com.swirlds.merkledb.files.hashmap.HalfDiskHashMap Maven / Gradle / Ivy
Show all versions of swirlds-merkledb Show documentation
/*
* Copyright (C) 2021-2024 Hedera Hashgraph, LLC
*
* 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 com.swirlds.merkledb.files.hashmap;
import static com.swirlds.common.threading.manager.AdHocThreadManager.getStaticThreadManager;
import static com.swirlds.logging.legacy.LogMarker.EXCEPTION;
import static com.swirlds.logging.legacy.LogMarker.MERKLE_DB;
import static com.swirlds.merkledb.MerkleDb.MERKLEDB_COMPONENT;
import com.hedera.pbj.runtime.io.buffer.BufferedData;
import com.hedera.pbj.runtime.io.buffer.Bytes;
import com.swirlds.common.config.singleton.ConfigurationHolder;
import com.swirlds.common.threading.framework.config.ThreadConfiguration;
import com.swirlds.merkledb.FileStatisticAware;
import com.swirlds.merkledb.Snapshotable;
import com.swirlds.merkledb.collections.CASableLongIndex;
import com.swirlds.merkledb.collections.LongList;
import com.swirlds.merkledb.collections.LongListDisk;
import com.swirlds.merkledb.collections.LongListOffHeap;
import com.swirlds.merkledb.collections.OffHeapUser;
import com.swirlds.merkledb.config.MerkleDbConfig;
import com.swirlds.merkledb.files.DataFileCollection;
import com.swirlds.merkledb.files.DataFileCollection.LoadedDataCallback;
import com.swirlds.merkledb.files.DataFileReader;
import edu.umd.cs.findbugs.annotations.Nullable;
import java.io.DataInputStream;
import java.io.DataOutputStream;
import java.io.IOException;
import java.nio.file.Files;
import java.nio.file.Path;
import java.util.Iterator;
import java.util.LongSummaryStatistics;
import java.util.Queue;
import java.util.concurrent.ConcurrentLinkedQueue;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import org.apache.logging.log4j.LogManager;
import org.apache.logging.log4j.Logger;
import org.eclipse.collections.api.tuple.primitive.IntObjectPair;
import org.eclipse.collections.impl.map.mutable.primitive.IntObjectHashMap;
/**
* This is a hash map implementation where the bucket index is in RAM and the buckets are on disk.
* It maps a VirtualKey to a long value. This allows very large maps with minimal RAM usage and the
* best performance profile as by using an in memory index we avoid the need for random disk writes.
* Random disk writes are horrible performance wise in our testing.
*
* This implementation depends on good hashCode() implementation on the keys, if there are too
* many hash collisions the performance can get bad.
*
*
IMPORTANT: This implementation assumes a single writing thread. There can be multiple
* readers while writing is happening.
*/
public class HalfDiskHashMap implements AutoCloseable, Snapshotable, FileStatisticAware, OffHeapUser {
private static final Logger logger = LogManager.getLogger(HalfDiskHashMap.class);
/** The version number for format of current data files */
private static final int METADATA_FILE_FORMAT_VERSION = 1;
/** Metadata file name suffix with extension. */
private static final String METADATA_FILENAME_SUFFIX = "_metadata.hdhm";
/** Bucket index file name suffix with extension */
private static final String BUCKET_INDEX_FILENAME_SUFFIX = "_bucket_index.ll";
/**
* A marker to indicate that a value should be deleted from the map, or that there is
* no old value to compare against in putIfEqual/deleteIfEqual
*/
protected static final long INVALID_VALUE = Long.MIN_VALUE;
/**
* This is the average number of entries per bucket we aim for when filled to mapSize. It is a
* heuristic used in calculation for how many buckets to create. The larger this number the
* slower lookups will be but the more even distribution of entries across buckets will be. So
* it is a matter of balance.
*/
private static final long GOOD_AVERAGE_BUCKET_ENTRY_COUNT = 32;
/** The limit on the number of concurrent read tasks in {@code endWriting()} */
private static final int MAX_IN_FLIGHT = 64;
/**
* Long list used for mapping bucketIndex(index into list) to disk location for latest copy of
* bucket
*/
private final LongList bucketIndexToBucketLocation;
/** DataFileCollection manages the files storing the buckets on disk */
private final DataFileCollection fileCollection;
/**
* This is the next power of 2 bigger than minimumBuckets. It needs to be a power of two, so
* that we can optimize and avoid the cost of doing a % to find the bucket index from hash code.
