parquet writer does not encode `null count = 0` correctly

[alamb]

Describe the bug
TLDR is that the Rust parquet writer does not write out null_counts correctly in row group statistics. However, the reader has the same mistake so systems that both write and read with Rust are unaffected

While reviewing #6216 from @Michael-J-Ward (🙏 ) I am pretty sure the arrow-rs parquet writer does not do the correct thing with respect to null statistics

Specifically, when there are no nulls in the data, the writer does not emit a value for null_count in the thrift metadata (it writes the equivalent of None) -- it should instead write the equivalent of Some(0)

This will not cause issues for people using parquet-rs to read and write data as the reader also (incorrectly) reports Some(0) when the thrift metadata has None

To Reproduce
TBD

Expected behavior

• When writing statistics for data without nulls, the parquet-rs writer should write Some(0)
• When reading statistics, the parquet-rs reader should read None

Additional context

[mapleFU]

When reading statistics, the parquet-rs reader should read None

I'm a bit curious here, since like in page-index ( https://github.com/apache/parquet-format/blob/master/src/main/thrift/parquet.thrift#L1088 ) and page header v2 ( https://github.com/apache/parquet-format/blob/master/src/main/thrift/parquet.thrift#L633 ), the null-count is better to have. Should it also Some(0) to represent "nullable type has null count, and it's 0" ?

[alamb]

When reading statistics, the parquet-rs reader should read None

I'm a bit curious here, since like in page-index ( https://github.com/apache/parquet-format/blob/master/src/main/thrift/parquet.thrift#L1088 ) and page header v2 ( https://github.com/apache/parquet-format/blob/master/src/main/thrift/parquet.thrift#L633 ), the null-count is better to have. Should it also Some(0) to represent "nullable type has null count, and it's 0" ?

I think the parquet-rs writer always has the null count when writing statistics so it is always possible to write Some(...) . At the moment, it writes None (aka missing)

I think we should change the writer to write Some(0)

The difference is when reading a parquet file from some other writer that may not have written a value for the null_count field.

The current parquet-rs code will assume a missing null count field means "0" which might not be correct

Previously to #6216 0 or "missing" were not distinguished in the rust parquet reader API so there was no way to communicate this difference. Now that the API supports the difference we can.

I will try and document this as clearly as I can

[mapleFU]

I think the parquet-rs writer always has the null count when writing statistics so it will always write Some(...)

I've checked parquet-java and parquet-c++, all of these seems to write null-count with 0, but I don't know the case of legacy data. Maybe @wgtmac knows?

[alamb]

I agree it is pretty unlikely to find one of these files in practice

[mapleFU]

seems parquet-java distinguishes None and 0. Maybe I can draft a pr in parquet-format to enforcing this or get more info about this...🤔

But otherplace it's checked: https://github.com/apache/parquet-java/blob/d4384d3f2e7703fab6363fd9cd80a001e9561db2/parquet-hadoop/src/main/java/org/apache/parquet/filter2/statisticslevel/StatisticsFilter.java#L128

[alamb]

seems parquet-java distinguishes None and 0. Maybe I can draft a pr in parquet-format to enforcing this or get more info about this...🤔

I am not sure how important it is to enforce that this statistics field is set

Note that all the parquet files written by the rust writer so far will have had their null counts set to None...

[mapleFU]

From my aspects, 0 and None is different, and parquet-c++ / parquet-java always writes the null_count if has any value. Perhaps there're some old file write it to 0, and make filter pruning becomes buggy, lol

Anyway, we all agree that

When writing statistics for data without nulls, the parquet-rs writer should write Some(0)

I'll draft a discuss in mailling list about this tomorrow, a bit late in utc-8 and I have to sleep :-)

[alamb]

I'll draft a discuss in mailling list about this tomorrow, a bit late in utc-8 and I have to sleep :-)

Thank you . Pleasant dreams 💤

[alamb]

From my aspects, 0 and None is different,

Yes, this makes sense to me

and parquet-c++ / parquet-java always writes the null_count if has any value.

Yes, this is what I think the rust writer should do too (I am making a PR to make it so)

[alamb]

I made a PR with a proposed fix: #6257 which turned out to be larger than I expected (though it is mostly tests)