为什么使用数据集比RDD要慢得多

Question

我使用较新的数据集API在Java中实现了Pagerank的示例。当我针对使用较旧的RDD API的示例对代码进行基准测试时，我发现我的代码需要186秒，而基线仅需109秒。是什么导致差异？（旁注：即使数据库仅包含少数条目，Spark也需要数百秒钟？）

我的代码：

Dataset<Row> outLinks = spark.read().jdbc("jdbc:postgresql://127.0.0.1:5432/postgres", "storagepage_outlinks", props);
Dataset<Row> page = spark.read().jdbc("jdbc:postgresql://127.0.0.1:5432/postgres", "pages", props);
outLinks = page.join(outLinks, page.col("id").equalTo(outLinks.col("storagepage_id")));
outLinks = outLinks.distinct().groupBy(outLinks.col("url")).agg(collect_set("outlinks")).cache();
Dataset<Row> ranks = outLinks.map(row -> new Tuple2<>(row.getString(0), 1.0), Encoders.tuple(Encoders.STRING(), Encoders.DOUBLE())).toDF("url", "rank");
for (int i = 0; i < iterations; i++) {
    Dataset<Row> joined = outLinks.join(ranks, new Set.Set1<>("url").toSeq());
    Dataset<Row> contribs = joined.flatMap(row -> {
        List<String> links = row.getList(1);
        double rank = row.getDouble(2);
        return links
                .stream()
                .map(s -> new Tuple2<>(s, rank / links.size()))
                .collect(Collectors.toList()).iterator();
    }, Encoders.tuple(Encoders.STRING(), Encoders.DOUBLE())).toDF("url", "num");
    Dataset<Tuple2<String, Double>> reducedByKey =
            contribs.groupByKey(r -> r.getString(0), Encoders.STRING())
            .mapGroups((s, iterator) -> {
                double sum = 0;
                while (iterator.hasNext()) {
                    sum += iterator.next().getDouble(1);
                }
                return new Tuple2<>(s, sum);
            }, Encoders.tuple(Encoders.STRING(), Encoders.DOUBLE()));
    ranks = reducedByKey.map(t -> new Tuple2<>(t._1, .15 + t._2 * .85),
            Encoders.tuple(Encoders.STRING(), Encoders.DOUBLE())).toDF("url", "rank");
}
ranks.show();

使用RDD的示例代码（适用于我的数据库读取）：

Dataset<Row> outLinks = spark.read().jdbc("jdbc:postgresql://127.0.0.1:5432/postgres", "storagepage_outlinks", props);
Dataset<Row> page = spark.read().jdbc("jdbc:postgresql://127.0.0.1:5432/postgres", "pages", props);
outLinks = page.join(outLinks, page.col("id").equalTo(outLinks.col("storagepage_id")));
outLinks = outLinks.distinct().groupBy(outLinks.col("url")).agg(collect_set("outlinks")).cache(); // TODO: play with this cache
JavaPairRDD<String, Iterable<String>> links = outLinks.javaRDD().mapToPair(row -> new Tuple2<>(row.getString(0), row.getList(1)));
// Loads all URLs with other URL(s) link to from input file and initialize ranks of them to one.
JavaPairRDD<String, Double> ranks = links.mapValues(rs -> 1.0);
// Calculates and updates URL ranks continuously using PageRank algorithm.
for (int current = 0; current < 20; current++) {
    // Calculates URL contributions to the rank of other URLs.
    JavaPairRDD<String, Double> contribs = links.join(ranks).values()
            .flatMapToPair(s -> {
                int urlCount = size(s._1());
                List<Tuple2<String, Double>> results = new ArrayList<>();
                for (String n : s._1) {
                    results.add(new Tuple2<>(n, s._2() / urlCount));
                }
                return results.iterator();
            });
    // Re-calculates URL ranks based on neighbor contributions.
    ranks = contribs.reduceByKey((x, y) -> x + y).mapValues(sum -> 0.15 + sum * 0.85);
}
// Collects all URL ranks and dump them to console.
List<Tuple2<String, Double>> output = ranks.collect();
for (Tuple2<?,?> tuple : output) {
    System.out.println(tuple._1() + " has rank: " + tuple._2() + ".");
}

Answer 1

tl; dr 可能是很好的旧避免groupbykey。

很难说，但是您的Dataset代码等于groupByKey：

groupByKey(...).mapGroups(...)

这意味着它首先进行洗牌，然后减少数据。

您的RDD使用reduceByKey-这应该通过应用局部减少来降低洗牌尺寸。如果您希望此代码有些等效，则应使用groupByKey(...).reduceGroups(...)重写groupByKey(...).mapGroups(...)。

另一个可能的候选人是配置。spark.sql.shuffle.partitions的默认值为200，将用于Dataset聚合。如果

数据库仅包含少数条目？

这是一个严重的过度杀伤。

rdd将根据父数据使用spark.default.parallelism或值，通常更适度。