Skip to content

Commit

Permalink
Improve CostBalancerStrategy, deprecate cachingCost (apache#14484)
Browse files Browse the repository at this point in the history
Changes to `cost` strategy:
- In every `ServerHolder`, track the number of segments per datasource per interval
- Perform cost computations for a given interval just once, and then multiply by a constant
factor to account for the total segment count in that interval
- Do not perform joint cost computations with segments that are outside the compute interval
(± 45 days) for the segment being considered for move
- Remove metrics `segment/cost/*` as they were coordinator killers! Turning on these metrics
(by setting `emitBalancingStats` to true) has often caused the coordinator to be stuck for hours.
Moreover, they are too complicated to decipher and do not provide any meaningful insight into
a Druid cluster.
- Add new simpler metrics `segment/balancer/compute/*` to track cost computation time,
count and errors.

Other changes:
- Remove flaky test from `CostBalancerStrategyTest`.
- Add tests to verify that computed cost has remained unchanged
- Remove usages of mock `BalancerStrategy` from `LoadRuleTest`, `BalanceSegmentsTest`
- Clean up `BalancerStrategy` interface
  • Loading branch information
kfaraz authored and sergioferragut committed Jul 21, 2023
1 parent 165810a commit ca3be43
Show file tree
Hide file tree
Showing 28 changed files with 515 additions and 567 deletions.
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,7 @@
package org.apache.druid.server.coordinator;

import com.google.common.annotations.VisibleForTesting;
import com.google.common.base.Stopwatch;
import com.google.common.collect.ImmutableList;
import com.google.common.collect.Ordering;
import com.google.common.collect.Sets;
Expand Down Expand Up @@ -98,6 +99,7 @@
import java.util.Set;
import java.util.concurrent.ConcurrentHashMap;
import java.util.concurrent.ScheduledExecutorService;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicInteger;
import java.util.stream.Collectors;

Expand Down Expand Up @@ -672,7 +674,7 @@ public void run()
{
try {
log.info("Starting coordinator run for group [%s]", dutyGroupName);
final long globalStart = System.currentTimeMillis();
final Stopwatch groupRunTime = Stopwatch.createStarted();

synchronized (lock) {
if (!coordLeaderSelector.isLeader()) {
Expand Down Expand Up @@ -719,23 +721,25 @@ public void run()
log.info("Coordination has been paused. Duties will not run until coordination is resumed.");
}

final Stopwatch dutyRunTime = Stopwatch.createUnstarted();
for (CoordinatorDuty duty : duties) {
// Don't read state and run state in the same duty otherwise racy conditions may exist
if (!coordinationPaused
&& coordLeaderSelector.isLeader()
&& startingLeaderCounter == coordLeaderSelector.localTerm()) {

final long start = System.currentTimeMillis();
dutyRunTime.reset().start();
params = duty.run(params);
final long end = System.currentTimeMillis();
dutyRunTime.stop();

final String dutyName = duty.getClass().getName();
if (params == null) {
log.info("Stopping run for group [%s] on request of duty [%s].", dutyGroupName, dutyName);
return;
} else {
final RowKey rowKey = RowKey.builder().add(Dimension.DUTY, dutyName).build();
params.getCoordinatorStats().add(Stats.CoordinatorRun.DUTY_RUN_TIME, rowKey, end - start);
final RowKey rowKey = RowKey.of(Dimension.DUTY, dutyName);
final long dutyRunMillis = dutyRunTime.elapsed(TimeUnit.MILLISECONDS);
params.getCoordinatorStats().add(Stats.CoordinatorRun.DUTY_RUN_TIME, rowKey, dutyRunMillis);
}
}
}
Expand All @@ -745,9 +749,9 @@ public void run()
if (allStats.rowCount() > 0) {
final AtomicInteger emittedCount = new AtomicInteger();
allStats.forEachStat(
(dimensionValues, stat, value) -> {
(stat, dimensions, value) -> {
if (stat.shouldEmit()) {
emitStat(stat, dimensionValues, value);
emitStat(stat, dimensions.getValues(), value);
emittedCount.incrementAndGet();
}
}
Expand All @@ -760,7 +764,7 @@ public void run()
}

// Emit the runtime of the full DutiesRunnable
final long runMillis = System.currentTimeMillis() - globalStart;
final long runMillis = groupRunTime.stop().elapsed(TimeUnit.MILLISECONDS);
emitStat(Stats.CoordinatorRun.GROUP_RUN_TIME, Collections.emptyMap(), runMillis);
log.info("Finished coordinator run for group [%s] in [%d] ms", dutyGroupName, runMillis);
}
Expand All @@ -771,10 +775,6 @@ public void run()

private void emitStat(CoordinatorStat stat, Map<Dimension, String> dimensionValues, long value)
{
if (stat.equals(Stats.Balancer.NORMALIZED_COST_X_1000)) {
value = value / 1000;
}

