io.trino.operator.join.LookupJoinPageBuilder Maven / Gradle / Ivy
/*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package io.trino.operator.join;
import io.trino.spi.Page;
import io.trino.spi.PageBuilder;
import io.trino.spi.block.Block;
import io.trino.spi.type.Type;
import it.unimi.dsi.fastutil.ints.IntArrayList;
import java.util.List;
import static com.google.common.base.MoreObjects.toStringHelper;
import static com.google.common.base.Verify.verify;
import static io.trino.operator.project.PageProcessor.MAX_BATCH_SIZE;
import static io.trino.spi.block.PageBuilderStatus.DEFAULT_MAX_PAGE_SIZE_IN_BYTES;
import static java.util.Objects.requireNonNull;
/**
* This page builder creates pages with dictionary blocks:
* normal dictionary blocks for the probe side and the original blocks for the build side.
*
* TODO use dictionary blocks (probably extended kind) to avoid data copying for build side
*/
public class LookupJoinPageBuilder
{
private final IntArrayList probeIndexBuilder = new IntArrayList();
private final PageBuilder buildPageBuilder;
private final int buildOutputChannelCount;
private long estimatedProbeBlockBytes;
private long estimatedProbeRowSize = -1;
private int previousPosition = -1;
private boolean isSequentialProbeIndices = true;
public LookupJoinPageBuilder(List buildTypes)
{
this.buildPageBuilder = new PageBuilder(requireNonNull(buildTypes, "buildTypes is null"));
this.buildOutputChannelCount = buildTypes.size();
}
public boolean isFull()
{
return estimatedProbeBlockBytes + buildPageBuilder.getSizeInBytes() >= DEFAULT_MAX_PAGE_SIZE_IN_BYTES ||
buildPageBuilder.getPositionCount() >= MAX_BATCH_SIZE ||
buildPageBuilder.isFull();
}
public boolean isEmpty()
{
return probeIndexBuilder.isEmpty() && buildPageBuilder.isEmpty();
}
public void reset()
{
// be aware that probeIndexBuilder will not clear its capacity
probeIndexBuilder.clear();
buildPageBuilder.reset();
estimatedProbeBlockBytes = 0;
estimatedProbeRowSize = -1;
previousPosition = -1;
isSequentialProbeIndices = true;
}
/**
* append the index for the probe and copy the row for the build
*/
public void appendRow(JoinProbe probe, LookupSource lookupSource, long joinPosition)
{
// probe side
appendProbeIndex(probe);
// build side
buildPageBuilder.declarePosition();
lookupSource.appendTo(joinPosition, buildPageBuilder, 0);
}
/**
* append the index for the probe and append nulls for the build
*/
public void appendNullForBuild(JoinProbe probe)
{
// probe side
appendProbeIndex(probe);
// build side
buildPageBuilder.declarePosition();
for (int i = 0; i < buildOutputChannelCount; i++) {
buildPageBuilder.getBlockBuilder(i).appendNull();
}
}
public Page build(JoinProbe probe)
{
int outputPositions = probeIndexBuilder.size();
verify(buildPageBuilder.getPositionCount() == outputPositions);
int[] probeOutputChannels = probe.getOutputChannels();
Block[] blocks = new Block[probeOutputChannels.length + buildOutputChannelCount];
Page probePage = probe.getPage();
if (!isSequentialProbeIndices || outputPositions == 0) {
int[] probeIndices = probeIndexBuilder.toIntArray();
for (int i = 0; i < probeOutputChannels.length; i++) {
blocks[i] = unwrapLoadedBlock(probePage.getBlock(probeOutputChannels[i]).getPositions(probeIndices, 0, outputPositions));
}
}
else {
// probeIndices are sequential without holes
int startRegion = probeIndexBuilder.getInt(0);
verify(previousPosition - startRegion == outputPositions - 1);
// probeIndices are a simple covering of the block, output the probe block directly
boolean outputProbeBlocksDirectly = startRegion == 0 && outputPositions == probePage.getPositionCount();
for (int i = 0; i < probeOutputChannels.length; i++) {
Block block = probePage.getBlock(probeOutputChannels[i]);
if (!outputProbeBlocksDirectly) {
// only a subregion of the block should be output
block = block.getRegion(startRegion, outputPositions);
}
blocks[i] = unwrapLoadedBlock(block);
}
}
int offset = probeOutputChannels.length;
for (int i = 0; i < buildOutputChannelCount; i++) {
blocks[offset + i] = buildPageBuilder.getBlockBuilder(i).build();
verify(blocks[offset + i].getPositionCount() == outputPositions);
}
return new Page(outputPositions, blocks);
}
@Override
public String toString()
{
return toStringHelper(this)
.add("estimatedSize", estimatedProbeBlockBytes + buildPageBuilder.getSizeInBytes())
.add("positionCount", buildPageBuilder.getPositionCount())
.toString();
}
private static Block unwrapLoadedBlock(Block filteredProbeBlock)
{
// Lazy blocks (e.g. used in filter condition) could be loaded during filter evaluation.
