Jozef, feel free to open a Parquet JIRA to give the issue more
details. Ideally the writer should recover itself and produce the
correct result, but I don't have enough context yet so not sure
whether it's doable.

On Mon, Aug 8, 2022 at 1:53 AM Jozef Vilcek <[email protected]> wrote:
>
> Found it. Problem in my case was not with Parquet but my implementation of
> the `OutputFile` wrapper providing `PositionOutputStream`.
>
> Would it make sense to do changes to the writer to crash on negative
> offsets rather than continue and produce unreadable results.
>
> On Fri, Aug 5, 2022 at 8:42 PM Chao Sun <[email protected]> wrote:
>
> > Seems the file was corrupted during write. There's a similar issue
> > https://issues.apache.org/jira/browse/PARQUET-2164 we found recently.
> >
> > On Fri, Aug 5, 2022 at 3:40 AM Steve Loughran
> > <[email protected]> wrote:
> > >
> > > tha has to be an integer wraparound...something is using a signed int for
> > > position, so when it goes above 2GB it goes negative, and a seek(negative
> > > value) is rejected.
> > >
> > > fix: find the variable and make it a long
> > >
> > >
> > >
> > > On Thu, 4 Aug 2022 at 11:09, Jozef Vilcek <[email protected]> wrote:
> > >
> > > > I came across a case where a job writes out a data set in parquet
> > format
> > > > and it can not be read back as it appears to be corrupted.
> > > >
> > > > Files fail to read back if their size is going over 2GB. If I set the
> > job
> > > > to produce more smaller files from exactly the same input, all is good.
> > > >
> > > > Job write to parquet Avro messages via `parquet-avro` and
> > `parquet-mr`. It
> > > > does happen with v1.10.1 and v1.12.0.
> > > >
> > > > Read error is:
> > > >
> > > > Cannot seek to negative offset
> > > > java.io.EOFException: Cannot seek to negative offset
> > > > at org.apache.hadoop.hdfs.DFSInputStream.seek(DFSInputStream.java:1454)
> > > > at
> > org.apache.hadoop.fs.FSDataInputStream.seek(FSDataInputStream.java:62)
> > > > at
> > > >
> > > >
> > org.apache.parquet.hadoop.util.H2SeekableInputStream.seek(H2SeekableInputStream.java:60)
> > > > at
> > > >
> > > >
> > org.apache.parquet.hadoop.ParquetFileReader$ConsecutiveChunkList.readAll(ParquetFileReader.java:1157)
> > > > at
> > > >
> > > >
> > org.apache.parquet.hadoop.ParquetFileReader.readNextRowGroup(ParquetFileReader.java:805)
> > > >
> > > > When digging a bit into the read, the code materializing
> > > > `ColumnChunkMetaData` is here [1] starting to see negative values for
> > > > `firstDataPage`. Printing some info from `reader.getRowGroups` yields:
> > > >
> > > >
> > > > startingPos=4, totalBytesSize=519551822, rowCount=2300100
> > > > startingPos=108156606, totalBytesSize=517597985, rowCount=2300100
> > > > ...
> > > > startingPos=1950017569, totalBytesSize=511705703, rowCount=2300100
> > > > startingPos=2058233752, totalBytesSize=521762439, rowCount=2300100
> > > > startingPos=-2128348908, totalBytesSize=508570588, rowCount=2300100
> > > > startingPos=-2020294298, totalBytesSize=518901187, rowCount=2300100
> > > > startingPos=-1911848035, totalBytesSize=512724804, rowCount=2300100
> > > > startingPos=-1803573306, totalBytesSize=510980877, rowCount=2300100
> > > > startingPos=-1695543557, totalBytesSize=525871692, rowCount=2300100
> > > > startingPos=-1587016600, totalBytesSize=519353830, rowCount=2300100
> > > > startingPos=-1478696427, totalBytesSize=451032173, rowCount=2090372
> > > >
> > > >
> > > >
> > > > Unfortunately, I was not able to reproduce it locally by taking avro
> > schema
> > > > and generating random inputs and writing them out to a local file.
> > Every
> > > > time, compressed or uncompressed, 3GB file was reading back correctly.
> > > >
> > > > I am looking for help in finding a solution of hints in debugging
> > this, as
> > > > I am out of clues to try to pinpoint and reproduce the problem.
> > > >
> > > > Thanks!
> > > >
> > > > [1]
> > > >
> > > >
> > https://github.com/apache/parquet-mr/blob/apache-parquet-1.12.0/parquet-hadoop/src/main/java/org/apache/parquet/hadoop/metadata/ColumnChunkMetaData.java#L127
> > > >
> >

Reply via email to