Shall we document the known issue on file stream sink and provide workaround? There's more than a couple of questions about this in a couple of months, and there have been 5 related issues. The workaround Burak provided looks nice to those who don't need to have end-to-end exactly once semantics (and in many cases they are OK with the semantics).

On Fri, Jun 19, 2020 at 8:05 AM Burak Yavuz <> wrote:
Hi Rachana,

If you don't need exactly once semantics, you can use foreachBatch to write your data.
df.writeStream.foreachBatch { case (df, batchId) =>

However, I would highly recommend upgrading to some ACID data store project like Delta Lake (which natively supports streaming), Iceberg or Hudi. 


On Thu, Jun 18, 2020 at 8:24 AM Rachana Srivastava <> wrote:
Thanks so much for your response.  I agree using Spark Streaming is not recommended.  But I want a stable system we cannot have a system that crashes every 5 days.  As seen in the picture below we have nearly 47 mb of data in the metadata folder.  Issue is when size of data increases to nearly 13 GB and driver memory is 5 GB that time we get OOM.  Not sure how to add TTL to metadata, if I delete metadata then I have to delete checkpoint hence loose the data.  

Inline image

On Thursday, June 18, 2020, 03:23:32 AM PDT, Jacek Laskowski <> wrote:

Hi Rachana,

> Should I go backward and use Spark Streaming DStream based.

No. Never. It's no longer supported (and should really be removed from the codebase once and for all - dreaming...).

Spark focuses on Spark SQL and Spark Structured Streaming as user-facing modules for batch and streaming queries, respectively.

Please note that I'm not a PMC member or even a committer so I'm speaking for myself only (not representing the project in an official way).

On Thu, Jun 18, 2020 at 12:03 AM Rachana Srivastava <> wrote:
Structured Stream Vs Spark Steaming (DStream)?

Which is recommended for system stability.  Exactly once is NOT first priority.  First priority is STABLE system.

I am I need to make a decision soon.  I need help.  Here is the question again.  Should I go backward and use Spark Streaming DStream based.  Write our own checkpoint and go from there.  At least we never encounter these metadata issues there.



On Wednesday, June 17, 2020, 02:02:20 PM PDT, Jungtaek Lim <> wrote:

Just in case if anyone prefers ASF projects then there are other alternative projects in ASF as well, alphabetically, Apache Hudi [1] and Apache Iceberg [2]. Both are recently graduated as top level projects. (DISCLAIMER: I'm not involved in both.)

BTW it would be nice if we make the metadata implementation on file stream source/sink be pluggable - from what I've seen, plugin approach has been selected as the way to go whenever some part is going to be complicated and it becomes arguable whether the part should be handled in Spark project vs should be outside. e.g. checkpoint manager, state store provider, etc. It would open up chances for the ecosystem to play with the challenge "without completely re-writing the file stream source and sink", focusing on scalability for metadata in a long run query. Alternative projects described above will still provide more higher-level features and look attractive, but sometimes it may be just "using a sledgehammer to crack a nut".

On Thu, Jun 18, 2020 at 2:34 AM Tathagata Das <> wrote:
Hello Rachana,

Getting exactly-once semantics on files and making it scale to a very large number of files are very hard problems to solve. While Structured Streaming + built-in file sink solves the exactly-once guarantee that DStreams could not, it is definitely limited in other ways (scaling in terms of files, combining batch and streaming writes in the same place, etc). And solving this problem requires a holistic solution that is arguably beyond the scope of the Spark project. 

There are other projects that are trying to solve this file management issue. For example, Delta Lake (full disclosure, I am involved in it) was built to exactly solve this problem - get exactly-once and ACID guarantees on files, but also scale to handling millions of files. Please consider it as part of your solution. 

On Wed, Jun 17, 2020 at 9:50 AM Rachana Srivastava <> wrote:
I have written a simple spark structured steaming app to move data from Kafka to S3. Found that in order to support exactly-once guarantee spark creates _spark_metadata folder, which ends up growing too large as the streaming app is SUPPOSE TO run FOREVER. But when the streaming app runs for a long time the metadata folder grows so big that we start getting OOM errors. Only way to resolve OOM is delete Checkpoint and Metadata folder and loose VALUABLE customer data.

Spark open JIRAs SPARK-24295 and SPARK-29995, SPARK-30462, and SPARK-24295)

Since Spark Streaming was NOT broken like this. Is Spark Streaming a BETTER choice?

To unsubscribe e-mail: