[
https://issues.apache.org/jira/browse/PARQUET-2249?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17737487#comment-17737487
]
ASF GitHub Bot commented on PARQUET-2249:
-----------------------------------------
gszadovszky commented on code in PR #196:
URL: https://github.com/apache/parquet-format/pull/196#discussion_r1243234931
##########
src/main/thrift/parquet.thrift:
##########
@@ -966,6 +985,23 @@ struct ColumnIndex {
/** A list containing the number of null values for each page **/
5: optional list<i64> null_counts
+
+ /**
+ * A list of Boolean values to determine pages that contain only NaNs. Only
+ * present for columns of type FLOAT and DOUBLE. If true, all non-null
+ * values in a page are NaN. Writers are suggested to set the corresponding
+ * entries in min_values and max_values to NaN, so that all lists have the
same
+ * length and contain valid values. If false, then either all values in the
+ * page are null or there is at least one non-null non-NaN value in the page.
+ * As readers are supposed to ignore all NaN values in bounds, legacy readers
+ * who do not consider nan_pages yet are still able to use the column index
+ * but are not able to skip only-NaN pages.
+ */
+ 6: optional list<bool> nan_pages
Review Comment:
@mapleFU, it seems to me that NaN is only checked for column indexes at the
write path in parquet-mr. (In this case the column index will be invalid and
won't be written to the file.) For the read path, though, there is no such
check. It means that legacy readers can come to incorrect results using
FLOAT/DOUBLE column indexes after we start writing NaN values. (Sorry for the
late conclusion, I've thought this check was implemented for both directions.)
The only way I can think of for backward compatible NaN handling is to
define a
[ColumnOrder](https://github.com/apache/parquet-format/blob/master/src/main/thrift/parquet.thrift#L863)
for FP values that includes NaNs as well. In case of we would also add support
to row-group level statistics with NaNs. parquet-mr currently skip all kinds of
min/max statistics for columns with not supported column orders.
> Parquet spec (parquet.thrift) is inconsistent w.r.t. ColumnIndex + NaNs
> -----------------------------------------------------------------------
>
> Key: PARQUET-2249
> URL: https://issues.apache.org/jira/browse/PARQUET-2249
> Project: Parquet
> Issue Type: Bug
> Components: parquet-format
> Reporter: Jan Finis
> Priority: Major
>
> Currently, the specification of {{ColumnIndex}} in {{parquet.thrift}} is
> inconsistent, leading to cases where it is impossible to create a parquet
> file that is conforming to the spec.
> The problem is with double/float columns if a page contains only NaN values.
> The spec mentions that NaN values should not be included in min/max bounds,
> so a page consisting of only NaN values has no defined min/max bound. To
> quote the spec:
> {noformat}
> * When writing statistics the following rules should be followed:
> * - NaNs should not be written to min or max statistics
> fields.{noformat}
> However, the comments in the ColumnIndex on the null_pages member states the
> following:
> {noformat}
> struct ColumnIndex {
> /**
> * A list of Boolean values to determine the validity of the corresponding
> * min and max values. If true, a page contains only null values, and
> writers
> * have to set the corresponding entries in min_values and max_values to
> * byte[0], so that all lists have the same length. If false, the
> * corresponding entries in min_values and max_values must be valid.
> */
> 1: required list<bool> null_pages{noformat}
> For a page with only NaNs, we now have a problem. The page definitly does
> *not* only contain null values, so {{null_pages}} should be {{false}} for
> this page. However, in this case the spec requires valid min/max values in
> {{min_values}} and {{max_values}} for this page. As the only value in the
> page is NaN, the only valid min/max value we could enter here is NaN, but as
> mentioned before, NaNs should never be written to min/max values.
> Thus, no writer can currently create a parquet file that conforms to this
> specification as soon as there is a only-NaN column and column indexes are to
> be written.
> I see three possible solutions:
> 1. A page consisting only of NaNs (or a mixture of NaNs and nulls) has it's
> null_pages entry set to {*}true{*}.
> 2. A page consisting of only NaNs (or a mixture of NaNs and nulls) has
> {{byte[0]}} as min/max, even though the null_pages entry is set to
> {*}false{*}.
> 3. A page consisting of only NaNs (or a mixture of NaNs and nulls) does have
> NaN as min & max in the column index.
> None of the solutions is perfect. But I guess solution 3. is the best of
> them. It gives us valid min/max bounds, makes null_pages compatible with
> this, and gives us a way to determine only-Nan pages (min=max=NaN).
> As a general note: I would say that it is a shortcoming that Parquet doesn't
> track NaN counts. E.g., Iceberg does track NaN counts and therefore doesn't
> have this inconsistency. In a future version, NaN counts could be introduced,
> but that doesn't help for backward compatibility, so we do need a solution
> for now.
> Any of the solutions is better than the current situation where engines
> writing such a page cannot write a conforming parquet file and will randomly
> pick any of the solutions.
> Thus, my suggestion would be to update parquet.thrift to use solution 3.
> I.e., rewrite the comments saying that NaNs shouldn't be included in min/max
> bounds by adding a clause stating that "if a page contains only NaNs or a
> mixture of NaNs and NULLs, then NaN should be written as min & max".
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
