$ kafka-avro-console-consumer --topic topic-test-5 --bootstrap-server localhost:9092 --property schema.registry.url=http://127.0.0.1:8081 --property print.key=true --from-beginning 
{"keyAttribute1":"key-attribute-1"}	{"valueAttribute1":"value-1"}
{"keyAttribute1":"key-attribute-1"}	{"valueAttribute1":"value-1"}
{"keyAttribute1":"key-attribute-1"}	{"valueAttribute1":"value-1"}
{"keyAttribute1":"key-attribute-1","keyAttribute2":null}	{"valueAttribute1":"value-1"}{code}



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)