Bump again.
Unlike file stream sink which has lots of limitations and many of us have
been suggesting alternatives, file stream source is the only way if end
users want to read the data from files. No alternative unless they
introduce another ETL & storage (probably Kafka).
On Fri, Jul 31, 2020 at 3:06 PM Jungtaek Lim
wrote:
> Hi German,
>
> option 1 isn't about "deleting" the old files, as your input directory may
> be accessed by multiple queries. Kafka centralizes the maintenance of input
> data hence possible to apply retention without problem.
> option 1 is more about "hiding" the old files being read, so that end
> users "may" be able to delete the files once they ensure "all queries
> accessing the input directory" don't see the old files.
>
> On Fri, Jul 31, 2020 at 2:57 PM German Schiavon
> wrote:
>
>> HI Jungtaek,
>>
>> I have a question, aren't both approaches compatible?
>>
>> How I see it, I think It would be interesting to have a retention period
>> to delete old files and/or the possibility of indicating an offset
>> (Timestamp). It would be very "similar" to how we do it with kafka.
>>
>> WDYT?
>>
>> On Thu, 30 Jul 2020 at 23:51, Jungtaek Lim
>> wrote:
>>
>>> (I'd like to keep the discussion thread focusing on the specific topic -
>>> let's initiate another discussion threads on different topics.)
>>>
>>> Thanks for the input. I'd like to emphasize that the point in discussion
>>> is the "latestFirst" option - the rationalization starts from
>>> growing metadata log issues. I hope your input is picking option 2, but
>>> could you please make clear your input represents OK to "replace" the
>>> "latestFirst" option with "starting from timestamp"?
>>>
>>>
>>> On Thu, Jul 30, 2020 at 4:48 PM vikram agrawal
>>> wrote:
>>>
If we compare file-stream source with other streaming sources such as
Kafka, the current behavior is indeed incomplete. Starting the streaming
from a custom offset/particular point of time is something that is missing.
Typically filestream sources don't have auto-deletion of the older
data/files. In kafka we can define the retention period. So even if we use
"Earliest" we won't end up reading from the time when the Kafka topic was
created. On the other hand, streaming sources can hold very old files. It's
very valid use-cases to read the bulk of the old files using a batch job
until a particular timestamp. And then use streaming jobs for real-time
updates.
So having support where we can specify a timestamp. and we would
consider files created post that timestamp can be useful.
Another concern which we need to consider is the listing cost. is there
any way we can avoid listing the entire base directory and then filtering
out the new files. if the data is organized as partitions using date, will
it help to list only those partitions where new files were added?
On Thu, Jul 30, 2020 at 11:22 AM Jungtaek Lim <
kabhwan.opensou...@gmail.com> wrote:
> bump, is there any interest on this topic?
>
> On Mon, Jul 20, 2020 at 6:21 AM Jungtaek Lim <
> kabhwan.opensou...@gmail.com> wrote:
>
>> (Just to add rationalization, you can refer the original mail thread
>> on dev@ list to see efforts on addressing problems in file stream
>> source / sink -
>> https://lists.apache.org/thread.html/r1cd548be1cbae91c67e5254adc0404a99a23930f8a6fde810b987285%40%3Cdev.spark.apache.org%3E
>> )
>>
>> On Mon, Jul 20, 2020 at 6:18 AM Jungtaek Lim <
>> kabhwan.opensou...@gmail.com> wrote:
>>
>>> Hi devs,
>>>
>>> As I have been going through the various issues on metadata log
>>> growing, it's not only the issue of sink, but also the issue of source.
>>> Unlike sink metadata log which entries should be available to the
>>> readers, the source metadata log is only for the streaming query
>>> starting
>>> from the checkpoint, hence in theory it should only memorize about
>>> minimal entries which prevent processing multiple times on the same
>>> file.
>>>
>>> This is not applied to the file stream source, and I think it's
>>> because of the existence of the "latestFirst" option which I haven't
>>> seen
>>> from any sources. The option works as reading files in "backward" order,
>>> which means Spark can read the oldest file and latest file together in a
>>> micro-batch, which ends up having to memorize all files previously read.
>>> The option can be changed during query restart, so even if the query is
>>> started with "latestFirst" being false, it's not safe to apply the
>>> logic of
>>> minimizing entries to memorize, as the option can be changed to true and
>>> then we'll read files again.
>>>
>>> I'm seeing two approaches here:
>>>
>>> 1) apply "retention" - unlike "maxFileAge", the option would apply
>>> to latestFirst as
