Fix Hadoop multi-value string null value handling to match native batch

[jtuglu1]

Description

1. Doing some more digging, I found another unfortunate data difference between native batch (on-cluster) and Hadoop batch ingest. Ingesting a multi-value string ["a","b",null] with Hadoop is treated as ["a","b","null"] and in native batch, this correctly ingests to ["a","b",null]. This is difference appears to be a bug in all Druid versions(even latest). While this will not affect the current null handling migration, this will affect the future Hadoop -> native batch ingestion migration that will also need to take place.
2. Hadoop doesn't allow for all-null columns in segments, it simply excludes them from the segment. I've updated the Hadoop job to support running druid.indexer.task.storeEmptyColumns=true, which allows us to store all NULL columns (how native/streaming ingest work today).

Related to:

• preserve Rows.objectToStrings behavior of translating null into "null" inside of lists and arrays #15190.
• Store null columns in the segments #12279.

Release note

Fix Hadoop null value handling to match native batch and allow v10 segment creation.

BREAKING CHANGES

1. Hadoop ingests will now process multi-value string inputs like ["a","b",null] -> ["a","b",null] instead of ["a","b","null"] to match native batch ingestion.
2. Hadoop ingests will now by default keep columns with all NULL values, instead of excluding them from the segment.
3. useStringValueOfNullInLists parameter in RowBasedColumnSelectorFactory.java‎ has been removed.

---

This PR has:

• been self-reviewed.
  • using the concurrency checklist (Remove this item if the PR doesn't have any relation to concurrency.)
• added documentation for new or modified features or behaviors.
• a release note entry in the PR description.
• added Javadocs for most classes and all non-trivial methods. Linked related entities via Javadoc links.
• added or updated version, license, or notice information in licenses.yaml
• added comments explaining the "why" and the intent of the code wherever would not be obvious for an unfamiliar reader.
• added unit tests or modified existing tests to cover new code paths, ensuring the threshold for code coverage is met.
• added integration tests.
• been tested in a test Druid cluster.

[gianm]

Seems like a good change to me. This behavior was reverted in #15190 but the additional changes in this patch look like they would deal with the original problem. Have you run this through a real world test to confirm that everything works as expected?

[jtuglu1]

@gianm another thing I discovered investigating this patch is that Hadoop by default does not create all-null columns in a segment(-Ddruid.indexer.task.storeEmptyColumns=false by default). Native batch in latest version does... . This is the key difference that showed up in the segment diff. #12279. Do you know why this is?

For example, if you were to ingest [null] multi-value string, native batch would see this correctly as null, whereas Hadoop would not even include this column in the segment which causes headaches when switching from Hadoop to native batch.

[gianm]

@gianm another thing I discovered investigating this patch is that Hadoop by default does not create all-null columns in a segment(-Ddruid.indexer.task.storeEmptyColumns=false by default). Native batch in latest version does... . This is the key difference that showed up in the segment diff. #12279. Do you know why this is?

There's a comment in IndexMergerV9 that says Hadoop indexing uses a constructor that doesn't support the storeEmptyColumns configuration yet. In that spot it's hard-coded to false. I suppose it would make more sense for that to be hard-coded to TaskConfig.DEFAULT_STORE_EMPTY_COLUMNS, i.e., true. It would make even more sense for it to respect the actual configuration, which would mean switching to some logic like the TaskToolboxFactory uses:

config.buildV10()
            ? indexMergerV10Factory.create()
            : indexMergerV9Factory.create(
                task.getContextValue(Tasks.STORE_EMPTY_COLUMNS_KEY, config.isStoreEmptyColumns())
            )

Rather than injecting the IndexMergerV9 directly.

[jtuglu1]

@gianm thanks, that's what I thought. I guess I just wanted to make sure there wasn't any critical piece of null merging that was missing/incompatible with Hadoop for both V9/V10.

[clintropolis]

i think this is a good change because the old behavior was wack, but I'm still tracing through to try to determine the actual impacts of this change.

Besides the hadoop impact, which you fix in this PR, this method seems like it will mostly impact callers of Row.getDimension as well as the toGroupKey method of this class since it calls getDimension.

luckily there are a relatively small number of 'production' callers of these methods

Row.getDimension:

Rows.toGroupKey:

which look mostly related to partitioning. I think we need to determine if the null -> 'null' coercion is important for these callers, and if so, do the coercion there. I'm uncertain currently but will keep trying to figure it out.

[jtuglu1]

@clintropolis I can scope this PR to hadoop-only by creating a separate implementation of the Rows.* methods?

[clintropolis]

it feels worth figuring this out since the old code seems quite odd to be doing what it is at this layer, so I want to keep looking. It probably would be fine though if we can't figure it out?

[jtuglu1]

I've summarized the breaking changes in the release notes section.

[jtuglu1]

I've gone through the callers of these methods and provided some extra tests where I can. From what I can tell, things should work.

[jtuglu1]

@clintropolis @gianm Question: do you know why this is null? I had to update this to get it to work properly (actually ingest null-only columns).

[gianm]

@clintropolis @gianm Question: do you know why this is null? I had to update this to get it to work properly (actually ingest null-only columns).

I do not know. It's the DimensionsSpec parameter, but I don't recall what happens when that isn't provided.

[jtuglu1]

@gianm @clintropolis Looks like with this change there's a single byte difference between native/hadoop-outputted segments in handling of this multi-value string: [null, "a", "b"]. With this change, Hadoop properly encodes the dictionary cardinality as 3, whereas native seems to incorrectly encode it as 2. The byte difference is in the dictionary cardinality (hadoop represents as 3, while native represents as 2). It doesn't show up in queries, but I believe difference is in VSizeColumnarMultiInts.