*/
private final int numOfBuckets;
/**
* The requested max size for the map, this is the maximum number of key/values expected to be
* stored in this map.
*/
private final long mapSize;
/** The name to use for the files prefix on disk */
private final String storeName;
/** Bucket pool used by this HDHM */
private final ReusableBucketPool bucketPool;
/** Store for session data during a writing transaction */
private IntObjectHashMap oneTransactionsData = null;
/**
* The thread that called startWriting. We use it to check that other writing calls are done on
* same thread
*/
private Thread writingThread;
/** Executor for parallel bucket reads/updates in {@link #endWriting()} */
private static volatile ExecutorService flushExecutor = null;
private static ExecutorService getFlushExecutor() {
ExecutorService exec = flushExecutor;
if (exec == null) {
synchronized (HalfDiskHashMap.class) {
exec = flushExecutor;
if (exec == null) {
final MerkleDbConfig config = ConfigurationHolder.getConfigData(MerkleDbConfig.class);
exec = Executors.newFixedThreadPool(
config.getNumHalfDiskHashMapFlushThreads(),
new ThreadConfiguration(getStaticThreadManager())
.setComponent(MERKLEDB_COMPONENT)
.setThreadName("HalfDiskHashMap Flushing")
.setExceptionHandler((t, ex) -> logger.error(
EXCEPTION.getMarker(), "Uncaught exception during HDHM flushing", ex))
.buildFactory());
flushExecutor = exec;
}
}
}
return exec;
}
/**
* Construct a new HalfDiskHashMap
*
* @param config MerkleDb config
* @param mapSize The maximum map number of entries. This should be more than big enough to
* avoid too many key collisions.
* @param storeDir The directory to use for storing data files.
* @param storeName The name for the data store, this allows more than one data store in a
* single directory.
* @param legacyStoreName Base name for the data store. If not null, the store will process
* files with this prefix at startup. New files in the store will be prefixed with {@code
* storeName}
* @param preferDiskBasedIndex When true we will use disk based index rather than ram where
* possible. This will come with a significant performance cost, especially for writing. It
* is possible to load a data source that was written with memory index with disk based
* index and vice versa.
* @throws IOException If there was a problem creating or opening a set of data files.
*/
public HalfDiskHashMap(
final MerkleDbConfig config,
final long mapSize,
final Path storeDir,
final String storeName,
final String legacyStoreName,
final boolean preferDiskBasedIndex)
throws IOException {
this.mapSize = mapSize;
this.storeName = storeName;
Path indexFile = storeDir.resolve(storeName + BUCKET_INDEX_FILENAME_SUFFIX);
// create bucket pool
this.bucketPool = new ReusableBucketPool(Bucket::new);
// load or create new
LoadedDataCallback loadedDataCallback;
if (Files.exists(storeDir)) {
// load metadata
Path metaDataFile = storeDir.resolve(storeName + METADATA_FILENAME_SUFFIX);
boolean loadedLegacyMetadata = false;
if (!Files.exists(metaDataFile)) {
metaDataFile = storeDir.resolve(legacyStoreName + METADATA_FILENAME_SUFFIX);
indexFile = storeDir.resolve(legacyStoreName + BUCKET_INDEX_FILENAME_SUFFIX);
loadedLegacyMetadata = true;
}
if (Files.exists(metaDataFile)) {
try (DataInputStream metaIn = new DataInputStream(Files.newInputStream(metaDataFile))) {
final int fileVersion = metaIn.readInt();
if (fileVersion != METADATA_FILE_FORMAT_VERSION) {
throw new IOException("Tried to read a file with incompatible file format version ["
+ fileVersion
+ "], expected ["
+ METADATA_FILE_FORMAT_VERSION
+ "].");
}
metaIn.readInt(); // backwards compatibility, was: minimumBuckets
numOfBuckets = metaIn.readInt();
}
if (loadedLegacyMetadata) {
Files.delete(metaDataFile);
}
} else {
logger.error(
EXCEPTION.getMarker(),
"Loading existing set of data files but no metadata file was found in [{}]",
storeDir.toAbsolutePath());
throw new IOException("Can not load an existing HalfDiskHashMap from ["
+ storeDir.toAbsolutePath()
+ "] because metadata file is missing");
}
// load or rebuild index
final boolean forceIndexRebuilding = config.indexRebuildingEnforced();
if (Files.exists(indexFile) && !forceIndexRebuilding) {
bucketIndexToBucketLocation =
preferDiskBasedIndex ? new LongListDisk(indexFile) : new LongListOffHeap(indexFile);
loadedDataCallback = null;
} else {
