This message was deleted.

Since you haven't received a response from someone with experience in hadoop ingestion, I'll venture a guess: Any difference in the lz4 compression libraries in hadoop vs druid deployments? Druid currently bundles version 1.8.0. Also, what version of Druid are you on? the lz4 version may be different.

we are running druid major version 24 with 1.8. We had a similar thought and have been pulling up yarn container classpaths in the yarn launcher logs and haven't found an example of a version other than 1.8 getting sucked in. Definitely seems like a possibility though

lz4 is meant to be a stable format— probably more likely that the segments are getting corrupted somehow, and that's coming out as lz4 errors b/c lz4 is the first code path that read data from the files

Thanks for the thoughts, Gian! We are definitely starting to trend in that direction - a bad hadoop node - instead of lz4 issues.