/**
*
* 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.hadoop.hbase.regionserver;
import java.io.IOException;
import java.util.ArrayList;
import java.util.Arrays;
import java.util.Collection;
import java.util.Collections;
import java.util.Comparator;
import java.util.HashMap;
import java.util.Iterator;
import java.util.List;
import java.util.Map;
import java.util.TreeMap;
import org.apache.commons.logging.Log;
import org.apache.commons.logging.LogFactory;
import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.hbase.Cell;
import org.apache.hadoop.hbase.CellComparator;
import org.apache.hadoop.hbase.CellUtil;
import org.apache.hadoop.hbase.HConstants;
import org.apache.hadoop.hbase.KeyValue;
import org.apache.hadoop.hbase.KeyValue.KeyOnlyKeyValue;
import org.apache.hadoop.hbase.classification.InterfaceAudience;
import org.apache.hadoop.hbase.regionserver.compactions.StripeCompactionPolicy;
import org.apache.hadoop.hbase.util.Bytes;
import org.apache.hadoop.hbase.util.ConcatenatedLists;
import org.apache.hadoop.util.StringUtils.TraditionalBinaryPrefix;
import com.google.common.collect.ImmutableCollection;
import com.google.common.collect.ImmutableList;
/**
* Stripe implementation of StoreFileManager.
* Not thread safe - relies on external locking (in HStore). Collections that this class
* returns are immutable or unique to the call, so they should be safe.
* Stripe store splits the key space of the region into non-overlapping stripes, as well as
* some recent files that have all the keys (level 0). Each stripe contains a set of files.
* When L0 is compacted, it's split into the files corresponding to existing stripe boundaries,
* that can thus be added to stripes.
* When scan or get happens, it only has to read the files from the corresponding stripes.
* See StripeCompationPolicy on how the stripes are determined; this class doesn't care.
*
* This class should work together with StripeCompactionPolicy and StripeCompactor.
* With regard to how they work, we make at least the following (reasonable) assumptions:
* - Compaction produces one file per new stripe (if any); that is easy to change.
* - Compaction has one contiguous set of stripes both in and out, except if L0 is involved.
*/
@InterfaceAudience.Private
public class StripeStoreFileManager
implements StoreFileManager, StripeCompactionPolicy.StripeInformationProvider {
private static final Log LOG = LogFactory.getLog(StripeStoreFileManager.class);
/**
* The file metadata fields that contain the stripe information.
*/
public static final byte[] STRIPE_START_KEY = Bytes.toBytes("STRIPE_START_KEY");
public static final byte[] STRIPE_END_KEY = Bytes.toBytes("STRIPE_END_KEY");
private final static Bytes.RowEndKeyComparator MAP_COMPARATOR = new Bytes.RowEndKeyComparator();
/**
* The key value used for range boundary, indicating that the boundary is open (i.e. +-inf).
*/
public final static byte[] OPEN_KEY = HConstants.EMPTY_BYTE_ARRAY;
final static byte[] INVALID_KEY = null;
/**
* The state class. Used solely to replace results atomically during
* compactions and avoid complicated error handling.
*/
private static class State {
/**
* The end rows of each stripe. The last stripe end is always open-ended, so it's not stored
* here. It is invariant that the start row of the stripe is the end row of the previous one
* (and is an open boundary for the first one).
*/
public byte[][] stripeEndRows = new byte[0][];
/**
* Files by stripe. Each element of the list corresponds to stripeEndRow element with the
* same index, except the last one. Inside each list, the files are in reverse order by
* seqNum. Note that the length of this is one higher than that of stripeEndKeys.
