[AURON #2080] Support Hive Parquet table to NativeParquetHiveTableScanExec

[guixiaowen]

Which issue does this PR close?

Closes #2080

Rationale for this change

Switch the Spark execution plan to HiveTableScanExec, and when the table storage type is Parquet, the native execution plan NativeParquetHiveTableScanExec executes.

Design parameters are as follows:
spark.auron.enable.hiveTable and spark.auron.enable.parquetHiveTableScanExec parameters

Eg：
spark.auron.enable.hiveTable default is true （When set to true, this conversion is enabled by default）
spark.auron.enable.parquetHiveTableScanExec default is false（When set to true, this conversion is enabled by default））

What changes are included in this PR?

Are there any user-facing changes?

NO.

How was this patch tested?

UT

[merrily01]

Hi @guixiaowen

Could you take a look and get the CI fixed?

https://github.com/apache/auron/actions/runs/26717696445/job/82051444906?pr=2081

[weiqingy]

Thanks for taking this on — reusing NativeHiveTableScanBase and wiring the scan in through the AuronConvertProvider SPI (mirroring PaimonConvertProvider) is a clean fit. A few questions inline.

[weiqingy]

AuronConvertProvider declares isEnabled(exec: SparkPlan): Boolean, and the dispatcher calls it with the plan (AuronConverters.scala:216 and :307). This override drops the parameter, so it doesn't implement the trait method — the abstract isEnabled(SparkPlan) stays unimplemented and override matches nothing, which the compiler rejects. That looks like the CI break flagged above. PaimonConvertProvider keeps the parameter and branches inside — would matching that signature work here?

override def isEnabled(exec: SparkPlan): Boolean =
  getBooleanConf("spark.auron.enable.hiveTable", defaultValue = true)

[weiqingy]

filesMaxPartitionBytes defaults to 128MB and filesOpenCostInBytes to 4MB, so Math.min here always returns the open-cost (4MB) and every split is capped at it — which would fan a large table out into many tiny partitions. Spark computes the cap as min(maxPartitionBytes, max(openCostInBytes, bytesPerCore)) rather than a straight min of the two. Was min intended here? FilePartition.maxSplitBytes already implements that formula if it helps, though note it's keyed on Seq[PartitionDirectory] rather than the Seq[PartitionedFile] you have here, so it's not a literal drop-in.

[weiqingy]

isSupported returns true for any table whose provider is hive, but convert only builds a native plan when isParquetTable holds and otherwise returns exec unchanged (line 49). So a non-Parquet Hive table (ORC, text) with both flags on would be reported as supported and then handed back un-converted, instead of falling through to the normal conversion path. Would gating isSupported on isParquetTable too keep the two in lockstep?

[weiqingy]

This io.netty relocation was added in #1597 to fix NoClassDefFoundError: io/netty/buffer/Unpooled under shaded-spark — that change bundled netty-buffer/netty-common into the assembly and relocated them so they wouldn't clash. This PR removes the relocation but keeps those bundled deps (pom.xml:48-55), so netty ships at its original io.netty.* coordinates again, which is the state #1597 fixed away. Does dropping it reopen that shaded-spark conflict? Since the relocation only affects the shaded assembly — not the unit tests or the Hive Parquet path here — it'd help to know what in this PR needs it gone, along with the arrow-memory-netty exclusions added to the shims pom.xml.