[
https://issues.apache.org/jira/browse/DRILL-6385?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16529168#comment-16529168
]
ASF GitHub Bot commented on DRILL-6385:
---------------------------------------
amansinha100 commented on a change in pull request #1334: DRILL-6385: Support
JPPD feature
URL: https://github.com/apache/drill/pull/1334#discussion_r199354240
##########
File path:
exec/java-exec/src/main/java/org/apache/drill/exec/physical/impl/ScanBatch.java
##########
@@ -226,6 +244,96 @@ public IterOutcome next() {
}
}
+ private void applyRuntimeFilter() throws SchemaChangeException {
+ RuntimeFilterWritable runtimeFilterWritable = context.getRuntimeFilter();
+ if (runtimeFilterWritable == null) {
+ return;
+ }
+ if (recordCount <= 0) {
+ return;
+ }
+ List<BloomFilter> bloomFilters = runtimeFilterWritable.unwrap();
+ if (hash64 == null) {
+ ValueVectorHashHelper hashHelper = new ValueVectorHashHelper(this,
context);
+ try {
+ //generate hash helper
+ this.toFilterFields =
runtimeFilterWritable.getRuntimeFilterBDef().getProbeFieldsList();
+ List<LogicalExpression> hashFieldExps = new ArrayList<>();
+ List<TypedFieldId> typedFieldIds = new ArrayList<>();
+ for (String toFilterField : toFilterFields) {
+ SchemaPath schemaPath = new SchemaPath(new
PathSegment.NameSegment(toFilterField), ExpressionPosition.UNKNOWN);
+ TypedFieldId typedFieldId = container.getValueVectorId(schemaPath);
+ this.field2id.put(toFilterField, typedFieldId.getFieldIds()[0]);
+ typedFieldIds.add(typedFieldId);
+ ValueVectorReadExpression toHashFieldExp = new
ValueVectorReadExpression(typedFieldId);
+ hashFieldExps.add(toHashFieldExp);
+ }
+ hash64 = hashHelper.getHash64(hashFieldExps.toArray(new
LogicalExpression[hashFieldExps.size()]), typedFieldIds.toArray(new
TypedFieldId[typedFieldIds.size()]));
+ } catch (Exception e) {
+ throw UserException.internalError(e).build(logger);
+ }
+ }
+ selectionVector2.allocateNew(recordCount);
+ BitSet bitSet = new BitSet(recordCount);
+ for (int i = 0; i < toFilterFields.size(); i++) {
+ BloomFilter bloomFilter = bloomFilters.get(i);
+ String fieldName = toFilterFields.get(i);
+ computeBitSet(field2id.get(fieldName), bloomFilter, bitSet);
+ }
+ int svIndex = 0;
+ int tmpFilterRows = 0;
+ for (int i = 0; i < recordCount; i++) {
+ boolean contain = bitSet.get(i);
+ if (contain) {
+ selectionVector2.setIndex(svIndex, i);
+ svIndex++;
+ } else {
+ tmpFilterRows++;
+ }
+ }
+ selectionVector2.setRecordCount(svIndex);
+ if (tmpFilterRows > 0 && tmpFilterRows == recordCount) {
+ recordCount = 0;
+ selectionVector2.clear();
+ logger.debug("filter {} rows by the RuntimeFilter", tmpFilterRows);
+ return;
+ }
+ if (tmpFilterRows > 0 && tmpFilterRows != recordCount ) {
+ totalFilterRows = totalFilterRows + tmpFilterRows;
+ recordCount = svIndex;
+ BatchSchema batchSchema = this.schema;
+ VectorContainer backUpContainer = new
VectorContainer(this.oContext.getAllocator(), batchSchema);
+ int fieldCount = batchSchema.getFieldCount();
+ for (int i = 0; i < fieldCount; i++) {
+ ValueVector from =
this.getContainer().getValueVector(i).getValueVector();
+ ValueVector to = backUpContainer.getValueVector(i).getValueVector();
+ to.setInitialCapacity(svIndex);
+ for (int r = 0; r < svIndex; r++) {
+ to.copyEntry(r, from, selectionVector2.getIndex(r));
Review comment:
To summarize the way you have implemented this part: suppose the original
ScanBatch contained 100 rows and 10 of them qualified the bloom filter, you
create an SV2 of size 10, set the original qualifying row's index in the SV2,
then for each ValueVector you copy the qualifying row's data into the
backupContainer, followed by exchanging it with the current output container.