*/
public ArrayList<ImmutableList<StoreFile>> stripeFiles = new ArrayList<>();
/** Level 0. The files are in reverse order by seqNum. */
public ImmutableList<StoreFile> level0Files = ImmutableList.<StoreFile>of();
/** Cached list of all files in the structure, to return from some calls */
public ImmutableList<StoreFile> allFilesCached = ImmutableList.<StoreFile>of();
private ImmutableList<StoreFile> allCompactedFilesCached = ImmutableList.<StoreFile>of();
}
private State state = null;
/** Cached file metadata (or overrides as the case may be) */
private HashMap<StoreFile, byte[]> fileStarts = new HashMap<>();
private HashMap<StoreFile, byte[]> fileEnds = new HashMap<>();
/** Normally invalid key is null, but in the map null is the result for "no key"; so use
* the following constant value in these maps instead. Note that this is a constant and
* we use it to compare by reference when we read from the map. */
private static final byte[] INVALID_KEY_IN_MAP = new byte[0];
private final CellComparator cellComparator;
private StripeStoreConfig config;
private final int blockingFileCount;
public StripeStoreFileManager(
CellComparator kvComparator, Configuration conf, StripeStoreConfig config) {
this.cellComparator = kvComparator;
this.config = config;
this.blockingFileCount = conf.getInt(
HStore.BLOCKING_STOREFILES_KEY, HStore.DEFAULT_BLOCKING_STOREFILE_COUNT);
}
@Override
public void loadFiles(List<StoreFile> storeFiles) {
loadUnclassifiedStoreFiles(storeFiles);
}
@Override
public Collection<StoreFile> getStorefiles() {
return state.allFilesCached;
}
@Override
public Collection<StoreFile> getCompactedfiles() {
return state.allCompactedFilesCached;
}
@Override
public void insertNewFiles(Collection<StoreFile> sfs) throws IOException {
CompactionOrFlushMergeCopy cmc = new CompactionOrFlushMergeCopy(true);
// Passing null does not cause NPE??
cmc.mergeResults(null, sfs);
debugDumpState("Added new files");
}
@Override
public ImmutableCollection<StoreFile> clearFiles() {
ImmutableCollection<StoreFile> result = state.allFilesCached;
this.state = new State();
this.fileStarts.clear();
this.fileEnds.clear();
return result;
}
@Override
public ImmutableCollection<StoreFile> clearCompactedFiles() {
ImmutableCollection<StoreFile> result = state.allCompactedFilesCached;
this.state = new State();
return result;
}
@Override
public int getStorefileCount() {
return state.allFilesCached.size();
}
/** See {@link StoreFileManager#getCandidateFilesForRowKeyBefore(KeyValue)}
* for details on this methods. */
@Override
public Iterator<StoreFile> getCandidateFilesForRowKeyBefore(final KeyValue targetKey) {
KeyBeforeConcatenatedLists result = new KeyBeforeConcatenatedLists();
// Order matters for this call.
result.addSublist(state.level0Files);
if (!state.stripeFiles.isEmpty()) {
int lastStripeIndex = findStripeForRow(CellUtil.cloneRow(targetKey), false);
for (int stripeIndex = lastStripeIndex; stripeIndex >= 0; --stripeIndex) {
result.addSublist(state.stripeFiles.get(stripeIndex));
}
}
return result.iterator();
}
/** See {@link StoreFileManager#getCandidateFilesForRowKeyBefore(KeyValue)} and
* {@link StoreFileManager#updateCandidateFilesForRowKeyBefore(Iterator, KeyValue, Cell)}
* for details on this methods. */
@Override
public Iterator<StoreFile> updateCandidateFilesForRowKeyBefore(
Iterator<StoreFile> candidateFiles, final KeyValue targetKey, final Cell candidate) {
KeyBeforeConcatenatedLists.Iterator original =
(KeyBeforeConcatenatedLists.Iterator)candidateFiles;
assert original != null;
ArrayList<List<StoreFile>> components = original.getComponents();
for (int firstIrrelevant = 0; firstIrrelevant < components.size(); ++firstIrrelevant) {
StoreFile sf = components.get(firstIrrelevant).get(0);
byte[] endKey = endOf(sf);
// Entries are ordered as such: L0, then stripes in reverse order. We never remove
// level 0; we remove the stripe, and all subsequent ones, as soon as we find the
// first one that cannot possibly have better candidates.
if (!isInvalid(endKey) && !isOpen(endKey)
&& (nonOpenRowCompare(targetKey, endKey) >= 0)) {
original.removeComponents(firstIrrelevant);
break;
}
}
return original;
}
@Override
/**
* Override of getSplitPoint that determines the split point as the boundary between two
* stripes, unless it causes significant imbalance between split sides' sizes. In that
* case, the split boundary will be chosen from the middle of one of the stripes to
* minimize imbalance.
