[SPARK-27216][CORE] Upgrade RoaringBitmap to 0.7.45

[LantaoJin]

What changes were proposed in this pull request?

HighlyCompressedMapStatus uses RoaringBitmap to record the empty blocks. But RoaringBitmap couldn't be ser/deser with unsafe KryoSerializer.

How was this patch tested?

Adding UT

 test("kryo serialization with RoaringBitmap") {
valbitmap=newRoaringBitmap
bitmap.add(1787)
valsafeSer=newKryoSerializer(conf).newInstance()
valbitmap2:RoaringBitmap= safeSer.deserialize(safeSer.serialize(bitmap))
assert(bitmap2.equals(bitmap))
conf.set("spark.kryo.unsafe", "true")
valunsafeSer=newKryoSerializer(conf).newInstance()
valbitmap3:RoaringBitmap= unsafeSer.deserialize(unsafeSer.serialize(bitmap))
assert(bitmap3.equals(bitmap)) // this will fail
}

[LantaoJin]

This UT only works after #24156 fixed. Now it's easy to reproduce by replacing conf.set("spark.kryo.unsafe", "true") to conf.set("spark.kyro.unsafe", "true").

[LantaoJin]

Since current RoaringBitmap couldn't be ser/deser correctly in unsafe KryoSerializer, first thing I could think out is replacing this data structure totally or when use unsafe Kryo. How do you think about it?
cc @dongjoon-hyun@vanzin@squito@gatorsmile

[srowen]

Does this need to be serialized? I wouldn't think so if it doesn't work!

[attilapiros]

You are changing the conf which also used by other tests within the suite and now the execution order of these tests are important. If the test execution starts with this test and others are executed latter they might fail.

[LantaoJin]

It can be move to a totally new Suite. I will update it.

[AmplabJenkins]

Can one of the admins verify this patch?

[squito]

@srowen I think this is only necessary with "spark.kryo.unsafe=true" -- it probably never worked with that configuration before, but did work with the default "spark.kryo.unsafe=false"

Err, scratch that, I was looking at entirely the wrong thing. I'm also confused here -- so far, this change is just the failing UT, right? you will add the actual fix to behavior as part of this pr?

[LantaoJin]

@srowen@squito I've added another UT which is the minimized dataset from our product issue.
In this UT, I roughly comment one line in ShuffleBlockFetcherIterator to avoid job fail.

if (buf.size ==0) {
// throwFetchFailedException(blockId, address, new IOException(msg))
}

After that, the testing fail zero-size blocks in ShuffleBlockFetcherIteratorSuite will fail. This was introduced by #21219. So in Spark2.3.x, this UT doesn't need this hard code commenting.

[squito]

@LantaoJin I'm still confused by the status of this -- it seems its just test changes, not behavior changes, but it sounds like you are saying some behavior is just broken. Its labeled as a WIP, but you've also pinged people for review. Are you looking for help in determining the right fix? If so, it would help us if you could give a more complete description of what goes wrong. I don't see anything obviously wrong with unsafe kryo and roaring bitmap -- you could try serializing a tiny bitmap and see if the bits make sense

Or do you believe this by itself is actually the complete change?

[LantaoJin]

How can I reopen this closed ticket? @srowen@srowen Now I have detected the root cause. It's a bug of RoaringBitmap and fixed in latest version. (Above unit tests could illustrate it observably. I renamed this PR and am going to push new commit.

[LantaoJin]

Other users also meet this problem. https://issues.apache.org/jira/browse/SPARK-27216?focusedCommentId=16804447&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-16804447

[LantaoJin]

Since SQLQueryWithKryoSuite is overkill for this PR but useful to illustrate this problem. I will delete it from code and keep it in comment here. After upgraded to latest version, below UT could pass.

packageorg.apache.spark.sqlimportorg.apache.spark.internal.configimportorg.apache.spark.internal.config.Kryo._importorg.apache.spark.internal.config.SERIALIZERimportorg.apache.spark.sql.internal.SQLConfimportorg.apache.spark.sql.test.SharedSQLContextclassSQLQueryWithKryoSuiteextendsQueryTestwithSharedSQLContext {
overrideprotecteddefsparkConf=super.sparkConf
.set(SERIALIZER, "org.apache.spark.serializer.KryoSerializer")
.set(KRYO_USE_UNSAFE, true)
test("kryo unsafe data quality issue") {
// This issue can be reproduced when// 1. Enable KryoSerializer// 2. Set spark.kryo.unsafe to true// 3. Use HighlyCompressedMapStatus since it uses RoaringBitmap// 4. Set spark.sql.shuffle.partitions to 6000, 6000 can trigger issue based the supplied data// 5. Comment the zero-size blocks fetch fail exception in ShuffleBlockFetcherIterator// or this job will failed with FetchFailedException.
withSQLConf(
SQLConf.SHUFFLE_PARTITIONS.key ->"6000",
config.SHUFFLE_MIN_NUM_PARTS_TO_HIGHLY_COMPRESS.key ->"-1") {
withTempView("t") {
valdf= spark.read.parquet(testFile("test-data/dates.parquet")).toDF("date")
df.createOrReplaceTempView("t")
checkAnswer(
sql("SELECT COUNT(*) FROM t"),
sql(
""" |SELECT SUM(a) FROM |( |SELECT COUNT(*) a, date |FROM t |GROUP BY date |)""".stripMargin))
}
}
}
}

[srowen]

@LantaoJin you should be able to reopen this, or it will reopen if you push a new commit.

[LantaoJin]

Sorry I can not reopen it since a force pushing. I open a new #24264 as a updating.

[srowen]

That's fine, I can reopen them too, but you already have a new PR