// create new index and setup call back to rebuild
bucketIndexToBucketLocation =
preferDiskBasedIndex ? new LongListDisk(indexFile) : new LongListOffHeap();
loadedDataCallback = (dataLocation, bucketData) -> {
final Bucket bucket = bucketPool.getBucket();
bucket.readFrom(bucketData);
bucketIndexToBucketLocation.put(bucket.getBucketIndex(), dataLocation);
};
}
} else {
// create store dir
Files.createDirectories(storeDir);
// create new index
bucketIndexToBucketLocation = preferDiskBasedIndex ? new LongListDisk(indexFile) : new LongListOffHeap();
// calculate number of entries we can store in a disk page
final int minimumBuckets = (int) (mapSize / GOOD_AVERAGE_BUCKET_ENTRY_COUNT);
// numOfBuckets is the nearest power of two greater than minimumBuckets with a min of 2
numOfBuckets = Math.max(Integer.highestOneBit(minimumBuckets) * 2, 2);
// we are new so no need for a loadedDataCallback
loadedDataCallback = null;
// write metadata
writeMetadata(storeDir);
logger.info(
MERKLE_DB.getMarker(),
"HalfDiskHashMap [{}] created with minimumBuckets={} and numOfBuckets={}",
storeName,
minimumBuckets,
numOfBuckets);
}
bucketIndexToBucketLocation.updateValidRange(0, numOfBuckets - 1);
// create file collection
fileCollection = new DataFileCollection(
// Need: propagate MerkleDb config from the database
config, storeDir, storeName, legacyStoreName, loadedDataCallback);
}
private void writeMetadata(final Path dir) throws IOException {
try (DataOutputStream metaOut =
new DataOutputStream(Files.newOutputStream(dir.resolve(storeName + METADATA_FILENAME_SUFFIX)))) {
metaOut.writeInt(METADATA_FILE_FORMAT_VERSION);
metaOut.writeInt(0); // backwards compatibility, was: minimumBuckets
metaOut.writeInt(numOfBuckets);
metaOut.flush();
}
}
/** {@inheritDoc} */
public void snapshot(final Path snapshotDirectory) throws IOException {
// create snapshot directory if needed
Files.createDirectories(snapshotDirectory);
// write index to file
bucketIndexToBucketLocation.writeToFile(snapshotDirectory.resolve(storeName + BUCKET_INDEX_FILENAME_SUFFIX));
// snapshot files
fileCollection.snapshot(snapshotDirectory);
// write metadata
writeMetadata(snapshotDirectory);
}
/**
* {@inheritDoc}
*/
@Override
public long getOffHeapConsumption() {
if (bucketIndexToBucketLocation instanceof LongListOffHeap offheapIndex) {
return offheapIndex.getOffHeapConsumption();
}
return 0;
}
/**
* {@inheritDoc}
*/
public LongSummaryStatistics getFilesSizeStatistics() {
return fileCollection.getAllCompletedFilesSizeStatistics();
}
/**
* Close this HalfDiskHashMap's data files. Once closed this HalfDiskHashMap can not be reused.
* You should make sure you call close before system exit otherwise any files being written
* might not be in a good state.
*
* @throws IOException If there was a problem closing the data files.
*/
@Override
public void close() throws IOException {
// Close the files first, then the index. If done in a different order, there may be
// file operations still running, but the index is already closed
fileCollection.close();
bucketIndexToBucketLocation.close();
}
// =================================================================================================================
// Writing API - Single thead safe
/**
* Start a writing session to the map. Each new writing session results in a new data file on
* disk, so you should ideally batch up map writes.
*/
public void startWriting() {
oneTransactionsData = new IntObjectHashMap<>();
writingThread = Thread.currentThread();
}
private BucketMutation findBucketForUpdate(
final Bytes keyBytes, final int keyHashCode, final long oldValue, final long value) {
if (keyBytes == null) {
throw new IllegalArgumentException("Can not write a null key");
}
if (oneTransactionsData == null) {
throw new IllegalStateException(
"Trying to write to a HalfDiskHashMap when you have not called startWriting().");
}
if (Thread.currentThread() != writingThread) {
throw new IllegalStateException("Tried to write with different thread to startWriting()");
}
// store key and value in transaction cache
final int bucketIndex = computeBucketIndex(keyHashCode);
return oneTransactionsData.getIfAbsentPut(
bucketIndex, () -> new BucketMutation(keyBytes, keyHashCode, oldValue, value));
}
/**
* Put a key/value during the current writing session. The value will not be retrievable until
* it is committed in the {@link #endWriting()} call.
