问题
There seem to be a few postings on this but none seem to answer what I understand.
The following code run on DataBricks:
spark.sparkContext.setCheckpointDir("/dbfs/FileStore/checkpoint/cp1/loc7")
val checkpointDir = spark.sparkContext.getCheckpointDir.get
val ds = spark.range(10).repartition(2)
ds.cache()
ds.checkpoint()
ds.count()
ds.rdd.isCheckpointed
Added an improvement of sorts:
...
val ds2 = ds.checkpoint(eager=true)
println(ds2.queryExecution.toRdd.toDebugString)
...
returns:
(2) MapPartitionsRDD[307] at toRdd at command-1835760423149753:13 []
| MapPartitionsRDD[305] at checkpoint at command-1835760423149753:12 []
| ReliableCheckpointRDD[306] at checkpoint at command-1835760423149753:12 []
checkpointDir: String = dbfs:/dbfs/FileStore/checkpoint/cp1/loc10/86cc77b5-27c3-4049-9136-503ddcab0fa9
ds: org.apache.spark.sql.Dataset[Long] = [id: bigint]
ds2: org.apache.spark.sql.Dataset[Long] = [id: bigint]
res53: Boolean = false
Question 1:
ds.rdd.isCheckpointed or ds2.rdd.isCheckpointed both return False even though with count I have a non-lazy situation. Why, when in particular the ../loc 7 & 10 are written with (part) files? Also we can see that ReliableCheckPoint!
Not well explained the whole concept. Trying to sort this out.
Question 2 - secondary question:
Is the cache really necessary or not with latest versions of Spark 2.4? A new branch on the ds, if not cached, will it cause re-computation or is that better now? Seems odd the checkpoint data would not be used, or could we say Spark does not really know what is better?
From High Performance Spark I get the mixed impression that check pointing is not so recommended, but then again it is.
回答1:
TL;DR: You don't inspect the object that is actually checkpointed:
ds2.queryExecution.toRdd.dependencies(0).rdd.isCheckpointed
// Boolean = true
ds.rdd.isCheckpointed or ds2.rdd.isCheckpointed both return False
That is an expected behavior. The object being checkpointed is not the converted RDD (which is a result of the additional transformations required to convert to external representation), that you reference, but the internal RDD object (in fact, as you see above, it is not even the latest internal RDD, but its parent).
Additionally, in the first case you just use a wrong Dataset object whatsoever - as explained in the linked answer Dataset.checkpoint returns a new Dataset
even though with count I have a non-lazy situation
That doesn't make much sense. The default checkpoint implementation is eager, therefore it force evaluates. Even if it wasn't for that, Dataset.count is not the right way to force evaluation.
Is the cache really necessary or not with latest version
As you can see in the linked source, Dataset.checkpoint uses RDD.checkpoint internally so the same rule apply. However you already execute a separate action to force checkpoint, so additional caching, especially considering the cost of Dataset persistence, could be an overkill.
Of course, if in doubt, you might consider benchmarking in a specific context.
来源：https://stackoverflow.com/questions/54005223/spark-scala-checkpointing-data-set-showing-ischeckpointed-false-after-action