emkornfield commented on code in PR #39677:
URL: https://github.com/apache/arrow/pull/39677#discussion_r1458184838
##########
cpp/src/parquet/column_writer.cc:
##########
@@ -323,6 +323,19 @@ class SerializedPageWriter : public PageWriter {
PARQUET_THROW_NOT_OK(sink_->Write(output_data_buffer, output_data_len));
+ // Add the dictionary page to the offsets
+ if (offset_index_builder_ != nullptr) {
+ const int64_t compressed_size = output_data_len + header_size;
+ if (compressed_size > std::numeric_limits<int32_t>::max()) {
+ throw ParquetException("Compressed dictionary page size overflows to
INT32_MAX.");
+ }
+
+ /// start_pos is a relative offset in the buffered mode. It should be
+ /// adjusted via OffsetIndexBuilder::Finish() after BufferedPageWriter
+ /// has flushed all data pages.
+ offset_index_builder_->AddPage(start_pos,
static_cast<int32_t>(compressed_size), -1);
+ }
Review Comment:
We should clarify the spec, but I believe dictionary pages were intended to
be retrieved via column metadata (which would be in the footer). I think it
would have been mentioned if it is, and I think the spec always assumes the
footer is parsed, and for page indexes there is a copy of the metadata there.
The OffsetIndex is meant to directly parallel the ColumnIndex which only keep
stats for data pages.
> The dictionary page, is in fact, a type of data page. So including it
should be backwards compatible since readers should recognize its type in the
list of pages provided by an OffsetIndex.
In the sense that everything is "data" that is correct. But in parquet
dictionaries are
[modeled](https://github.com/apache/parquet-format/blob/master/src/main/thrift/parquet.thrift#L603)
as a separate concept. The actual type of page is not stored in either
index, it is therefore not really possible to make this a compatible change
because old readers would likely error out here. [Parquet-mr does not touch
the index when writing a dictionary
page](https://github.com/apache/parquet-mr/blob/master/parquet-hadoop/src/main/java/org/apache/parquet/hadoop/ParquetFileWriter.java#L1042)
either it does for [data
pages](https://github.com/apache/parquet-mr/blob/master/parquet-hadoop/src/main/java/org/apache/parquet/hadoop/ParquetFileWriter.java#L781).
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