* @return The split point, or null if no split is possible.
*/
public byte[] getSplitPoint() throws IOException {
if (this.getStorefileCount() == 0) return null;
if (state.stripeFiles.size() <= 1) {
return getSplitPointFromAllFiles();
}
int leftIndex = -1, rightIndex = state.stripeFiles.size();
long leftSize = 0, rightSize = 0;
long lastLeftSize = 0, lastRightSize = 0;
while (rightIndex - 1 != leftIndex) {
if (leftSize >= rightSize) {
--rightIndex;
lastRightSize = getStripeFilesSize(rightIndex);
rightSize += lastRightSize;
} else {
++leftIndex;
lastLeftSize = getStripeFilesSize(leftIndex);
leftSize += lastLeftSize;
}
}
if (leftSize == 0 || rightSize == 0) {
String errMsg = String.format("Cannot split on a boundary - left index %d size %d, "
+ "right index %d size %d", leftIndex, leftSize, rightIndex, rightSize);
debugDumpState(errMsg);
LOG.warn(errMsg);
return getSplitPointFromAllFiles();
}
double ratio = (double)rightSize / leftSize;
if (ratio < 1) {
ratio = 1 / ratio;
}
if (config.getMaxSplitImbalance() > ratio) return state.stripeEndRows[leftIndex];
// If the difference between the sides is too large, we could get the proportional key on
// the a stripe to equalize the difference, but there's no proportional key method at the
// moment, and it's not extremely important.
// See if we can achieve better ratio if we split the bigger side in half.
boolean isRightLarger = rightSize >= leftSize;
double newRatio = isRightLarger
? getMidStripeSplitRatio(leftSize, rightSize, lastRightSize)
: getMidStripeSplitRatio(rightSize, leftSize, lastLeftSize);
if (newRatio < 1) {
newRatio = 1 / newRatio;
}
if (newRatio >= ratio) return state.stripeEndRows[leftIndex];
LOG.debug("Splitting the stripe - ratio w/o split " + ratio + ", ratio with split "
+ newRatio + " configured ratio " + config.getMaxSplitImbalance());
// Ok, we may get better ratio, get it.
return StoreUtils.getLargestFile(state.stripeFiles.get(
isRightLarger ? rightIndex : leftIndex)).getFileSplitPoint(this.cellComparator);
}
private byte[] getSplitPointFromAllFiles() throws IOException {
ConcatenatedLists<StoreFile> sfs = new ConcatenatedLists<>();
sfs.addSublist(state.level0Files);
sfs.addAllSublists(state.stripeFiles);
if (sfs.isEmpty()) return null;
return StoreUtils.getLargestFile(sfs).getFileSplitPoint(this.cellComparator);
}
private double getMidStripeSplitRatio(long smallerSize, long largerSize, long lastLargerSize) {
return (double)(largerSize - lastLargerSize / 2f) / (smallerSize + lastLargerSize / 2f);
}
@Override
public Collection<StoreFile> getFilesForScan(byte[] startRow, boolean includeStartRow,
byte[] stopRow, boolean includeStopRow) {
if (state.stripeFiles.isEmpty()) {
return state.level0Files; // There's just L0.
}
int firstStripe = findStripeForRow(startRow, true);
int lastStripe = findStripeForRow(stopRow, false);
assert firstStripe <= lastStripe;
if (firstStripe == lastStripe && state.level0Files.isEmpty()) {
return state.stripeFiles.get(firstStripe); // There's just one stripe we need.
}
if (firstStripe == 0 && lastStripe == (state.stripeFiles.size() - 1)) {
return state.allFilesCached; // We need to read all files.
}
ConcatenatedLists<StoreFile> result = new ConcatenatedLists<>();
result.addAllSublists(state.stripeFiles.subList(firstStripe, lastStripe + 1));
result.addSublist(state.level0Files);
return result;
}
@Override
public void addCompactionResults(
Collection<StoreFile> compactedFiles, Collection<StoreFile> results) throws IOException {
// See class comment for the assumptions we make here.
