Checkout Hudi (https://github.com/apache/incubator-hudi) which adds upsert functionality on top of columnar data such as Parquet.
Chao On Mon, Apr 15, 2019 at 10:49 AM Vinod Kumar Vavilapalli <vino...@apache.org> wrote: > If one uses HDFS as raw file storage where a single file intermingles data > from all users, it's not easy to achieve what you are trying to do. > > Instead, using systems (e.g. HBase, Hive) that support updates and deletes > to individual records is the only way to go. > > +Vinod > > On Apr 15, 2019, at 1:32 AM, Ivan Panico <iv.pan...@gmail.com> wrote: > > Hi, > > Recent GDPR introduced a new right for people : the right to be forgotten. > This right means that if an organization is asked by a customer to delete > all his data, the organization have to comply most of the time (there are > conditions which can suspend this right but that's besides my point). > > Now HDFS being WORM (Write Once Read Multpliple Times), I guess you see > where I'm going. What would be the best way to implement this line deletion > feature (supposing that when a customer asks for a delete of all his data, > the organization would have to delete some lines in some HDFS files). > > Right now I'm going for the following : > > - Create a key-value base (user, [files]) > - On file writing, feed this base with the users and file location (by > appending or updating a key). > - When the deletion is requested by the user "john", look in that base > and rewrite all the files of the "john" key (read the file in memmory, > suppress the lines of "john", rewrite the files) > > > Would this be the most hadoop way to do that ? > I discarded some cryptoshredding like solution because the HDFS data has > to be readable by some mutliple proprietary softwares and by users at some > point and I'm not sur how to incorporate a decyphering step for all those > uses cases. > Also, I came up with this table solution because a violent grep for some > key on the whole HDFS tree seemed unlikely to scale but maybe I'm mistaken ? > > Thanks for your help, > Best regards > > >