spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Cheng, Hao" <>
Subject RE: Very high latency to initialize a DataFrame from partitioned parquet database.
Date Wed, 12 Aug 2015 00:56:37 GMT
Definitely worth to try. And you can sort the record before writing out, and then you will
get the parquet files without overlapping keys.
Let us know if that helps.


From: Philip Weaver []
Sent: Wednesday, August 12, 2015 4:05 AM
To: Cheng Lian
Cc: user
Subject: Re: Very high latency to initialize a DataFrame from partitioned parquet database.

Do you think it might be faster to put all the files in one directory but still partitioned
the same way? I don't actually need to filter on the values of the partition keys, but I need
to rely on there be no overlap in the value of the keys between any two parquet files.

On Fri, Aug 7, 2015 at 8:23 AM, Philip Weaver <<>>
Thanks, I also confirmed that the partition discovery is slow by writing a non-Spark application
that uses the parquet library directly to load that partitions.

It's so slow that my colleague's Python application can read the entire contents of all the
parquet data files faster than my application can even discover the partitions!

On Fri, Aug 7, 2015 at 2:09 AM, Cheng Lian <<>>
However, it's weird that the partition discovery job only spawns 2 tasks. It should use the
default parallelism, which is probably 8 according to the logs of the next Parquet reading
job. Partition discovery is already done in a distributed manner via a Spark job. But the
parallelism is mysteriously low...


On 8/7/15 3:32 PM, Cheng Lian wrote:

Hi Philip,

Thanks for providing the log file. It seems that most of the time are spent on partition discovery.
The code snippet you provided actually issues two jobs. The first one is for listing the input
directories to find out all leaf directories (and this actually requires listing all leaf
files, because we can only assert that a directory is a leaf one when it contains no sub-directories).
Then partition information is extracted from leaf directory paths. This process starts at:

10:51:44 INFO sources.HadoopFsRelation: Listing leaf files and directories in parallel under:
file:/home/pweaver/work/parquet/day=20150225, …

and ends at:

10:52:31 INFO scheduler.TaskSchedulerImpl: Removed TaskSet 0.0, whose tasks have all completed,
from pool

The actual tasks execution time is about 36s:

10:51:54 INFO scheduler.TaskSetManager: Starting task 0.0 in stage 0.0 (TID 0, lindevspark5,
PROCESS_LOCAL, 3087 bytes)
10:52:30 INFO scheduler.TaskSetManager: Finished task 0.0 in stage 0.0 (TID 0) in 36107 ms
on lindevspark5 (1/2)

You mentioned that your dataset has about 40,000+ partitions, so there are a lot of leaf directories
and files out there. My guess is that the local file system spent lots of time listing FileStatus-es
of all these files.

I also noticed that Mesos job scheduling takes more time then expected. It is probably because
this is the first Spark job executed in the application, and the system is not warmed up yet.
For example, there’s a 6s gap between these two adjacent lines:

10:51:45 INFO scheduler.TaskSchedulerImpl: Adding task set 0.0 with 2 tasks
10:51:51 INFO mesos.CoarseMesosSchedulerBackend: Mesos task 0 is now TASK_RUNNING

The 2nd Spark job is the real Parquet reading job, and this one actually finishes pretty quickly,
only 3s (note that the Mesos job scheduling latency is also included):

10:52:32 INFO scheduler.DAGScheduler: Got job 1 (parquet at App.scala:182) with 8 output partitions
10:52:32 INFO scheduler.TaskSetManager: Starting task 0.0 in stage 1.0 (TID 2, lindevspark4,
PROCESS_LOCAL, 2058 bytes)
10:52:32 INFO scheduler.TaskSetManager: Starting task 1.0 in stage 1.0 (TID 3, lindevspark5,
PROCESS_LOCAL, 2058 bytes)
10:52:32 INFO scheduler.TaskSetManager: Starting task 2.0 in stage 1.0 (TID 4, lindevspark4,
PROCESS_LOCAL, 2058 bytes)
10:52:34 INFO scheduler.TaskSetManager: Finished task 6.0 in stage 1.0 (TID 8) in 1527 ms
on lindevspark4 (6/8)
10:52:34 INFO scheduler.TaskSetManager: Finished task 4.0 in stage 1.0 (TID 6) in 1533 ms
on lindevspark4 (7/8)
10:52:35 INFO scheduler.TaskSetManager: Finished task 7.0 in stage 1.0 (TID 9) in 2886 ms
on lindevspark5 (8/8)

That might be the reason why you observed that the C parquet library you mentioned (is it
parquet-cpp?) is an order of magnitude faster?


On 8/7/15 2:02 AM, Philip Weaver wrote:
With DEBUG, the log output was over 10MB, so I opted for just INFO output. The (sanitized)
log is attached.

The driver is essentially this code:


    val t = System.currentTimeMillis
    val df =

    val elapsed = System.currentTimeMillis - t
    info(s"Init time: ${elapsed} ms")

We've also observed that it is very slow to read the contents of the parquet files. My colleague
wrote a PySpark application that gets the list of files, parallelizes it, maps across it and
reads each file manually using a C parquet library, and aggregates manually in the loop. Ignoring
the 1-2 minute initialization cost, compared to a Spark SQL or DataFrame query in Scala, his
is an order of magnitude faster. Since he is parallelizing the work through Spark, and that
isn't causing any performance issues, it seems to be a problem with the parquet reader. I
may try to do what he did to construct a DataFrame manually, and see if I can query it with
Spark SQL with reasonable performance.

- Philip

On Thu, Aug 6, 2015 at 8:37 AM, Cheng Lian <<>>
Would you mind to provide the driver log?

On 8/6/15 3:58 PM, Philip Weaver wrote:
I built spark from the v1.5.0-snapshot-20150803 tag in the repo and tried again.

The initialization time is about 1 minute now, which is still pretty terrible.

On Wed, Aug 5, 2015 at 9:08 PM, Philip Weaver <<>>
Absolutely, thanks!

On Wed, Aug 5, 2015 at 9:07 PM, Cheng Lian <<>>
We've fixed this issue in 1.5

Could you give it a shot to see whether it helps in your case? We've observed ~50x performance
boost with schema merging turned on.


On 8/6/15 8:26 AM, Philip Weaver wrote:
I have a parquet directory that was produced by partitioning by two keys, e.g. like this:

df.write.partitionBy("a", "b").parquet("asdf")

There are 35 values of "a", and about 1100-1200 values of "b" for each value of "a", for a
total of over 40,000 partitions.

Before running any transformations or actions on the DataFrame, just initializing it like
this takes 2 minutes:

val df ="asdf")

Is this normal? Is this because it is doing some bookeeping to discover all the partitions?
Is it perhaps having to merge the schema from each partition? Would you expect it to get better
or worse if I subpartition by another key?

- Philip


View raw message