ServiceMetricEvent.Builder eventBuilder = new ServiceMetricEvent.Builder()
.setDimension(Dimension.DUTY_GROUP.reportedName(), dutyGroupName);
dimensionValues.forEach(
Expand Down
Original file line number Diff line number Diff line change
@@ -0,0 +1,67 @@
/*
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you 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 org.apache.druid.server.coordinator;

import it.unimi.dsi.fastutil.objects.Object2IntMap;
import it.unimi.dsi.fastutil.objects.Object2IntMaps;
import it.unimi.dsi.fastutil.objects.Object2IntOpenHashMap;
import org.apache.druid.timeline.DataSegment;
import org.joda.time.Interval;

import java.util.HashMap;
import java.util.Map;

/**
* Maintains a count of segments for each datasource and interval.
*/
public class SegmentCountsPerInterval
{
private final Map<String, Object2IntMap<Interval>>
datasourceIntervalToSegmentCount = new HashMap<>();
private final Object2IntMap<Interval> intervalToTotalSegmentCount = new Object2IntOpenHashMap<>();

public void addSegment(DataSegment segment)
{
updateCountInInterval(segment, 1);
}

public void removeSegment(DataSegment segment)
{
updateCountInInterval(segment, -1);
}

public Object2IntMap<Interval> getIntervalToSegmentCount(String datasource)
{
return datasourceIntervalToSegmentCount.getOrDefault(datasource, Object2IntMaps.emptyMap());
}

public Object2IntMap<Interval> getIntervalToTotalSegmentCount()
{
return intervalToTotalSegmentCount;
}

private void updateCountInInterval(DataSegment segment, int delta)
{
intervalToTotalSegmentCount.mergeInt(segment.getInterval(), delta, Integer::sum);
datasourceIntervalToSegmentCount
.computeIfAbsent(segment.getDataSource(), ds -> new Object2IntOpenHashMap<>())
.mergeInt(segment.getInterval(), delta, Integer::sum);
}
}
Original file line number Diff line number Diff line change
Expand Up @@ -32,11 +32,9 @@
import java.util.Collection;
import java.util.Comparator;
import java.util.HashMap;
import java.util.HashSet;
import java.util.List;
import java.util.Map;
import java.util.Objects;
import java.util.Set;
import java.util.concurrent.atomic.AtomicInteger;

/**
Expand Down Expand Up @@ -75,11 +73,7 @@ public class ServerHolder implements Comparable<ServerHolder>
*/
private final Map<DataSegment, SegmentAction> queuedSegments = new HashMap<>();

/**
* Segments that are expected to be loaded on this server once all the
* operations in progress have completed.
*/
private final Set<DataSegment> projectedSegments = new HashSet<>();
private final SegmentCountsPerInterval projectedSegments = new SegmentCountsPerInterval();

public ServerHolder(ImmutableDruidServer server, LoadQueuePeon peon)
{
Expand Down Expand Up @@ -133,31 +127,31 @@ private void initializeQueuedSegments(
AtomicInteger loadingReplicaCount
)
{
projectedSegments.addAll(server.iterateAllSegments());
for (DataSegment segment : server.iterateAllSegments()) {
projectedSegments.addSegment(segment);
}

final List<SegmentHolder> expiredSegments = new ArrayList<>();
peon.getSegmentsInQueue().forEach(
(holder) -> {
int runsInQueue = holder.incrementAndGetRunsInQueue();
if (runsInQueue > maxLifetimeInQueue) {
expiredSegments.add(holder);
}

final SegmentAction action = holder.getAction();
addToQueuedSegments(holder.getSegment(), simplify(action));

if (action == SegmentAction.MOVE_TO) {
movingSegmentCount.incrementAndGet();
}
if (action == SegmentAction.REPLICATE) {
loadingReplicaCount.incrementAndGet();
}
}
);

peon.getSegmentsMarkedToDrop().forEach(
segment -> addToQueuedSegments(segment, SegmentAction.MOVE_FROM)
);
for (SegmentHolder holder : peon.getSegmentsInQueue()) {
int runsInQueue = holder.incrementAndGetRunsInQueue();
if (runsInQueue > maxLifetimeInQueue) {
expiredSegments.add(holder);
}

final SegmentAction action = holder.getAction();
addToQueuedSegments(holder.getSegment(), simplify(action));

if (action == SegmentAction.MOVE_TO) {
movingSegmentCount.incrementAndGet();
}
if (action == SegmentAction.REPLICATE) {
loadingReplicaCount.incrementAndGet();
}
}

for (DataSegment segment : peon.getSegmentsMarkedToDrop()) {
addToQueuedSegments(segment, SegmentAction.MOVE_FROM);
}

if (!expiredSegments.isEmpty()) {
List<SegmentHolder> expiredSegmentsSubList =
Expand Down Expand Up @@ -251,11 +245,21 @@ public Map<DataSegment, SegmentAction> getQueuedSegments()
* Segments that are expected to be loaded on this server once all the
* operations in progress have completed.
*/
public Set<DataSegment> getProjectedSegments()
public SegmentCountsPerInterval getProjectedSegments()
{
return projectedSegments;
}

public boolean isProjectedSegment(DataSegment segment)
{
SegmentAction action = getActionOnSegment(segment);
if (action == null) {
return hasSegmentLoaded(segment.getId());
} else {
return action.isLoad();
}
}

