Re: DataFrameReader bottleneck in DataSource#checkAndGlobPathIfNecessary when reading S3 files

[Arwin Tio <ar...@hotmail.com>]

Hello,

I have a ticket/PR out for this issue:

https://issues.apache.org/jira/browse/SPARK-29089
https://github.com/apache/spark/pull/25899

Can somebody please take a look/anything else I can do to get this through the door?

Thanks,

Arwin

________________________________
From: Steve Loughran <st...@cloudera.com>
Sent: September 7, 2019 9:22 AM
To: Arwin Tio <ar...@hotmail.com>
Cc: Sean Owen <sr...@gmail.com>; dev@spark.apache.org <de...@spark.apache.org>
Subject: Re: DataFrameReader bottleneck in DataSource#checkAndGlobPathIfNecessary when reading S3 files



On Fri, Sep 6, 2019 at 10:56 PM Arwin Tio <ar...@hotmail.com>> wrote:
I think the problem is calling globStatus to expand all 300K files.
In my particular case I did not use any glob patterns, so my bottleneck came from the FileSystem#exists specifically. I do concur that the globStatus expansion could also be problematic.

But you might
consider, if possible, running a lot of .csv jobs in parallel to query
subsets of all the files, and union the results. At least there you
parallelize the reading from the object store.
That is a great solution! I think that's what I will do as a workaround for the moment. Right now I'm thinking that a potential improvement here is to parallelize the SparkHadoopUtil#globPathIfNecessary and FileSystem#exists calls whenever possible (i.e. when multiple paths are specified), so that the client doesn't have to.


The other tactic though it'd go through a lot more of the code would be to postpone the exists check until the work is scheduled, which is implicitly in open() on the workers, or explicit when the RDD does the split calculation and calls getFileBlockLocations(). If you are confident that that always happens (and you will have to trace back from those calls in things like org.apache.spark.streaming.util.HdfsUtils and ParallelizedWithLocalityRDD) then you get those scans in the driver ... but I fear regression handling there gets harder.

* have SparkHadoopUtils differentiate between files returned by globStatus(), and which therefore exist, and those which it didn't glob for -it will only need to check those.
* then worry about parallel execution of the scan, again
Okay sounds good, I will take a crack at this and open a ticket. Any thoughts on the parallelism; should it be configurable?

For file input formats (parquet, orc, ...) there is an option, default == 8. Though its also off by default...maybe i should change that.


Another possible QoL improvement here is to show progress log messages - something that indicates to the user that the cluster is stuck while the driver is listing S3 files, maybe even including the FS getStorageStatistics?

yeah. If you want some examples of this, take a look at https://github.com/steveloughran/cloudstore . the locatedfilestatus command replicates what happens during FileInputFormat scans, so is how I'm going to tune IOPs there. It might also be good to have those bits of the hadoop MR classes which spark uses to log internally @ debug, so everything gets this logging if they ask for it.

Happy to take contribs there as Hadoop JIRAs & PRs

Thanks,

Arwin
________________________________
From: Steve Loughran <st...@cloudera.com>>
Sent: September 6, 2019 4:15 PM
To: Sean Owen <sr...@gmail.com>>
Cc: Arwin Tio <ar...@hotmail.com>>; dev@spark.apache.org<ma...@spark.apache.org> <de...@spark.apache.org>>
Subject: Re: DataFrameReader bottleneck in DataSource#checkAndGlobPathIfNecessary when reading S3 files



On Fri, Sep 6, 2019 at 2:50 PM Sean Owen <sr...@gmail.com>> wrote:
I think the problem is calling globStatus to expand all 300K files.
This is a general problem for object stores and huge numbers of files.
Steve L. may have better thoughts on real solutions. But you might
consider, if possible, running a lot of .csv jobs in parallel to query
subsets of all the files, and union the results. At least there you
parallelize the reading from the object store.

yeah, avoid globs and small files, especially small files in deep trees.

