baohe-zhang commented on a change in pull request #29425:
URL: https://github.com/apache/spark/pull/29425#discussion_r476790437
##########
File path:
common/kvstore/src/main/java/org/apache/spark/util/kvstore/LevelDB.java
##########
@@ -164,35 +165,39 @@ public void writeAll(List<?> values) throws Exception {
Preconditions.checkArgument(values != null && !values.isEmpty(),
"Non-empty values required.");
- // Group by class, in case there are values from different classes in the
values
+ // Group by class, in case there are values from different classes in the
values.
// Typical usecase is for this to be a single class.
// A NullPointerException will be thrown if values contain null object.
for (Map.Entry<? extends Class<?>, ? extends List<?>> entry :
values.stream().collect(Collectors.groupingBy(Object::getClass)).entrySet()) {
-
- final Iterator<?> valueIter = entry.getValue().iterator();
- final Iterator<byte[]> serializedValueIter;
-
- // Deserialize outside synchronized block
- List<byte[]> list = new ArrayList<>(entry.getValue().size());
- for (Object value : values) {
- list.add(serializer.serialize(value));
- }
- serializedValueIter = list.iterator();
-
final Class<?> klass = entry.getKey();
- final LevelDBTypeInfo ti = getTypeInfo(klass);
- synchronized (ti) {
- final LevelDBTypeInfo.Index naturalIndex = ti.naturalIndex();
- final Collection<LevelDBTypeInfo.Index> indices = ti.indices();
+ // Partition the value list to a set of the 128-values batches. It can
reduce the
+ // memory pressure caused by serialization and give fairness to other
writing threads
+ // when writing a very large list.
+ for (List<?> batchList : Iterables.partition(entry.getValue(), 128)) {
+ final Iterator<?> valueIter = batchList.iterator();
+ final Iterator<byte[]> serializedValueIter;
+
+ // Deserialize outside synchronized block
+ List<byte[]> serializedValueList = new ArrayList<>(batchList.size());
+ for (Object value : batchList) {
+ serializedValueList.add(serializer.serialize(value));
+ }
+ serializedValueIter = serializedValueList.iterator();
+
+ final LevelDBTypeInfo ti = getTypeInfo(klass);
+ synchronized (ti) {
+ final LevelDBTypeInfo.Index naturalIndex = ti.naturalIndex();
+ final Collection<LevelDBTypeInfo.Index> indices = ti.indices();
- try (WriteBatch batch = db().createWriteBatch()) {
while (valueIter.hasNext()) {
- updateBatch(batch, valueIter.next(), serializedValueIter.next(),
klass,
- naturalIndex, indices);
+ try (WriteBatch batch = db().createWriteBatch()) {
+ updateBatch(batch, valueIter.next(), serializedValueIter.next(),
klass,
+ naturalIndex, indices);
+ db().write(batch);
Review comment:
@HeartSaVioR Here I put db().write(batch) inside the while loop, that
means we will call db().write() for every entity, in a way similar to
removeAllByIndexValues(). I did this change because after I added the unit
test, I found a bug that the count cannot be updated properly when we use the
original code of writeAll(). The count is updated in
https://github.com/apache/spark/blob/master/common/kvstore/src/main/java/org/apache/spark/util/kvstore/LevelDBTypeInfo.java#L395.
Since the delta is overridden to 1 every time we add a new entity, we can only
increase the count by 1 for every db().write(batch) call, even though we write
multiple entities in that call.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
