[SPARK-22357][CORE] SparkContext.binaryFiles ignore minPartitions parameter by bomeng · Pull Request #21638 · apache/spark

bomeng · 2018-06-25T20:08:02Z

What changes were proposed in this pull request?

Fix the issue that minPartitions was not used in the method. This is a simple fix and I am not trying to make it complicated. The purpose is to still allow user to control the defaultParallelism through the value of minPartitions, while also via sc.defaultParallelism parameters.

How was this patch tested?

I have not provided the additional test since the fix is very straightforward.

SparkQA · 2018-06-26T00:37:51Z

Test build #92309 has finished for PR 21638 at commit 0fc35d4.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

viirya · 2018-06-26T01:16:38Z

+ val defaultMaxSplitBytes = Math.max(
+ sc.getConf.get(config.FILES_MAX_PARTITION_BYTES), minPartitions)
 val openCostInBytes = sc.getConf.get(config.FILES_OPEN_COST_IN_BYTES)
 val defaultParallelism = sc.defaultParallelism


hmm, shouldn't minPartitions be used like this?
valdefaultParallelism=Math.max(sc.defaultParallelism, if (minPartitions ==0) 1else minPartitions)

Could you describe the use case when you need to take into account minPartitions. By default, FILES_MAX_PARTITION_BYTES is 128MB. Let's say it is even set to 1000, and minPartitions equals to 10 000. What is the reason to set the max size of splits in bytes to the min number of partition. Why should bigger number of partitions require bigger split size? Could you add more details to the PR description, please.

MaxGekk · 2018-06-26T12:21:13Z

It seems there is similar code there:

spark/sql/core/src/main/scala/org/apache/spark/sql/execution/DataSourceScanExec.scala

Lines 424 to 433 in e76b012

valdefaultMaxSplitBytes=
 fsRelation.sparkSession.sessionState.conf.filesMaxPartitionBytes 
valopenCostInBytes= fsRelation.sparkSession.sessionState.conf.filesOpenCostInBytes 
valdefaultParallelism= fsRelation.sparkSession.sparkContext.defaultParallelism 
valtotalBytes= selectedPartitions.flatMap(_.files.map(_.getLen + openCostInBytes)).sum 
valbytesPerCore= totalBytes / defaultParallelism 
valmaxSplitBytes=Math.min(defaultMaxSplitBytes, Math.max(openCostInBytes, bytesPerCore)) 
 logInfo(s"Planning scan with bin packing, max size: $maxSplitBytes bytes, "+
s"open cost is considered as scanning $openCostInBytes bytes.") 

. Should it be changed in the same way?

HyukjinKwon · 2018-06-26T13:27:12Z

Not sure yet but let's leave that out of this PR.

MaxGekk · 2018-06-26T17:59:27Z

+ val defaultMaxSplitBytes = sc.getConf.get(config.FILES_MAX_PARTITION_BYTES)
 val openCostInBytes = sc.getConf.get(config.FILES_OPEN_COST_IN_BYTES)
- val defaultParallelism = sc.defaultParallelism
+ val defaultParallelism = Math.max(sc.defaultParallelism, minPartitions)


Now it makes much more sense.

SparkQA · 2018-06-26T22:25:05Z

Test build #92350 has finished for PR 21638 at commit c24fbe5.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

bomeng · 2018-06-27T17:34:19Z

@HyukjinKwon please review. thanks.

HyukjinKwon · 2018-07-16T02:49:23Z

retest this please

SparkQA · 2018-07-16T07:05:01Z

Test build #93067 has finished for PR 21638 at commit c24fbe5.

This patch fails due to an unknown error code, -9.
This patch merges cleanly.
This patch adds no public classes.

jiangxb1987 · 2018-07-16T15:58:50Z

 val defaultMaxSplitBytes = sc.getConf.get(config.FILES_MAX_PARTITION_BYTES)
 val openCostInBytes = sc.getConf.get(config.FILES_OPEN_COST_IN_BYTES)
- val defaultParallelism = sc.defaultParallelism
+ val defaultParallelism = Math.max(sc.defaultParallelism, minPartitions)


If sc.defaultParallelism < 2, and minParititions is not set in BinaryFileRDD, then previously defaultParallelism shall be the same as sc.defaultParallelism, and after the change it will be 2. Have you already consider this case and feel it's right behavior change to make?

you need to pass in the minPartitions to use this method, what do you mean minParititions is not set?