*
* This method may be called multiple times for the same key in a single writing
* session. The value from the last call will be stored in this map after the session is
* ended.
*
* @param keyBytes the key to store the value for
* @param keyHashCode the key hash code
* @param value the value to store for given key
*/
public void put(final Bytes keyBytes, final int keyHashCode, final long value) {
final BucketMutation bucketMap = findBucketForUpdate(keyBytes, keyHashCode, INVALID_VALUE, value);
bucketMap.put(keyBytes, keyHashCode, value);
}
/**
* Put a key/value during the current writing session. This method is similar to {@link
* #put(Bytes, int, long)}, but the new value is set only if the current value is equal to
* the given {@code oldValue}.
*
*
This method may be called multiple times for the same key in a single writing
* session. If the new value from the first call is equal to the old value in the second
* call, the new value from the second call will be stored in this map after the session
* is ended, otherwise the value from the second call will be ignored.
*
*
If the value for {@code oldValue} is {@link #INVALID_VALUE}, it's ignored, and this
* method is identical to {@link #put(Bytes, int, long)}.
*
* @param keyBytes the key to store the value for
* @param keyHashCode the key hash code
* @param oldValue the value to check the current value against, or {@link #INVALID_VALUE}
* if no current value check is needed
* @param value the value to store for the given key
*/
public void putIfEqual(final Bytes keyBytes, final int keyHashCode, final long oldValue, final long value) {
final BucketMutation bucketMap = findBucketForUpdate(keyBytes, keyHashCode, oldValue, value);
bucketMap.putIfEqual(keyBytes, keyHashCode, oldValue, value);
}
/**
* Delete a key entry from the map.
*
* @param keyBytes The key to delete entry for
*/
public void delete(final Bytes keyBytes, final int keyHashCode) {
put(keyBytes, keyHashCode, INVALID_VALUE);
}
/**
* Delete a key entry from the map, if the current value is equal to the given {@code oldValue}.
* If {@code oldValue} is {@link #INVALID_VALUE}, no current value check is performed, and this
* method is identical to {@link #delete(Bytes, int)}.
*
* @param keyBytes the key to delete the entry for
* @param oldValue the value to check the current value against, or {@link #INVALID_VALUE}
* if no current value check is needed
*/
public void deleteIfEqual(final Bytes keyBytes, final int keyHashCode, final long oldValue) {
putIfEqual(keyBytes, keyHashCode, oldValue, INVALID_VALUE);
}
/**
* End current writing session, committing all puts to data store.
*
* @return Data file reader for the file written
* @throws IOException If there was a problem committing data to store
*/
@Nullable
public DataFileReader endWriting() throws IOException {
/* FUTURE WORK - https://github.com/swirlds/swirlds-platform/issues/3943 */
if (Thread.currentThread() != writingThread) {
throw new IllegalStateException("Tried calling endWriting with different thread to startWriting()");
}
writingThread = null;
final int size = oneTransactionsData.size();
logger.info(
MERKLE_DB.getMarker(),
"Finishing writing to {}, num of changed bins = {}, num of changed keys = {}",
storeName,
size,
oneTransactionsData.stream().mapToLong(BucketMutation::size).sum());
final ExecutorService flushExecutor = getFlushExecutor();
final DataFileReader dataFileReader;
if (size > 0) {
final Queue queue = new ConcurrentLinkedQueue<>();
final Iterator> iterator =
oneTransactionsData.keyValuesView().iterator();
// read and update all buckets in parallel, write sequentially in random order
fileCollection.startWriting();
int processed = 0;
int inFlight = 0;
while (processed < size) {
// submit read tasks
while (inFlight < MAX_IN_FLIGHT && iterator.hasNext()) {
IntObjectPair keyValue = iterator.next();
final int bucketIndex = keyValue.getOne();
final BucketMutation bucketMap = keyValue.getTwo();
flushExecutor.execute(() -> readUpdateQueueBucket(bucketIndex, bucketMap, queue));
++inFlight;
}
ReadBucketResult res;
while ((res = queue.poll()) != null) {
--inFlight;
if (res.error != null) {
throw new RuntimeException(res.error);
}
try (final Bucket bucket = res.bucket) {
final int bucketIndex = bucket.getBucketIndex();
if (bucket.isEmpty()) {
// bucket is missing or empty, remove it from the index
bucketIndexToBucketLocation.remove(bucketIndex);
} else {
// save bucket
final long bucketLocation =
fileCollection.storeDataItem(bucket::writeTo, bucket.sizeInBytes());
// update bucketIndexToBucketLocation
bucketIndexToBucketLocation.put(bucketIndex, bucketLocation);
}
} finally {
++processed;
}
}
}
// close files session
dataFileReader = fileCollection.endWriting(0, numOfBuckets);
// we have updated all indexes so the data file can now be included in merges
dataFileReader.setFileCompleted();
} else {
dataFileReader = null;
}
// clear put cache
oneTransactionsData = null;
return dataFileReader;
}
/**
* Reads a bucket with a given index from disk, updates given keys in it, and puts the bucket to
* a queue. If an exception is thrown, it's put to the queue instead, so the number of {@code
* ReadBucketResult} objects in the queue is consistent.
