Hi Vinoth, First of all thank you so much for the response.
Now let me explain in detail my use case: We get 5.6 million telemetry data every day and each telemetry data produces 500-1000 Kafka messages and these high frequency Kafka topics are compacted, means TTL for 1 week on Kafka storage. And these messages land on different tables (HUDI tables on S3) based on types. Now we get a delete request (yes its small fraction) [due to either they wrongly send data to an incorrect user or it’s a GDPR request], we get only the userId as the attribute in the delete request. Based on this userId we have to query those tables and pull the records which are for that user and delete them. So if I understand you correctly you are asking whether I can query those tables and pull the corresponding records by spark job and write it back (output) to the same topic which we use for ingestion and the existing streaming ingestion can handle those accordingly by performing delete/upsert based on flags? Other option you are saying is we can have a batch job which can delete the records while streaming jobs are still ingesting data with the new OCC feature. And one of the job fails if both touches the same file. But which one will fail, either streaming or delete batch job? I am AWS EMR and it doesn’t have latest (0.8.0) yet so it would be nice to get a custom build for EMR. So I could test the feature and if get stuck, I can borrow some help from one of you. Once again thank you. Regards, Felix K Jose From: Vinoth Chandar <vin...@apache.org> Date: Wednesday, April 14, 2021 at 3:26 PM To: dev <dev@hudi.apache.org> Subject: Re: GDPR deletes and Consenting deletes of data from hudi table Caution: This e-mail originated from outside of Philips, be careful for phishing. Hi Felix, Most people I think are publishing this data into Kafka,and apply the deletes as a part of the streaming job itself. The reason why this works is because typically, only a small fraction of users leave the service (say << 0.1% weekly is what I have heard). So, the cost of storage on Kafka is not much. Is that not the case for you? Are you looking for one time scrubbing of data for e.g? The benefit of this approach is that you eliminate any concurrency issues that arise from streaming job producing data for a user, while the deletes are also issued for that user. On concurrency control, Hudi now supports multiple writers, if you want to write a background job that will perform these deletes for you. it's in 0.8.0, see https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fhudi.apache.org%2Fdocs%2Fconcurrency_control.html&data=04%7C01%7C%7Ca6cfa9c8d2c9486aef6908d8ff7b241a%7C1a407a2d76754d178692b3ac285306e4%7C0%7C0%7C637540251633946765%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=JlylbgCfMM7NSDJ4KkBSQlxOm%2B0vVBTJBgFYsMWAw2A%3D&reserved=0. One of us can help you out with trying this and rolling out. (Nishith is the feature author). Here, if the delete job touches same files, that the streaming job is writing to, then only one of them will succeed. We are working on a design for true lock free concurrency control, which provides the benefits of both models. But, won't be there for another month or two. Thanks Vinoth On Tue, Apr 13, 2021 at 1:45 PM Kizhakkel Jose, Felix <felix.j...@philips.com.invalid> wrote: > Hi All, > > I have 100s of HUDI tables (AWS S3) where each of those are populated via > Spark structured streaming from kafka streams. Now I have to delete records > for a given user (userId) from all the tables which has data for that user. > Meaning all tables where we have reference to that specific userId. I > cannot republish all the events/records for that user to kafka to perform > delete, since its around 10-15 year’s worth of data for each user and is > going to be so costly and time consuming. So I am wondering how everybody > is performing GDPR on the their HUDI tables? > > > How I get delete request? > On a delete kafka topic we get a delete event [which just contains the > userId of the user to delete], so we have to use that as filter condition > and read all the records from HUDI tables and write it back with data > source operation as ‘delete’. But while performing/running this delete > spark job on the table if the streaming job continues to ingest new > arriving data- what will be the side effect? Will it work, since seems like > multi writers are not currently supported. > > Could you help me with a solution? > > Regards, > Felix K Jose > > ________________________________ > The information contained in this message may be confidential and legally > protected under applicable law. The message is intended solely for the > addressee(s). If you are not the intended recipient, you are hereby > notified that any use, forwarding, dissemination, or reproduction of this > message is strictly prohibited and may be unlawful. If you are not the > intended recipient, please contact the sender by return e-mail and destroy > all copies of the original message. > ________________________________ The information contained in this message may be confidential and legally protected under applicable law. The message is intended solely for the addressee(s). If you are not the intended recipient, you are hereby notified that any use, forwarding, dissemination, or reproduction of this message is strictly prohibited and may be unlawful. If you are not the intended recipient, please contact the sender by return e-mail and destroy all copies of the original message.