I metioned BinaryFileRDD not this method, you can check the code to see how it handles the default value.

BinaryFileRDD will set minPartitions, which will either be defaultMinPartitions, or the values you can set via binaryFiles(path, minPartitions) method. Eventually, this minPartitions value will be passed to setMinPartitions() method.

srowen · 2018-07-18T14:32:38Z

Because this method is internal to Spark, why not just take out the parameter? Yes it's superfluous now, but it's been this way for a while, and seems perhaps better to avoid a behavior change. In fact you can pull a minPartitions parameter out of several private methods then. You can't remove the parameter to binaryFiles, sure, but it can be documented as doing nothing.

bomeng · 2018-07-18T21:09:38Z

Either way works for me, but I think since this is not a private method, so people may use it in their own approach. The minimal change will be the best.

srowen · 2018-07-18T21:14:41Z

Except for binaryFiles, everything else that needs to change is private to Spark. I know it's public in the bytecode, but only Java callers could accidentally exploit that. Still I don't personally care too much either way, as long as all the unused args are documented, I guess, for completeness.

HyukjinKwon · 2018-07-19T02:23:53Z

Yea, it's internal to Spark. Might be good to keep it but that concern should be secondary IMHO.

SparkQA · 2018-08-26T01:37:52Z

Test build #4290 has finished for PR 21638 at commit c24fbe5.

This patch fails Spark unit tests.
This patch does not merge cleanly.
This patch adds no public classes.

SparkQA · 2018-08-27T21:03:28Z

Test build #95295 has finished for PR 21638 at commit 5e46efb.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

srowen · 2018-08-29T00:40:11Z

Merged to master

gatorsmile · 2018-09-03T05:48:38Z

 val defaultMaxSplitBytes = sc.getConf.get(config.FILES_MAX_PARTITION_BYTES)
 val openCostInBytes = sc.getConf.get(config.FILES_OPEN_COST_IN_BYTES)
- val defaultParallelism = sc.defaultParallelism
+ val defaultParallelism = Math.max(sc.defaultParallelism, minPartitions)


We should have a test case; otherwise, we could hit the same issue again.

BTW, it is easy to add such a test case. We can even test the behaviors of the boundary cases. cc @srowen @HyukjinKwon @MaxGekk @jiangxb1987

I think it's hard to test, technically, because setMinPartitions is only a hint. In the case of binaryFiles we know it will put a hard limit on the number of partitions, but it isn't true of other implementations. We can still make a simple test for all of these, it just may be asserting behavior that could change in the future in Hadoop, though I strongly doubt it would.

I agree it is hard to test. I appreciate If anyone can give me some hints of how to do these (how to verify and where to put my test cases).

Would you mind following up with a test that just asserts that asking for, say, 20 partitions results in 20 partitions? This is technically too specific as a test, but is probably fine for now.

From the codes, you can see the calculation is just the intermediate result and this method won't return any value. Checking the split size does not make sense for this test case because it depends on multiple variables and this is just one of them.

sc = new SparkContext(new SparkConf().setAppName("test").setMaster("local") .set(config.FILES_OPEN_COST_IN_BYTES.key, "0") .set("spark.default.parallelism", "1")) println(sc.binaryFiles(dirpath1, minPartitions = 50).getNumPartitions) println(sc.binaryFiles(dirpath1, minPartitions = 1).getNumPartitions)
It is not hard to verify whether the parameter minPartitions takes an effect. Currently, the description of this parameter is not clear. We need to document it clear which factors impact the actual number of partitions; otherwise, users will not understand how to use it.

gatorsmile · 2018-09-03T05:54:34Z

@bomeng Could you submit a follow-up PR to add a test case?