A couple of thoughts about this:
- Why not produce the SV2 in the output batch and let the downstream
operator handle it ? Quite often there may be a Filter operator above the Scan
which would be applying other filters (i.e not the run-time filters) and it can
combine its own SV2 with the SV2 produced by the Scan. That way you avoid the
extra copying and let Filter handle it (note that Filter does code-gen so it is
more efficient way to handle bulk filtering).
- There is clearly trade-offs with the extra copy approach : it depends on
selectivity, i.e how many rows actually get eliminated by the run-time filter.
I suppose you have mentioned this as a TODO depending on the NDV statistics ?
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
> Support JPPD (Join Predicate Push Down)
> ---------------------------------------
>
> Key: DRILL-6385
> URL: https://issues.apache.org/jira/browse/DRILL-6385
> Project: Apache Drill
> Issue Type: New Feature
> Components: Server, Execution - Flow
> Affects Versions: 1.14.0
> Reporter: weijie.tong
> Assignee: weijie.tong
> Priority: Major
>
> This feature is to support the JPPD (Join Predicate Push Down). It will
> benefit the HashJoin ,Broadcast HashJoin performance by reducing the number
> of rows to send across the network ,the memory consumed. This feature is
> already supported by Impala which calls it RuntimeFilter
> ([https://www.cloudera.com/documentation/enterprise/5-9-x/topics/impala_runtime_filtering.html]).
> The first PR will try to push down a bloom filter of HashJoin node to
> Parquet’s scan node. The propose basic procedure is described as follow:
> # The HashJoin build side accumulate the equal join condition rows to
> construct a bloom filter. Then it sends out the bloom filter to the foreman
> node.
> # The foreman node accept the bloom filters passively from all the fragments
> that has the HashJoin operator. It then aggregates the bloom filters to form
> a global bloom filter.
> # The foreman node broadcasts the global bloom filter to all the probe side
> scan nodes which maybe already have send out partial data to the hash join
> nodes(currently the hash join node will prefetch one batch from both sides ).
> 4. The scan node accepts a global bloom filter from the foreman node.
> It will filter the rest rows satisfying the bloom filter.
>
> To implement above execution flow, some main new notion described as below:
> 1. RuntimeFilter
> It’s a filter container which may contain BloomFilter or MinMaxFilter.
> 2. RuntimeFilterReporter
> It wraps the logic to send hash join’s bloom filter to the foreman.The
> serialized bloom filter will be sent out through the data tunnel.This object
> will be instanced by the FragmentExecutor and passed to the
> FragmentContext.So the HashJoin operator can obtain it through the
> FragmentContext.
> 3. RuntimeFilterRequestHandler
> It is responsible to accept a SendRuntimeFilterRequest RPC to strip the
> actual BloomFilter from the network. It then translates this filter to the
> WorkerBee’s new interface registerRuntimeFilter.
> Another RPC type is BroadcastRuntimeFilterRequest. It will register the
> accepted global bloom filter to the WorkerBee by the registerRuntimeFilter
> method and then propagate to the FragmentContext through which the probe side
> scan node can fetch the aggregated bloom filter.
> 4.RuntimeFilterManager
> The foreman will instance a RuntimeFilterManager .It will indirectly get
> every RuntimeFilter by the WorkerBee. Once all the BloomFilters have been
> accepted and aggregated . It will broadcast the aggregated bloom filter to
> all the probe side scan nodes through the data tunnel by a
> BroadcastRuntimeFilterRequest RPC.
> 5. RuntimeFilterEnableOption
> A global option will be added to decide whether to enable this new feature.
>
> Welcome suggestion and advice from you.The related PR will be presented as
> soon as possible.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
