spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Tathagata Das <tathagata.das1...@gmail.com>
Subject Re: "spark.streaming.unpersist" and "spark.cleaner.ttl"
Date Sun, 27 Jul 2014 00:01:51 GMT
Yeah, I wrote those lines a while back, I wanted to contrast storage
levels with and without serialization. Should have realized that
StorageLevel.MEMORY_ONLY_SER can be confused to be the default level.

TD

On Wed, Jul 23, 2014 at 5:12 AM, Shao, Saisai <saisai.shao@intel.com> wrote:
> Yeah, the document may not be precisely aligned with latest code, so the best way is
to check the code.
>
> -----Original Message-----
> From: Haopu Wang [mailto:HWang@qilinsoft.com]
> Sent: Wednesday, July 23, 2014 5:56 PM
> To: user@spark.apache.org
> Subject: RE: "spark.streaming.unpersist" and "spark.cleaner.ttl"
>
> Jerry, thanks for the response.
>
> For the default storage level of DStream, it looks like Spark's document is wrong. In
this link: http://spark.apache.org/docs/latest/streaming-programming-guide.html#memory-tuning
> It mentions:
> "Default persistence level of DStreams: Unlike RDDs, the default persistence level of
DStreams serializes the data in memory (that is, StorageLevel.MEMORY_ONLY_SER for DStream
compared to StorageLevel.MEMORY_ONLY for RDDs). Even though keeping the data serialized incurs
higher serialization/deserialization overheads, it significantly reduces GC pauses."
>
> I will take a look at DStream.scala although I have no Scala experience.
>
> -----Original Message-----
> From: Shao, Saisai [mailto:saisai.shao@intel.com]
> Sent: 2014年7月23日 15:13
> To: user@spark.apache.org
> Subject: RE: "spark.streaming.unpersist" and "spark.cleaner.ttl"
>
> Hi Haopu,
>
> Please see the inline comments.
>
> Thanks
> Jerry
>
> -----Original Message-----
> From: Haopu Wang [mailto:HWang@qilinsoft.com]
> Sent: Wednesday, July 23, 2014 3:00 PM
> To: user@spark.apache.org
> Subject: "spark.streaming.unpersist" and "spark.cleaner.ttl"
>
> I have a DStream receiving data from a socket. I'm using local mode.
> I set "spark.streaming.unpersist" to "false" and leave "
> spark.cleaner.ttl" to be infinite.
> I can see files for input and shuffle blocks under "spark.local.dir"
> folder and the size of folder keeps increasing, although JVM's memory usage seems to
be stable.
>
> [question] In this case, because input RDDs are persisted but they don't fit into memory,
so write to disk, right? And where can I see the details about these RDDs? I don't see them
in web UI.
>
> [answer] Yes, if memory is not enough to put input RDDs, this data will be flush to disk,
because the default storage level is "MEMORY_AND_DISK_SER_2" as you can see in StreamingContext.scala.
Actually you cannot not see the input RDD in web UI, you can only see the cached RDD in web
UI.
>
> Then I set "spark.streaming.unpersist" to "true", the size of "spark.local.dir" folder
and JVM's used heap size are reduced regularly.
>
> [question] In this case, because I didn't change "spark.cleaner.ttl", which component
is doing the cleanup? And what's the difference if I set "spark.cleaner.ttl" to some duration
in this case?
>
> [answer] If you set "spark.streaming.unpersist" to true, old unused rdd will be deleted,
as you can see in DStream.scala. While "spark.cleaner.ttl" is timer-based spark cleaner, not
only clean streaming data, but also broadcast, shuffle and other data.
>
> Thank you!
>

Mime
View raw message