/**
* Segments that are currently in the queue for being loaded on this server.
* This does not include segments that are being moved to this server.
Expand Down Expand Up @@ -362,10 +366,10 @@ private void addToQueuedSegments(DataSegment segment, SegmentAction action)

// Add to projected if load is started, remove from projected if drop has started
if (action.isLoad()) {
projectedSegments.add(segment);
projectedSegments.addSegment(segment);
sizeOfLoadingSegments += segment.getSize();
} else {
projectedSegments.remove(segment);
projectedSegments.removeSegment(segment);
if (action == SegmentAction.DROP) {
sizeOfDroppingSegments += segment.getSize();
}
Expand All @@ -379,10 +383,10 @@ private void removeFromQueuedSegments(DataSegment segment, SegmentAction action)
queuedSegments.remove(segment);

if (action.isLoad()) {
projectedSegments.remove(segment);
projectedSegments.removeSegment(segment);
sizeOfLoadingSegments -= segment.getSize();
} else {
projectedSegments.add(segment);
projectedSegments.addSegment(segment);
if (action == SegmentAction.DROP) {
sizeOfDroppingSegments -= segment.getSize();
}
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -20,70 +20,60 @@
package org.apache.druid.server.coordinator.balancer;

import org.apache.druid.server.coordinator.ServerHolder;
import org.apache.druid.server.coordinator.duty.BalanceSegments;
import org.apache.druid.server.coordinator.stats.CoordinatorRunStats;
import org.apache.druid.timeline.DataSegment;

import javax.annotation.Nullable;
import java.util.Iterator;
import java.util.List;
import java.util.NavigableSet;

/**
* This interface describes the coordinator balancing strategy, which is responsible for making decisions on where
* to place {@link DataSegment}s on historical servers (described by {@link ServerHolder}). The balancing strategy
* is used by {@link org.apache.druid.server.coordinator.rules.LoadRule} to assign and drop segments, and by
* {@link BalanceSegments} to migrate segments between historicals.
* Segment balancing strategy, used in every coordinator run by
* {@link org.apache.druid.server.coordinator.loading.StrategicSegmentAssigner}
* to choose optimal servers to load, move or drop a segment.
*/
public interface BalancerStrategy
{

/**
* Finds the best server to move a segment to according to the balancing strategy.
* Finds the best server from the list of {@code destinationServers} to load
* the {@code segmentToMove}, if it is moved from the {@code sourceServer}.
* <p>
* In order to avoid unnecessary moves when the segment is already optimally placed,
* include the {@code sourceServer} in the list of {@code destinationServers}.
*
* @param proposalSegment segment to move
* @param sourceServer Server the segment is currently placed on.
* @param destinationServers servers to consider as move destinations
* @return The server to move to, or null if no move should be made or no server is suitable
* @return The server to move to, or null if the segment is already optimally placed.
*/
@Nullable
ServerHolder findDestinationServerToMoveSegment(
DataSegment proposalSegment,
DataSegment segmentToMove,
ServerHolder sourceServer,
List<ServerHolder> destinationServers
);

/**
* Finds the best servers on which to place the {@code proposalSegment}.
* This method can be used both for placing the first copy of a segment
* in the tier or a replica of the segment.
* Finds the best servers to load the given segment. This method can be used
* both for placing the first copy of a segment in a tier or a replica of an
* already available segment.
*
* @param proposalSegment segment to place on servers
* @param serverHolders servers to consider as segment homes
* @return Iterator over the best servers (in order) on which the segment
* can be placed.
* @return Iterator over the best servers (in order of preference) to load
* the segment.
*/
Iterator<ServerHolder> findServersToLoadSegment(
DataSegment proposalSegment,
DataSegment segmentToLoad,
List<ServerHolder> serverHolders
);

/**
* Returns an iterator for a set of servers to drop from, ordered by preference of which server to drop from first
* for a given drop strategy. One or more segments may be dropped, depending on how much the segment is
* over-replicated.
* @param toDropSegment segment to drop from one or more servers
* @param serverHolders set of historicals to consider dropping from
* @return Iterator for set of historicals, ordered by drop preference
* Finds the best servers to drop the given segment.
*
* @return Iterator over the servers (in order of preference) to drop the segment
*/
Iterator<ServerHolder> pickServersToDropSegment(DataSegment toDropSegment, NavigableSet<ServerHolder> serverHolders);
Iterator<ServerHolder> findServersToDropSegment(DataSegment segmentToDrop, List<ServerHolder> serverHolders);

/**
* Add balancing strategy stats during the 'balanceTier' operation of
* {@link BalanceSegments} to be included
* @param tier historical tier being balanced
* @param stats stats object to add balancing strategy stats to
* @param serverHolderList servers in tier being balanced
* Returns the stats collected by the strategy in the current run and resets
* the stats collector for the next run.
*/
void emitStats(String tier, CoordinatorRunStats stats, List<ServerHolder> serverHolderList);
CoordinatorRunStats getAndResetStats();
}
Loading

0 comments on commit ca3be43

Please sign in to comment.