LOG.debug("Attempting to merge compaction results: " + compactedFiles.size()
+ " files replaced by " + results.size());
// In order to be able to fail in the middle of the operation, we'll operate on lazy
// copies and apply the result at the end.
CompactionOrFlushMergeCopy cmc = new CompactionOrFlushMergeCopy(false);
cmc.mergeResults(compactedFiles, results);
markCompactedAway(compactedFiles);
debugDumpState("Merged compaction results");
}
// Mark the files as compactedAway once the storefiles and compactedfiles list is finalised
// Let a background thread close the actual reader on these compacted files and also
// ensure to evict the blocks from block cache so that they are no longer in
// cache
private void markCompactedAway(Collection<StoreFile> compactedFiles) {
for (StoreFile file : compactedFiles) {
file.markCompactedAway();
}
}
@Override
public void removeCompactedFiles(Collection<StoreFile> compactedFiles) throws IOException {
// See class comment for the assumptions we make here.
LOG.debug("Attempting to delete compaction results: " + compactedFiles.size());
// In order to be able to fail in the middle of the operation, we'll operate on lazy
// copies and apply the result at the end.
CompactionOrFlushMergeCopy cmc = new CompactionOrFlushMergeCopy(false);
cmc.deleteResults(compactedFiles);
debugDumpState("Deleted compaction results");
}
@Override
public int getStoreCompactionPriority() {
// If there's only L0, do what the default store does.
// If we are in critical priority, do the same - we don't want to trump all stores all
// the time due to how many files we have.
int fc = getStorefileCount();
if (state.stripeFiles.isEmpty() || (this.blockingFileCount <= fc)) {
return this.blockingFileCount - fc;
}
// If we are in good shape, we don't want to be trumped by all other stores due to how
// many files we have, so do an approximate mapping to normal priority range; L0 counts
// for all stripes.
int l0 = state.level0Files.size(), sc = state.stripeFiles.size();
int priority = (int)Math.ceil(((double)(this.blockingFileCount - fc + l0) / sc) - l0);
return (priority <= HStore.PRIORITY_USER) ? (HStore.PRIORITY_USER + 1) : priority;
}
/**
* Gets the total size of all files in the stripe.
* @param stripeIndex Stripe index.
* @return Size.
*/
private long getStripeFilesSize(int stripeIndex) {
long result = 0;
for (StoreFile sf : state.stripeFiles.get(stripeIndex)) {
result += sf.getReader().length();
}
return result;
}
/**
* Loads initial store files that were picked up from some physical location pertaining to
* this store (presumably). Unlike adding files after compaction, assumes empty initial
* sets, and is forgiving with regard to stripe constraints - at worst, many/all files will
* go to level 0.
* @param storeFiles Store files to add.
*/
private void loadUnclassifiedStoreFiles(List<StoreFile> storeFiles) {
LOG.debug("Attempting to load " + storeFiles.size() + " store files.");
TreeMap<byte[], ArrayList<StoreFile>> candidateStripes = new TreeMap<>(MAP_COMPARATOR);
ArrayList<StoreFile> level0Files = new ArrayList<>();
// Separate the files into tentative stripes; then validate. Currently, we rely on metadata.
// If needed, we could dynamically determine the stripes in future.
for (StoreFile sf : storeFiles) {
byte[] startRow = startOf(sf), endRow = endOf(sf);
// Validate the range and put the files into place.
if (isInvalid(startRow) || isInvalid(endRow)) {
insertFileIntoStripe(level0Files, sf); // No metadata - goes to L0.
ensureLevel0Metadata(sf);
} else if (!isOpen(startRow) && !isOpen(endRow) &&
nonOpenRowCompare(startRow, endRow) >= 0) {
LOG.error("Unexpected metadata - start row [" + Bytes.toString(startRow) + "], end row ["
+ Bytes.toString(endRow) + "] in file [" + sf.getPath() + "], pushing to L0");
insertFileIntoStripe(level0Files, sf); // Bad metadata - goes to L0 also.
