Replace multipart download with parallel file download (opensearch-pr…

…oject#10519) (opensearch-project#10549) There are a few open issues with the multi-stream download approach: - Recovery stats are not being reported correctly - It is incompatible (short of reopening and re-reading the entire file) with the existing Lucene checksum validation logic - There are some issues with integrating it with the pending client side encryption work Given this, I attempted an experiment where I replaced with multi-stream-within-a-single-file approach with simply parallelizing downloads across files (this is how snapshot restore works). I actually got better results with this approach: recovering a ~52GiB shard took about 4.7 minutes with the multi-stream code versus 3.9 minutes with the parallel file approach (r7g.4xlarge EC2 instance, 500MiB/s EBS volume, S3 as remote repository). I think this is the right approach as it leverages the more battle-tested code path and addresses the three issues listed above. The multi-stream approach still has promise as it will allow us to download very large files faster (whereas this approach they can be the long poll on the transfer operation). However, given that 5GB segments (made up of multiple files in practice) are the norm, we generally aren't dealing with huge files. (cherry picked from commit 00ccfc4) Signed-off-by: Andrew Ross <[email protected]>
ashking94 · Oct 11, 2023 · 2997702 · 2997702
1 parent 38df519
commit 2997702
Show file tree

Hide file tree

Showing 20 changed files with 333 additions and 312 deletions.
diff --git a/server/src/main/java/org/opensearch/index/IndexModule.java b/server/src/main/java/org/opensearch/index/IndexModule.java
@@ -81,6 +81,7 @@
 import org.opensearch.indices.IndicesQueryCache;
 import org.opensearch.indices.fielddata.cache.IndicesFieldDataCache;
 import org.opensearch.indices.mapper.MapperRegistry;
+import org.opensearch.indices.recovery.RecoverySettings;
 import org.opensearch.indices.recovery.RecoveryState;
 import org.opensearch.plugins.IndexStorePlugin;
 import org.opensearch.repositories.RepositoriesService;
@@ -602,7 +603,8 @@ public IndexService newIndexService(
         IndexStorePlugin.DirectoryFactory remoteDirectoryFactory,
         BiFunction<IndexSettings, ShardRouting, TranslogFactory> translogFactorySupplier,
         Supplier<TimeValue> clusterDefaultRefreshIntervalSupplier,
-        Supplier<TimeValue> clusterRemoteTranslogBufferIntervalSupplier
+        Supplier<TimeValue> clusterRemoteTranslogBufferIntervalSupplier,
+        RecoverySettings recoverySettings
     ) throws IOException {
         final IndexEventListener eventListener = freeze();
         Function<IndexService, CheckedFunction<DirectoryReader, DirectoryReader, IOException>> readerWrapperFactory = indexReaderWrapper
@@ -660,7 +662,8 @@ public IndexService newIndexService(
                 recoveryStateFactory,
                 translogFactorySupplier,
                 clusterDefaultRefreshIntervalSupplier,
-                clusterRemoteTranslogBufferIntervalSupplier
+                clusterRemoteTranslogBufferIntervalSupplier,
+                recoverySettings
             );
             success = true;
             return indexService;

diff --git a/server/src/main/java/org/opensearch/index/IndexService.java b/server/src/main/java/org/opensearch/index/IndexService.java
@@ -89,13 +89,13 @@
 import org.opensearch.index.shard.ShardNotInPrimaryModeException;
 import org.opensearch.index.shard.ShardPath;
 import org.opensearch.index.similarity.SimilarityService;
-import org.opensearch.index.store.RemoteSegmentStoreDirectoryFactory;
 import org.opensearch.index.store.Store;
 import org.opensearch.index.translog.Translog;
 import org.opensearch.index.translog.TranslogFactory;
 import org.opensearch.indices.cluster.IndicesClusterStateService;
 import org.opensearch.indices.fielddata.cache.IndicesFieldDataCache;
 import org.opensearch.indices.mapper.MapperRegistry;
+import org.opensearch.indices.recovery.RecoverySettings;
 import org.opensearch.indices.recovery.RecoveryState;
 import org.opensearch.indices.replication.checkpoint.SegmentReplicationCheckpointPublisher;
 import org.opensearch.plugins.IndexStorePlugin;
@@ -179,6 +179,7 @@ public class IndexService extends AbstractIndexComponent implements IndicesClust
     private final BiFunction<IndexSettings, ShardRouting, TranslogFactory> translogFactorySupplier;
     private final Supplier<TimeValue> clusterDefaultRefreshIntervalSupplier;
     private final Supplier<TimeValue> clusterRemoteTranslogBufferIntervalSupplier;
+    private final RecoverySettings recoverySettings;
 