HyukjinKwon · 2018-09-03T06:32:14Z

Yea, let's add a regression test.

bomeng · 2018-09-05T00:41:26Z

Here is the test code, not sure it is right or not ---

 test("Number of partitions") {
sc = new SparkContext(new SparkConf().setAppName("test").setMaster("local")
.set("spark.files.maxPartitionBytes", "10")
.set("spark.files.openCostInBytes", "0")
.set("spark.default.parallelism", "1"))
val dir1 = Utils.createTempDir()
val dirpath1 = dir1.getAbsolutePath
val dir2 = Utils.createTempDir()
val dirpath2 = dir2.getAbsolutePath
val file1 = new File(dir1, "part-00000")
val file2 = new File(dir1, "part-00001")
Files.write("someline1 in file1\nsomeline2 in file1\nsomeline3 in file1", file1,
StandardCharsets.UTF_8)
Files.write("someline1 in file2\nsomeline2 in file2\nsomeline3 in file2", file2,
StandardCharsets.UTF_8)
assert(sc.binaryFiles(dirpath1, minPartitions = 1).getNumPartitions == 2)
assert(sc.binaryFiles(dirpath1, minPartitions = 2).getNumPartitions == 2)
assert(sc.binaryFiles(dirpath1, minPartitions = 50).getNumPartitions == 2)
}

srowen · 2018-09-05T00:54:02Z

Ideally the last test should have 50 partitions? is it because we really need the test data to be at least 50 bytes? ideally a multiple of 50, I guess.

…itions parameter ## What changes were proposed in this pull request? This adds a test following #21638 ## How was this patch tested? Existing tests and new test. Closes#22356 from srowen/SPARK-22357.2. Authored-by: Sean Owen <sean.owen@databricks.com> Signed-off-by: Sean Owen <sean.owen@databricks.com> (cherry picked from commit 4e3365b) Signed-off-by: Sean Owen <sean.owen@databricks.com>

bomeng added 2 commits June 25, 2018 13:02

fix: SparkContext.binaryFiles ignore minPartitions parameter
b9eea49

should be max()
0fc35d4

viirya reviewed Jun 26, 2018
View reviewed changes

fix the issue
c24fbe5

MaxGekk reviewed Jun 26, 2018
View reviewed changes

jiangxb1987 reviewed Jul 16, 2018
View reviewed changes

srowen approved these changes Aug 25, 2018
View reviewed changes

Merge remote-tracking branch 'upstream/master' into 22357
5e46efb

HyukjinKwon approved these changes Aug 28, 2018
View reviewed changes

asfgit closed this in bbbf814Aug 29, 2018

gatorsmile reviewed Sep 3, 2018
View reviewed changes

srowen mentioned this pull request Sep 6, 2018
[SPARK-22357][CORE][FOLLOWUP] SparkContext.binaryFiles ignore minPartitions parameter #22356
Closed

Uh oh!

Conversation

bomeng commented Jun 25, 2018• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What changes were proposed in this pull request?

How was this patch tested?

Uh oh!

SparkQA commented Jun 26, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

MaxGekk commented Jun 26, 2018

Uh oh!

HyukjinKwon commented Jun 26, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SparkQA commented Jun 26, 2018

Uh oh!

bomeng commented Jun 27, 2018

Uh oh!

HyukjinKwon commented Jul 16, 2018

Uh oh!

SparkQA commented Jul 16, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

srowen commented Jul 18, 2018

Uh oh!

bomeng commented Jul 18, 2018

Uh oh!

srowen commented Jul 18, 2018

Uh oh!

HyukjinKwon commented Jul 19, 2018

Uh oh!

SparkQA commented Aug 26, 2018

Uh oh!

SparkQA commented Aug 27, 2018

Uh oh!

srowen commented Aug 29, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

gatorsmile commented Sep 3, 2018

Uh oh!

HyukjinKwon commented Sep 3, 2018

Uh oh!

bomeng commented Sep 5, 2018

Uh oh!

srowen commented Sep 5, 2018

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

8 participants

bomeng commented Jun 25, 2018•
edited
Loading