Thanks for the pointer, Vadim. However, I just tried it with Spark 2.4 and
get the same failure. (I was previously testing with 2.2 and/or 2.3.) And I
don't see this particular issue referred to there. The ticket that Harel
commented on indeed appears to be the most similar one to this issue:
There's an umbrella ticket for various 2GB limitations
https://issues.apache.org/jira/browse/SPARK-6235
On Fri, Nov 8, 2019 at 4:11 PM Jacob Lynn wrote:
>
> Sorry for the noise, folks! I understand that reducing the number of
> partitions works around the issue (at the scale I'm working at,
Sorry for the noise, folks! I understand that reducing the number of
partitions works around the issue (at the scale I'm working at, anyway) --
as I mentioned in my initial email -- and I understand the root cause. I'm
not looking for advice on how to resolve my issue. I'm just pointing out
that
Basically, the driver tracks partitions and sends it over to
executors, so what it's trying to do is to serialize and compress the
map but because it's so big, it goes over 2GiB and that's Java's limit
on the max size of byte arrays, so the whole thing drops.
The size of data doesn't matter here
File system is HDFS. Executors are 2 cores, 14GB RAM. But I don't think
either of these relate to the problem -- this is a memory allocation issue
on the driver side, and happens in an intermediate stage that has no HDFS
read/write.
On Fri, Nov 8, 2019 at 10:01 AM Spico Florin wrote:
> Hi!
>
Hi!
What file system are you using: EMRFS or HDFS?
Also what memory are you using for the reducer ?
On Thu, Nov 7, 2019 at 8:37 PM abeboparebop wrote:
> I ran into the same issue processing 20TB of data, with 200k tasks on both
> the map and reduce sides. Reducing to 100k tasks each resolved
I ran into the same issue processing 20TB of data, with 200k tasks on both
the map and reduce sides. Reducing to 100k tasks each resolved the issue.
But this could/would be a major problem in cases where the data is bigger or
the computation is heavier, since reducing the number of partitions may
I understand the error is because the number of partitions is very high,
yet when processing 40 TB (and this number is expected to grow) this number
seems reasonable:
40TB / 300,000 will result in partitions size of ~ 130MB (data should be
evenly distributed).
On Fri, Sep 7, 2018 at 6:28 PM Vadim
You have too many partitions, so when the driver is trying to gather
the status of all map outputs and send back to executors it chokes on
the size of the structure that needs to be GZipped, and since it's
bigger than 2GiB, it produces OOM.
On Fri, Sep 7, 2018 at 10:35 AM Harel Gliksman wrote:
>
Hi,
We are running a Spark (2.3.1) job on an EMR cluster with 500 r3.2xlarge
(60 GB, 8 vcores, 160 GB SSD ). Driver memory is set to 25GB.
It processes ~40 TB of data using aggregateByKey in which we specify
numPartitions = 300,000.
Map side tasks succeed, but reduce side tasks all fail.
We
10 matches
Mail list logo