     public IndexService(
         IndexSettings indexSettings,
@@ -213,7 +214,8 @@ public IndexService(
         IndexStorePlugin.RecoveryStateFactory recoveryStateFactory,
         BiFunction<IndexSettings, ShardRouting, TranslogFactory> translogFactorySupplier,
         Supplier<TimeValue> clusterDefaultRefreshIntervalSupplier,
-        Supplier<TimeValue> clusterRemoteTranslogBufferIntervalSupplier
+        Supplier<TimeValue> clusterRemoteTranslogBufferIntervalSupplier,
+        RecoverySettings recoverySettings
     ) {
         super(indexSettings);
         this.allowExpensiveQueries = allowExpensiveQueries;
@@ -290,6 +292,7 @@ public IndexService(
         this.retentionLeaseSyncTask = new AsyncRetentionLeaseSyncTask(this);
         this.translogFactorySupplier = translogFactorySupplier;
         this.clusterRemoteTranslogBufferIntervalSupplier = clusterRemoteTranslogBufferIntervalSupplier;
+        this.recoverySettings = recoverySettings;
         updateFsyncTaskIfNecessary();
     }
 
@@ -522,7 +525,7 @@ public synchronized IndexShard createShard(
                 remoteStoreStatsTrackerFactory,
                 clusterRemoteTranslogBufferIntervalSupplier,
                 nodeEnv.nodeId(),
-                (RemoteSegmentStoreDirectoryFactory) remoteDirectoryFactory
+                recoverySettings
             );
             eventListener.indexShardStateChanged(indexShard, null, indexShard.state(), "shard created");
             eventListener.afterIndexShardCreated(indexShard);

diff --git a/server/src/main/java/org/opensearch/index/shard/IndexShard.java b/server/src/main/java/org/opensearch/index/shard/IndexShard.java
@@ -63,7 +63,6 @@
 import org.opensearch.action.admin.indices.flush.FlushRequest;
 import org.opensearch.action.admin.indices.forcemerge.ForceMergeRequest;
 import org.opensearch.action.admin.indices.upgrade.post.UpgradeRequest;
-import org.opensearch.action.support.PlainActionFuture;
 import org.opensearch.action.support.replication.PendingReplicationActions;
 import org.opensearch.action.support.replication.ReplicationResponse;
 import org.opensearch.cluster.metadata.DataStream;
@@ -162,9 +161,8 @@
 import org.opensearch.index.seqno.SequenceNumbers;
 import org.opensearch.index.shard.PrimaryReplicaSyncer.ResyncTask;
 import org.opensearch.index.similarity.SimilarityService;
-import org.opensearch.index.store.DirectoryFileTransferTracker;
 import org.opensearch.index.store.RemoteSegmentStoreDirectory;
-import org.opensearch.index.store.RemoteSegmentStoreDirectoryFactory;
+import org.opensearch.index.store.RemoteStoreFileDownloader;
 import org.opensearch.index.store.Store;
 import org.opensearch.index.store.Store.MetadataSnapshot;
 import org.opensearch.index.store.StoreFileMetadata;
@@ -185,6 +183,7 @@
 import org.opensearch.indices.recovery.PeerRecoveryTargetService;
 import org.opensearch.indices.recovery.RecoveryFailedException;
 import org.opensearch.indices.recovery.RecoveryListener;
+import org.opensearch.indices.recovery.RecoverySettings;
 import org.opensearch.indices.recovery.RecoveryState;
 import org.opensearch.indices.recovery.RecoveryTarget;
 import org.opensearch.indices.replication.checkpoint.ReplicationCheckpoint;
@@ -202,7 +201,6 @@
 import java.nio.channels.FileChannel;
 import java.nio.charset.StandardCharsets;
 import java.nio.file.NoSuchFileException;
-import java.nio.file.Path;
 import java.util.ArrayList;
 import java.util.Arrays;
 import java.util.Collections;
@@ -344,7 +342,7 @@ Runnable getGlobalCheckpointSyncer() {
     private final RemoteStoreStatsTrackerFactory remoteStoreStatsTrackerFactory;
 
     private final List<ReferenceManager.RefreshListener> internalRefreshListener = new ArrayList<>();
-    private final RemoteSegmentStoreDirectoryFactory remoteSegmentStoreDirectoryFactory;
+    private final RemoteStoreFileDownloader fileDownloader;
 