ensureLevel0Metadata(sf);
} else {
ArrayList<StoreFile> stripe = candidateStripes.get(endRow);
if (stripe == null) {
stripe = new ArrayList<>();
candidateStripes.put(endRow, stripe);
}
insertFileIntoStripe(stripe, sf);
}
}
// Possible improvement - for variable-count stripes, if all the files are in L0, we can
// instead create single, open-ended stripe with all files.
boolean hasOverlaps = false;
byte[] expectedStartRow = null; // first stripe can start wherever
Iterator<Map.Entry<byte[], ArrayList<StoreFile>>> entryIter =
candidateStripes.entrySet().iterator();
while (entryIter.hasNext()) {
Map.Entry<byte[], ArrayList<StoreFile>> entry = entryIter.next();
ArrayList<StoreFile> files = entry.getValue();
// Validate the file start rows, and remove the bad ones to level 0.
for (int i = 0; i < files.size(); ++i) {
StoreFile sf = files.get(i);
byte[] startRow = startOf(sf);
if (expectedStartRow == null) {
expectedStartRow = startRow; // ensure that first stripe is still consistent
} else if (!rowEquals(expectedStartRow, startRow)) {
hasOverlaps = true;
LOG.warn("Store file doesn't fit into the tentative stripes - expected to start at ["
+ Bytes.toString(expectedStartRow) + "], but starts at [" + Bytes.toString(startRow)
+ "], to L0 it goes");
StoreFile badSf = files.remove(i);
insertFileIntoStripe(level0Files, badSf);
ensureLevel0Metadata(badSf);
--i;
}
}
// Check if any files from the candidate stripe are valid. If so, add a stripe.
byte[] endRow = entry.getKey();
if (!files.isEmpty()) {
expectedStartRow = endRow; // Next stripe must start exactly at that key.
} else {
entryIter.remove();
}
}
// In the end, there must be open ends on two sides. If not, and there were no errors i.e.
// files are consistent, they might be coming from a split. We will treat the boundaries
// as open keys anyway, and log the message.
// If there were errors, we'll play it safe and dump everything into L0.
if (!candidateStripes.isEmpty()) {
StoreFile firstFile = candidateStripes.firstEntry().getValue().get(0);
boolean isOpen = isOpen(startOf(firstFile)) && isOpen(candidateStripes.lastKey());
if (!isOpen) {
LOG.warn("The range of the loaded files does not cover full key space: from ["
+ Bytes.toString(startOf(firstFile)) + "], to ["
+ Bytes.toString(candidateStripes.lastKey()) + "]");
if (!hasOverlaps) {
ensureEdgeStripeMetadata(candidateStripes.firstEntry().getValue(), true);
ensureEdgeStripeMetadata(candidateStripes.lastEntry().getValue(), false);
} else {
LOG.warn("Inconsistent files, everything goes to L0.");
for (ArrayList<StoreFile> files : candidateStripes.values()) {
for (StoreFile sf : files) {
insertFileIntoStripe(level0Files, sf);
ensureLevel0Metadata(sf);
}
}
candidateStripes.clear();
}
}
}
// Copy the results into the fields.