I think it's hard to optimize this case from the Spark side as it's
not clear how big a glob like s3://foo/* is going to be. I think it
would take reimplementing some logic to expand the glob incrementally
or something. Or maybe I am overlooking optimizations that have gone
into Spark 3.

A long time ago I actually tried to move Filesystem.globFiles off its own recursive treewalk into supporting the option of flat-list-chlldren + filter. But while you can get some great speedups in some layouts, you can get pathological collapses in perf elsewhere, which makes the people running those queries very sad. So I gave up.

Parallelized scans can do speedup; look at the code in org.apache.hadoop.mapred.LocatedFileStatusFetcher to see what it does there. I've only just started exploring what we can do to tune that, with
HADOOP-16458, HADOOP-16465<https://issues.apache.org/jira/browse/HADOOP-16465>, which should speed up ORC/Parquet scans) . These are designed to cut 1-2 HEAD requests off per directory list, which may seem small but from my early measurements, can be significant.

That's why cutting things like an exists check makes a big difference, especially if you are going to call some list() or open() operation straight after -just call the operation and rely on the FileNotFoundException to tell you when it's not there.

Now, looking at the code, if the list has already come from a real call to globPath, then yes, the existsCall is wasteful, where waste = 500+ mills per file: http://steveloughran.blogspot.com/2016/12/how-long-does-filesystemexists-take.html

For speedup then
* have SparkHadoopUtils differentiate between files returned by globStatus(), and which therefore exist, and those which it didn't glob for -it will only need to check those.
* then worry about parallel execution of the scan, again

Why not file a JIRA on the spark work; send me a ref so I can look at your patch.

One thing to know here is that not only does the S3A FS class have counters for all operations you can get from getStorageStatistics, if you call toString() on it it will print out the current stats. So you can just log the fs string value before and after an operation and see what's gone on. We track FS API calls (op_*) and actual http requests of the store (object_*); both are interesting. object_ to see what is expensive (and in the S3A FS code, what we should cut), the op_ values what API calls are used a lot and should somehow be eliminated or, if you have insights, optimised better. Removal is usually the best, as it speeds up everything.

Long term, relying on directory trees to list your source data, commit algorithms which move/instantiate changes isn't sustainable. Things like Apache Iceberg are where data should go ... things for which S3 can be viewed as a fault-injecting test infrastructure. It's the Chaos Monkey of object storage.


On Fri, Sep 6, 2019 at 7:09 AM Arwin Tio <ar...@hotmail.com>> wrote:
>
> Hello,
>
> On Spark 2.4.4, I am using DataFrameReader#csv to read about 300000 files on S3, and I've noticed that it takes about an hour for it to load the data on the Driver. You can see the timestamp difference when the log from InMemoryFileIndex occurs from 7:45 to 8:54:
>
> 19/09/06 07:44:42 INFO SparkContext: Running Spark version 2.4.4
> 19/09/06 07:44:42 INFO SparkContext: Submitted application: LoglineParquetGenerator
> ...
> 19/09/06 07:45:40 INFO StateStoreCoordinatorRef: Registered StateStoreCoordinator endpoint
> 19/09/06 08:54:57 INFO InMemoryFileIndex: Listing leaf files and directories in parallel under: [300K files...]
>
>
> I believe that the issue comes from DataSource#checkAndGlobPathIfNecessary [0], specifically from when it is calling FileSystem#exists. Unlike bulkListLeafFiles, the existence check here happens in a single-threaded flatMap, which is a blocking network call if your files are stored on S3.
>
> I believe that there is a fairly straightforward opportunity for improvement here, which is to parallelize the existence check perhaps with a configurable number of threads. If that seems reasonable, I would like to create a JIRA ticket and submit a patch. Please let me know!
>
> Cheers,
>
> Arwin
>
> [0] https://github.com/apache/spark/blob/branch-2.4/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/DataSource.scala#L557

---------------------------------------------------------------------
To unsubscribe e-mail: dev-unsubscribe@spark.apache.org<ma...@spark.apache.org>