     public IndexShard(
         final ShardRouting shardRouting,
@@ -373,10 +371,7 @@ public IndexShard(
         final RemoteStoreStatsTrackerFactory remoteStoreStatsTrackerFactory,
         final Supplier<TimeValue> clusterRemoteTranslogBufferIntervalSupplier,
         final String nodeId,
-        // Wiring a directory factory here breaks some intended abstractions, but this remote directory
-        // factory is used not as a Lucene directory but instead to copy files from a remote store when
-        // restoring a shallow snapshot.
-        @Nullable final RemoteSegmentStoreDirectoryFactory remoteSegmentStoreDirectoryFactory
+        final RecoverySettings recoverySettings
     ) throws IOException {
         super(shardRouting.shardId(), indexSettings);
         assert shardRouting.initializing();
@@ -472,7 +467,7 @@ public boolean shouldCache(Query query) {
             ? false
             : mapperService.documentMapper().mappers().containsTimeStampField();
         this.remoteStoreStatsTrackerFactory = remoteStoreStatsTrackerFactory;
-        this.remoteSegmentStoreDirectoryFactory = remoteSegmentStoreDirectoryFactory;
+        this.fileDownloader = new RemoteStoreFileDownloader(shardRouting.shardId(), threadPool, recoverySettings);
     }
 
     public ThreadPool getThreadPool() {
@@ -570,6 +565,10 @@ public String getNodeId() {
         return translogConfig.getNodeId();
     }
 
+    public RemoteStoreFileDownloader getFileDownloader() {
+        return fileDownloader;
+    }
+
     @Override
     public void updateShardState(
         final ShardRouting newRouting,
@@ -2711,15 +2710,15 @@ public void restoreFromRemoteStore(ActionListener<Boolean> listener) {
 
     public void restoreFromSnapshotAndRemoteStore(
         Repository repository,
-        RemoteSegmentStoreDirectoryFactory remoteSegmentStoreDirectoryFactory,
+        RepositoriesService repositoriesService,
         ActionListener<Boolean> listener
     ) {
         try {
             assert shardRouting.primary() : "recover from store only makes sense if the shard is a primary shard";
             assert recoveryState.getRecoverySource().getType() == RecoverySource.Type.SNAPSHOT : "invalid recovery type: "
                 + recoveryState.getRecoverySource();
             StoreRecovery storeRecovery = new StoreRecovery(shardId, logger);
-            storeRecovery.recoverFromSnapshotAndRemoteStore(this, repository, remoteSegmentStoreDirectoryFactory, listener);
+            storeRecovery.recoverFromSnapshotAndRemoteStore(this, repository, repositoriesService, listener, threadPool);
         } catch (Exception e) {
             listener.onFailure(e);
         }
@@ -3565,7 +3564,7 @@ public void startRecovery(
                         "from snapshot and remote store",
                         recoveryState,
                         recoveryListener,
-                        l -> restoreFromSnapshotAndRemoteStore(repositoriesService.repository(repo), remoteSegmentStoreDirectoryFactory, l)
+                        l -> restoreFromSnapshotAndRemoteStore(repositoriesService.repository(repo), repositoriesService, l)
                     );
                     // indicesService.indexService(shardRouting.shardId().getIndex()).addMetadataListener();
                 } else {
@@ -4921,7 +4920,7 @@ private String copySegmentFiles(
 
             if (toDownloadSegments.isEmpty() == false) {
                 try {
-                    downloadSegments(storeDirectory, sourceRemoteDirectory, targetRemoteDirectory, toDownloadSegments, onFileSync);
+                    fileDownloader.download(sourceRemoteDirectory, storeDirectory, targetRemoteDirectory, toDownloadSegments, onFileSync);
                 } catch (Exception e) {
                     throw new IOException("Error occurred when downloading segments from remote store", e);
                 }
@@ -4934,26 +4933,6 @@ private String copySegmentFiles(
         return segmentNFile;
     }
 