State state = new State();
state.level0Files = ImmutableList.copyOf(level0Files);
state.stripeFiles = new ArrayList<>(candidateStripes.size());
state.stripeEndRows = new byte[Math.max(0, candidateStripes.size() - 1)][];
ArrayList<StoreFile> newAllFiles = new ArrayList<>(level0Files);
int i = candidateStripes.size() - 1;
for (Map.Entry<byte[], ArrayList<StoreFile>> entry : candidateStripes.entrySet()) {
state.stripeFiles.add(ImmutableList.copyOf(entry.getValue()));
newAllFiles.addAll(entry.getValue());
if (i > 0) {
state.stripeEndRows[state.stripeFiles.size() - 1] = entry.getKey();
}
--i;
}
state.allFilesCached = ImmutableList.copyOf(newAllFiles);
this.state = state;
debugDumpState("Files loaded");
}
private void ensureEdgeStripeMetadata(ArrayList<StoreFile> stripe, boolean isFirst) {
HashMap<StoreFile, byte[]> targetMap = isFirst ? fileStarts : fileEnds;
for (StoreFile sf : stripe) {
targetMap.put(sf, OPEN_KEY);
}
}
private void ensureLevel0Metadata(StoreFile sf) {
if (!isInvalid(startOf(sf))) this.fileStarts.put(sf, INVALID_KEY_IN_MAP);
if (!isInvalid(endOf(sf))) this.fileEnds.put(sf, INVALID_KEY_IN_MAP);
}
private void debugDumpState(String string) {
if (!LOG.isDebugEnabled()) return;
StringBuilder sb = new StringBuilder();
sb.append("\n" + string + "; current stripe state is as such:");
sb.append("\n level 0 with ")
.append(state.level0Files.size())
.append(
" files: "
+ TraditionalBinaryPrefix.long2String(
StripeCompactionPolicy.getTotalFileSize(state.level0Files), "", 1) + ";");
for (int i = 0; i < state.stripeFiles.size(); ++i) {
String endRow = (i == state.stripeEndRows.length)
? "(end)" : "[" + Bytes.toString(state.stripeEndRows[i]) + "]";
sb.append("\n stripe ending in ")
.append(endRow)
.append(" with ")
.append(state.stripeFiles.get(i).size())
.append(
" files: "
+ TraditionalBinaryPrefix.long2String(
StripeCompactionPolicy.getTotalFileSize(state.stripeFiles.get(i)), "", 1) + ";");
}
sb.append("\n").append(state.stripeFiles.size()).append(" stripes total.");
sb.append("\n").append(getStorefileCount()).append(" files total.");
LOG.debug(sb.toString());
}
/**
* Checks whether the key indicates an open interval boundary (i.e. infinity).
*/
private static final boolean isOpen(byte[] key) {
return key != null && key.length == 0;
}
private static final boolean isOpen(Cell key) {
return key != null && key.getRowLength() == 0;
}
/**
* Checks whether the key is invalid (e.g. from an L0 file, or non-stripe-compacted files).
*/
private static final boolean isInvalid(byte[] key) {
// No need to use Arrays.equals because INVALID_KEY is null
return key == INVALID_KEY;
}
/**
* Compare two keys for equality.
*/
private final boolean rowEquals(byte[] k1, byte[] k2) {
return Bytes.equals(k1, 0, k1.length, k2, 0, k2.length);
}
/**
* Compare two keys. Keys must not be open (isOpen(row) == false).
*/
private final int nonOpenRowCompare(byte[] k1, byte[] k2) {
assert !isOpen(k1) && !isOpen(k2);
return Bytes.compareTo(k1, k2);
}
private final int nonOpenRowCompare(Cell k1, byte[] k2) {
assert !isOpen(k1) && !isOpen(k2);
return cellComparator.compareRows(k1, k2, 0, k2.length);
}
/**
* Finds the stripe index by end row.
*/
private final int findStripeIndexByEndRow(byte[] endRow) {
assert !isInvalid(endRow);
if (isOpen(endRow)) return state.stripeEndRows.length;
return Arrays.binarySearch(state.stripeEndRows, endRow, Bytes.BYTES_COMPARATOR);
}
/**
* Finds the stripe index for the stripe containing a row provided externally for get/scan.
*/
private final int findStripeForRow(byte[] row, boolean isStart) {
if (isStart && Arrays.equals(row, HConstants.EMPTY_START_ROW)) return 0;
if (!isStart && Arrays.equals(row, HConstants.EMPTY_END_ROW)) {
return state.stripeFiles.size() - 1;
}
// If there's an exact match below, a stripe ends at "row". Stripe right boundary is
// exclusive, so that means the row is in the next stripe; thus, we need to add one to index.
