[SPARK-3623] Graph should support the checkpoint operation

• Type: Improvement
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 1.0.2, 1.1.0
• Fix Version/s: 1.2.0
• Component/s: GraphX
• Assignee: Guoqiang Li
• Reporter: Guoqiang Li
• Created: 21/Sep/14 08:13
• Updated: 06/Dec/14 08:58
• Resolved: 06/Dec/14 08:57

Description

Consider the following code:

    for (i <- 0 until totalIter) {
      val previousCorpus = corpus
      logInfo("Start Gibbs sampling (Iteration %d/%d)".format(i, totalIter))

      val corpusTopicDist = collectTermTopicDist(corpus, globalTopicCounter, sumTerms,
        numTerms, numTopics, alpha, beta).persist(storageLevel)

      val corpusSampleTopics = sampleTopics(corpusTopicDist, globalTopicCounter, sumTerms, numTerms,
        numTopics, alpha, beta).persist(storageLevel)

      corpus = updateCounter(corpusSampleTopics, numTopics).persist(storageLevel)

      globalTopicCounter = collectGlobalCounter(corpus, numTopics)
      assert(bsum(globalTopicCounter) == sumTerms)
      previousCorpus.unpersistVertices()
      corpusTopicDist.unpersistVertices()
      corpusSampleTopics.unpersistVertices()
    }

If there is no checkpoint operation will appear the following problems.
1. The RDD of corpus dependencies are too deep
2. The shuffle files are too large.
3. Any of a server crash will cause the algorithm to recalculate

Issue Links

• is related to: SPARK-4672 Cut off the super long serialization chain in GraphX to avoid the StackOverflow error
• relates to: SPARK-3625 In some cases, the RDD.checkpoint does not work
• links to: [Github] Pull Request #2631 (witgo)