// Unwrap them to reduce overhead of further processing.
return filteredProbeBlock.isLoaded() ? filteredProbeBlock.getLoadedBlock() : filteredProbeBlock;
}
private void appendProbeIndex(JoinProbe probe)
{
int position = probe.getPosition();
// positions to be appended should be in ascending order
verify(position >= 0 && previousPosition <= position);
isSequentialProbeIndices &= position == previousPosition + 1 || previousPosition == -1;
// Update probe indices and size
probeIndexBuilder.add(position);
estimatedProbeBlockBytes += Integer.BYTES;
// Update memory usage for probe side.
//
// The size of the probe cannot be easily calculated given
// (1) the structure of Block is recursive,
// (2) an inner block can serve as multiple views (e.g., in a dictionary block).
// Without a dedup at the granularity of rows, we cannot tell if we are overcounting, and
// (3) even we are able to dedup magically, calling getRegionSizeInBytes can be expensive.
// For example, consider a dictionary block inside an array block;
// calling getRegionSizeInBytes(p, 1) of the array block can lead to calling getRegionSizeInBytes with an arbitrary length for the dictionary block,
// which is very expensive.
//
// To workaround the memory accounting complexity yet having a relatively reasonable estimation, we use sizeInBytes / positionCount as the size for each row.
// It can be shown that the output page is bounded within range [buildPageBuilder.getSizeInBytes(), buildPageBuilder.getSizeInBytes + probe.getPage().getSizeInBytes()].
//
// This is under the assumption that the position of a probe is non-decreasing.
// if position > previousPosition, we know it is a new row to append and we accumulate the estimated row size (sizeInBytes / positionCount);
// otherwise we do not count because we know it is duplicated with the previous appended row.
// So in the worst case, we can only accumulate up to the sizeInBytes of the probe page.
//
// On the other hand, we do not want to produce a page that is too small if the build size is too small (e.g., the build side is with all nulls).
// That means we only appended a few small rows in the probe and reached the probe end.
// But that is going to happen anyway because we have to flush the page whenever we reach the probe end.
// So with or without precise memory accounting, the output page is small anyway.
if (previousPosition != position) {
previousPosition = position;
estimatedProbeBlockBytes += getEstimatedProbeRowSize(probe);
}
}
private long getEstimatedProbeRowSize(JoinProbe probe)
{
if (estimatedProbeRowSize != -1) {
return estimatedProbeRowSize;
}
long estimatedProbeRowSize = 0;
for (int index : probe.getOutputChannels()) {
Block block = probe.getPage().getBlock(index);
// Estimate the size of the probe row
// TODO: improve estimation for unloaded blocks by making it similar as in PageProcessor
estimatedProbeRowSize += block.getSizeInBytes() / block.getPositionCount();
}
this.estimatedProbeRowSize = estimatedProbeRowSize;
return estimatedProbeRowSize;
}
}