Hi Steve,
Le 26/10/2026 à 23:26, Steve Loughran a écrit :
The slides can be read online at
https://pitrou.net/coc-glasgow-2026/Fuzzing.html
nice!
one question: are there specific classes/modules which seems to crop up
more often than others in the fuzz scans? as these would seem to be the
targets for close examination by humans, maybe consideration for rework.
I don't have statistics on this but a human scan through our git log
suggests the following classes of issues:
- lack of validation of Parquet metadata values (such as negative
lengths, or invalid enum values)
- unbounded memory allocation driven by values in the Parquet metadata,
regardless of physical file/data size
- unchecked overflow integer on arithmetic, with values coming from
Parquet metadata
Note that "Parquet metadata" in the above should be understood in a
broad sense, for example the 4-byte length prefixes in a PLAIN-encoded
BYTE_ARRAY column. Basically any piece of information from a Parquet
file that can be used to derive information about data layout.
Regards
Antoine.