*
* @param bucketIndex The bucket index
* @param keyUpdates Key/value updates to apply to the bucket
* @param queue The queue to put the bucket or exception to
*/
private void readUpdateQueueBucket(
final int bucketIndex, final BucketMutation keyUpdates, final Queue queue) {
try {
// The bucket will be closed on the lifecycle thread
final Bucket bucket;
BufferedData bucketData = fileCollection.readDataItemUsingIndex(bucketIndexToBucketLocation, bucketIndex);
if (bucketData == null) {
// create a new bucket
bucket = bucketPool.getBucket();
bucket.setBucketIndex(bucketIndex);
} else {
bucket = bucketPool.getBucket();
bucket.readFrom(bucketData);
}
// for each changed key in bucket, update bucket
keyUpdates.forEachKeyValue(bucket::putValue);
queue.offer(new ReadBucketResult(bucket, null));
} catch (final Exception e) {
logger.error(EXCEPTION.getMarker(), "Failed to read / update bucket", e);
queue.offer(new ReadBucketResult(null, e));
}
}
// =================================================================================================================
// Reading API - Multi thead safe
/**
* Get a value from this map
*
* @param keyBytes the key to get value for
* @param keyHashCode the key hash code
* @param notFoundValue the value to return if the key was not found
* @return the value retrieved from the map or {notFoundValue} if no value was stored for the
* given key
* @throws IOException If there was a problem reading from the map
*/
public long get(final Bytes keyBytes, final int keyHashCode, final long notFoundValue) throws IOException {
if (keyBytes == null) {
throw new IllegalArgumentException("Can not get a null key");
}
final int bucketIndex = computeBucketIndex(keyHashCode);
try (final Bucket bucket = readBucket(bucketIndex)) {
if (bucket != null) {
return bucket.findValue(keyHashCode, keyBytes, notFoundValue);
}
}
return notFoundValue;
}
private Bucket readBucket(final int bucketIndex) throws IOException {
final BufferedData bucketData = fileCollection.readDataItemUsingIndex(bucketIndexToBucketLocation, bucketIndex);
if (bucketData == null) {
return null;
}
final Bucket bucket = bucketPool.getBucket();
bucket.readFrom(bucketData);
return bucket;
}
// =================================================================================================================
// Debugging Print API
/** Debug dump stats for this map */
public void printStats() {
logger.info(
MERKLE_DB.getMarker(),
"""
HalfDiskHashMap Stats {
mapSize = {}
numOfBuckets = {}
GOOD_AVERAGE_BUCKET_ENTRY_COUNT = {}
}""",
mapSize,
numOfBuckets,
GOOD_AVERAGE_BUCKET_ENTRY_COUNT);
}
public DataFileCollection getFileCollection() {
return fileCollection;
}
public CASableLongIndex getBucketIndexToBucketLocation() {
return bucketIndexToBucketLocation;
}
// =================================================================================================================
// Private API
/**
* Computes which bucket a key with the given hash falls. Depends on the fact the numOfBuckets
* is a power of two. Based on same calculation that is used in java HashMap.
*
* @param keyHash the int hash for key
* @return the index of the bucket that key falls in
*/
private int computeBucketIndex(final int keyHash) {
return (numOfBuckets - 1) & keyHash;
}
private record ReadBucketResult(Bucket bucket, Throwable error) {
public ReadBucketResult {
assert (bucket != null) ^ (error != null);
}
}
}