-    private void downloadSegments(
-        Directory storeDirectory,
-        RemoteSegmentStoreDirectory sourceRemoteDirectory,
-        RemoteSegmentStoreDirectory targetRemoteDirectory,
-        Set<String> toDownloadSegments,
-        final Runnable onFileSync
-    ) throws IOException {
-        final Path indexPath = store.shardPath() == null ? null : store.shardPath().resolveIndex();
-        final DirectoryFileTransferTracker fileTransferTracker = store.getDirectoryFileTransferTracker();
-        for (String segment : toDownloadSegments) {
-            final PlainActionFuture<String> segmentListener = PlainActionFuture.newFuture();
-            sourceRemoteDirectory.copyTo(segment, storeDirectory, indexPath, fileTransferTracker, segmentListener);
-            segmentListener.actionGet();
-            onFileSync.run();
-            if (targetRemoteDirectory != null) {
-                targetRemoteDirectory.copyFrom(storeDirectory, segment, segment, IOContext.DEFAULT);
-            }
-        }
-    }
-
     private boolean localDirectoryContains(Directory localDirectory, String file, long checksum) {
         try (IndexInput indexInput = localDirectory.openInput(file, IOContext.DEFAULT)) {
             if (checksum == CodecUtil.retrieveChecksum(indexInput)) {

diff --git a/server/src/main/java/org/opensearch/index/shard/StoreRecovery.java b/server/src/main/java/org/opensearch/index/shard/StoreRecovery.java
@@ -70,7 +70,9 @@
 import org.opensearch.indices.recovery.RecoveryState;
 import org.opensearch.indices.replication.common.ReplicationLuceneIndex;
 import org.opensearch.repositories.IndexId;
+import org.opensearch.repositories.RepositoriesService;
 import org.opensearch.repositories.Repository;
+import org.opensearch.threadpool.ThreadPool;
 
 import java.io.IOException;
 import java.nio.channels.FileChannel;
@@ -360,8 +362,9 @@ void recoverFromRepository(final IndexShard indexShard, Repository repository, A
     void recoverFromSnapshotAndRemoteStore(
         final IndexShard indexShard,
         Repository repository,
-        RemoteSegmentStoreDirectoryFactory directoryFactory,
-        ActionListener<Boolean> listener
+        RepositoriesService repositoriesService,
+        ActionListener<Boolean> listener,
+        ThreadPool threadPool
     ) {
         try {
             if (canRecover(indexShard)) {
@@ -389,6 +392,10 @@ void recoverFromSnapshotAndRemoteStore(
                     remoteStoreRepository = shallowCopyShardMetadata.getRemoteStoreRepository();
                 }
 
+                RemoteSegmentStoreDirectoryFactory directoryFactory = new RemoteSegmentStoreDirectoryFactory(
+                    () -> repositoriesService,
+                    threadPool
+                );
                 RemoteSegmentStoreDirectory sourceRemoteDirectory = (RemoteSegmentStoreDirectory) directoryFactory.newDirectory(
                     remoteStoreRepository,
                     indexUUID,

diff --git a/server/src/main/java/org/opensearch/index/store/RemoteDirectory.java b/server/src/main/java/org/opensearch/index/store/RemoteDirectory.java
@@ -333,10 +333,6 @@ public boolean copyFrom(
         return false;
     }
 
-    protected UnaryOperator<InputStream> getDownloadRateLimiter() {
-        return downloadRateLimiter;
-    }
-
     private void uploadBlob(
         Directory from,
         String src,

diff --git a/server/src/main/java/org/opensearch/index/store/RemoteSegmentStoreDirectory.java b/server/src/main/java/org/opensearch/index/store/RemoteSegmentStoreDirectory.java
@@ -24,8 +24,6 @@
 import org.apache.lucene.store.IndexOutput;
 import org.apache.lucene.util.Version;
 import org.opensearch.common.UUIDs;
-import org.opensearch.common.blobstore.AsyncMultiStreamBlobContainer;
-import org.opensearch.common.blobstore.stream.read.listener.ReadContextListener;
 import org.opensearch.common.collect.Tuple;
 import org.opensearch.common.io.VersionedCodecStreamWrapper;
 import org.opensearch.common.logging.Loggers;
@@ -38,15 +36,13 @@
 import org.opensearch.index.store.lockmanager.RemoteStoreLockManager;
 import org.opensearch.index.store.remote.metadata.RemoteSegmentMetadata;
 import org.opensearch.index.store.remote.metadata.RemoteSegmentMetadataHandler;
-import org.opensearch.indices.recovery.RecoverySettings;
 import org.opensearch.indices.replication.checkpoint.ReplicationCheckpoint;
 import org.opensearch.threadpool.ThreadPool;
 
 import java.io.FileNotFoundException;
 import java.io.IOException;
 import java.io.InputStream;
 import java.nio.file.NoSuchFileException;
-import java.nio.file.Path;
 import java.util.ArrayList;
 import java.util.Collection;
 import java.util.Collections;
@@ -93,8 +89,6 @@ public final class RemoteSegmentStoreDirectory extends FilterDirectory implement
 
     private final ThreadPool threadPool;
 
