GH-40407: [JS] Fix string coercion in MapRowProxyHandler.ownKeys

[mbostock]

Rationale for this change

The ProxyHandler.ownKeys implementation must return strings or symbols. Because of this bug, it was returning numbers, causing the in operator to crash when trying to iterate over the keys of a MapRow object.

An example of this is a DuckDB SQL query using the HISTOGRAM operator:

SELECT HISTOGRAM(phot_g_mean_mag) FROM gaia

What changes are included in this PR?

Instead of calling array.map(String), which returns a typed array of non-strings when array is a typed array, call Array.from which is guaranteed to return strings.

Are these changes tested?

Apologies, but I don’t know how to test this.

Are there any user-facing changes?

This fixes a crash when using the in operator on a MapRow object.

• GitHub Issue: [JS] MapRowProxyHandler.ownKeys can return a typed array #40407

[github-actions]

⚠️ GitHub issue #40407has been automatically assigned in GitHub to PR creator.

[mbostock]

(I’m not fixing this in this PR, but I suspect that there’s also a bug here: if key is a string, and rows[kKeys] is a typed array, then includes(key) will always return false. You’d need to coerce the key to a number, or coerce rows[kKeys] to strings, for this test to return true. Same with other usage of row[kKeys].indexOf(key) below, I expect.)

[trxcllnt]

@mbostock any reason not to fix it in this PR?

[mbostock]

I’m not familiar enough with this codebase to do it confidently and to test. I’d welcome someone else doing it.

[trxcllnt]

LGTM, thanks @mbostock!

[mbostock]

Woot, thanks for the approval @trxcllnt. I’m excited to contribute. 😁

[trxcllnt]

IIRC the Map keys are required to be strings (tho I'm not sure we prevent someone from using a different type). Or has newer Arrow has loosened this requirement? That's the only reason I can think we'd have made the assumption the keys would always be strings here.

[trxcllnt]

I personally prefer supporting maps with keys of any Arrow dtype, but not sure how compatible with other implementations that will be. I do see the integration tests seem to only test maps with string keys.

[domoritz]

Python (which uses C++ which could be considered the canonical implementation) only allows string keys.

Traceback (mostrecentcalllast):
File"/Users/dominik/Code/ramsch/map.py", line8, in<module>table=pa.table({'a': range(10), 'b': np.random.randn(10), 'c': [{x: x} forxinrange(10)]})
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^File"pyarrow/table.pxi", line5204, inpyarrow.lib.tableFile"pyarrow/table.pxi", line1813, inpyarrow.lib._Tabular.from_pydictFile"pyarrow/table.pxi", line5339, inpyarrow.lib._from_pydictFile"pyarrow/array.pxi", line374, inpyarrow.lib.asarrayFile"pyarrow/array.pxi", line344, inpyarrow.lib.arrayFile"pyarrow/array.pxi", line42, inpyarrow.lib._sequence_to_arrayFile"pyarrow/error.pxi", line154, inpyarrow.lib.pyarrow_internal_check_statusFile"pyarrow/error.pxi", line91, inpyarrow.lib.check_statuspyarrow.lib.ArrowTypeError: Expecteddictkeyof type strorbytes, got'int'

Rust doesn't seem to enforce string keys and we couldn't find anything in the spec so it seems to be a gray area right now.

[trxcllnt]

Yeah, I'm inclined to tell DuckDB their Map implementation is non-conformant. Generally whatever libarrow/pyarrow does (and what's in the integration tests) is the source of truth.

Even if JS allows non-string keys, other implementations probably won't. This has been a significant source of confusion in the past, so we try to align with C++ as much as possible.

[trxcllnt]

Changing to "request changes" until we can get some clarification about non-string Map keys. Maybe @pitrou or @kou have thoughts about this?

[CurtHagenlocher]

It is absolutely not the case that maps must have string keys. As just one example, part of the definition of ADBC metadata includes a map<int32, list<int32>>.

[trxcllnt]

You're right, I don't see C++ checking the keys type in MapArray::ValidateChildData(). Maybe the string keys is just a thing pyarrow enforces. Either way, it'd be good to get some integration tests to settle any disagreements.

[pitrou]

This is not how you would create a Map array with PyArrow. Example:

>>>ty=pa.map_(pa.int8(), pa.int32())
>>>a=pa.array([{1: 1000, 2: 10000}, {3: -1000}], type=ty)
>>>a<pyarrow.lib.MapArrayobjectat0x7f7ab948dd80>
[
keys:
[
1,
2
]
values:
[
1000,
10000
],
keys:
[
3
]
values:
[
-1000
]
]

[domoritz]

Oh, pa.table doesn't recognize the dictionary. Thanks for the corrected code!

[mbostock]

Anything else I can do to help this land? Seems like we still want this fix.

[mbostock]

Apologies for bumping, but I’d love to help move this forward as this is currently preventing us from using DuckDB’s HISTOGRAM function (e.g.SELECT HISTOGRAM(phot_g_mean_mag) crashes per observablehq/framework#1014). Please let me know if there is anything I can do.

[conbench-apache-arrow]

After merging your PR, Conbench analyzed the 7 benchmarking runs that have been run so far on merge-commit 117460b.

There were no benchmark performance regressions. 🎉

The full Conbench report has more details. It also includes information about 1 possible false positive for unstable benchmarks that are known to sometimes produce them.