[HIVE-15527] Memory usage is unbound in SortByShuffler for Spark

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Not A Problem
• Affects Version/s: 1.1.0
• Component/s: Spark
• Assignee: Chao Sun
• Reporter: Xuefu Zhang
• Created: 31/Dec/16 06:02
• Updated: 20/Jan/17 21:00
• Resolved: 20/Jan/17 21:00

Description

In SortByShuffler.java, an ArrayList is used to back the iterator for values that have the same key in shuffled result produced by spark transformation sortByKey. It's possible that memory can be exhausted because of a large key group.

            @Override
            public Tuple2<HiveKey, Iterable<BytesWritable>> next() {
              // TODO: implement this by accumulating rows with the same key into a list.
              // Note that this list needs to improved to prevent excessive memory usage, but this
              // can be done in later phase.
              while (it.hasNext()) {
                Tuple2<HiveKey, BytesWritable> pair = it.next();
                if (curKey != null && !curKey.equals(pair._1())) {
                  HiveKey key = curKey;
                  List<BytesWritable> values = curValues;
                  curKey = pair._1();
                  curValues = new ArrayList<BytesWritable>();
                  curValues.add(pair._2());
                  return new Tuple2<HiveKey, Iterable<BytesWritable>>(key, values);
                }
                curKey = pair._1();
                curValues.add(pair._2());
              }
              if (curKey == null) {
                throw new NoSuchElementException();
              }
              // if we get here, this should be the last element we have
              HiveKey key = curKey;
              curKey = null;
              return new Tuple2<HiveKey, Iterable<BytesWritable>>(key, curValues);
            }

Since the output from sortByKey is already sorted on key, it's possible to backup the value iterable using the same input iterator.

Issue Links

• is part of: HIVE-15580 Eliminate unbounded memory usage for orderBy and groupBy in Hive on Spark