-    private final RecoverySettings recoverySettings;
-
     /**
      * Keeps track of local segment filename to uploaded filename along with other attributes like checksum.
      * This map acts as a cache layer for uploaded segment filenames which helps avoid calling listAll() each time.
@@ -127,15 +121,13 @@ public RemoteSegmentStoreDirectory(
         RemoteDirectory remoteMetadataDirectory,
         RemoteStoreLockManager mdLockManager,
         ThreadPool threadPool,
-        ShardId shardId,
-        RecoverySettings recoverySettings
+        ShardId shardId
     ) throws IOException {
         super(remoteDataDirectory);
         this.remoteDataDirectory = remoteDataDirectory;
         this.remoteMetadataDirectory = remoteMetadataDirectory;
         this.mdLockManager = mdLockManager;
         this.threadPool = threadPool;
-        this.recoverySettings = recoverySettings;
         this.logger = Loggers.getLogger(getClass(), shardId);
         init();
     }
@@ -472,70 +464,6 @@ public void copyFrom(Directory from, String src, IOContext context, ActionListen
         }
     }
 
-    /**
-     * Copies an existing {@code source} file from this directory to a non-existent file (also
-     * named {@code source}) in either {@code destinationDirectory} or {@code destinationPath}.
-     * If the blob container backing this directory supports multipart downloads, the {@code source}
-     * file will be downloaded (potentially in multiple concurrent parts) directly to
-     * {@code destinationPath}. This method will return immediately and {@code fileCompletionListener}
-     * will be notified upon completion.
-     * <p>
-     * If multipart downloads are not supported, then {@code source} file will be copied to a file named
-     * {@code source} in a single part to {@code destinationDirectory}. The download will happen on the
-     * calling thread and {@code fileCompletionListener} will be notified synchronously before this
-     * method returns.
-     *
-     * @param source The source file name
-     * @param destinationDirectory The destination directory (if multipart is not supported)
-     * @param destinationPath The destination path (if multipart is supported)
-     * @param fileTransferTracker Tracker used for file transfer stats
-     * @param fileCompletionListener The listener to notify of completion
-     */
-    public void copyTo(
-        String source,
-        Directory destinationDirectory,
-        Path destinationPath,
-        DirectoryFileTransferTracker fileTransferTracker,
-        ActionListener<String> fileCompletionListener
-    ) {
-        final String blobName = getExistingRemoteFilename(source);
-        if (destinationPath != null && remoteDataDirectory.getBlobContainer() instanceof AsyncMultiStreamBlobContainer) {
-            long length = 0L;
-            try {
-                length = fileLength(source);
-            } catch (IOException ex) {
-                logger.error("Unable to fetch segment length for stats tracking", ex);
-            }
-            final long fileLength = length;
-            final long startTime = System.currentTimeMillis();
-            fileTransferTracker.addTransferredBytesStarted(fileLength);
-            final AsyncMultiStreamBlobContainer blobContainer = (AsyncMultiStreamBlobContainer) remoteDataDirectory.getBlobContainer();
-            final Path destinationFilePath = destinationPath.resolve(source);
-            final ActionListener<String> completionListener = ActionListener.wrap(response -> {
-                fileTransferTracker.addTransferredBytesSucceeded(fileLength, startTime);
-                fileCompletionListener.onResponse(response);
-            }, e -> {
-                fileTransferTracker.addTransferredBytesFailed(fileLength, startTime);
-                fileCompletionListener.onFailure(e);
-            });
-            final ReadContextListener readContextListener = new ReadContextListener(
-                blobName,
-                destinationFilePath,
-                completionListener,
-                threadPool,
-                remoteDataDirectory.getDownloadRateLimiter(),
-                recoverySettings.getMaxConcurrentRemoteStoreStreams()
-            );
-            blobContainer.readBlobAsync(blobName, readContextListener);
-        } else {
-            // Fallback to older mechanism of downloading the file
-            ActionListener.completeWith(fileCompletionListener, () -> {
-                destinationDirectory.copyFrom(this, source, source, IOContext.DEFAULT);
-                return source;
-            });
-        }
-    }
-
     /**
      * This acquires a lock on a given commit by creating a lock file in lock directory using {@code FileLockInfo}
      *