// If there's no match, the return value of binarySearch is (-(insertion point) - 1), where
// insertion point is the index of the next greater element, or list size if none. The
// insertion point happens to be exactly what we need, so we need to add one to the result.
return Math.abs(Arrays.binarySearch(state.stripeEndRows, row, Bytes.BYTES_COMPARATOR) + 1);
}
@Override
public final byte[] getStartRow(int stripeIndex) {
return (stripeIndex == 0 ? OPEN_KEY : state.stripeEndRows[stripeIndex - 1]);
}
@Override
public final byte[] getEndRow(int stripeIndex) {
return (stripeIndex == state.stripeEndRows.length
? OPEN_KEY : state.stripeEndRows[stripeIndex]);
}
private byte[] startOf(StoreFile sf) {
byte[] result = fileStarts.get(sf);
// result and INVALID_KEY_IN_MAP are compared _only_ by reference on purpose here as the latter
// serves only as a marker and is not to be confused with other empty byte arrays.
// See Javadoc of INVALID_KEY_IN_MAP for more information
return (result == null)
? sf.getMetadataValue(STRIPE_START_KEY)
: result == INVALID_KEY_IN_MAP ? INVALID_KEY : result;
}
private byte[] endOf(StoreFile sf) {
byte[] result = fileEnds.get(sf);
// result and INVALID_KEY_IN_MAP are compared _only_ by reference on purpose here as the latter
// serves only as a marker and is not to be confused with other empty byte arrays.
// See Javadoc of INVALID_KEY_IN_MAP for more information
return (result == null)
? sf.getMetadataValue(STRIPE_END_KEY)
: result == INVALID_KEY_IN_MAP ? INVALID_KEY : result;
}
/**
* Inserts a file in the correct place (by seqnum) in a stripe copy.
* @param stripe Stripe copy to insert into.
* @param sf File to insert.
*/
private static void insertFileIntoStripe(ArrayList<StoreFile> stripe, StoreFile sf) {
// The only operation for which sorting of the files matters is KeyBefore. Therefore,
// we will store the file in reverse order by seqNum from the outset.
for (int insertBefore = 0; ; ++insertBefore) {
if (insertBefore == stripe.size()
|| (StoreFile.Comparators.SEQ_ID.compare(sf, stripe.get(insertBefore)) >= 0)) {
stripe.add(insertBefore, sf);
break;
}
}
}
/**
* An extension of ConcatenatedLists that has several peculiar properties.
* First, one can cut the tail of the logical list by removing last several sub-lists.
* Second, items can be removed thru iterator.
* Third, if the sub-lists are immutable, they are replaced with mutable copies when needed.
* On average KeyBefore operation will contain half the stripes as potential candidates,
* but will quickly cut down on them as it finds something in the more likely ones; thus,
* the above allow us to avoid unnecessary copying of a bunch of lists.
*/
private static class KeyBeforeConcatenatedLists extends ConcatenatedLists<StoreFile> {
@Override
public java.util.Iterator<StoreFile> iterator() {
return new Iterator();
}
public class Iterator extends ConcatenatedLists<StoreFile>.Iterator {
public ArrayList<List<StoreFile>> getComponents() {
return components;
}
public void removeComponents(int startIndex) {
List<List<StoreFile>> subList = components.subList(startIndex, components.size());
for (List<StoreFile> entry : subList) {
size -= entry.size();
}
assert size >= 0;
subList.clear();
}
@Override
public void remove() {
if (!this.nextWasCalled) {
throw new IllegalStateException("No element to remove");
}
this.nextWasCalled = false;
List<StoreFile> src = components.get(currentComponent);
if (src instanceof ImmutableList<?>) {
src = new ArrayList<>(src);
components.set(currentComponent, src);
}
src.remove(indexWithinComponent);
--size;
--indexWithinComponent;
if (src.isEmpty()) {
components.remove(currentComponent); // indexWithinComponent is already -1 here.
}
}
}
}
/**
* Non-static helper class for merging compaction or flush results.
* Since we want to merge them atomically (more or less), it operates on lazy copies,
* then creates a new state object and puts it in place.
*/
private class CompactionOrFlushMergeCopy {
private ArrayList<List<StoreFile>> stripeFiles = null;
private ArrayList<StoreFile> level0Files = null;
private ArrayList<byte[]> stripeEndRows = null;
private Collection<StoreFile> compactedFiles = null;
private Collection<StoreFile> results = null;
private List<StoreFile> l0Results = new ArrayList<>();
private final boolean isFlush;
public CompactionOrFlushMergeCopy(boolean isFlush) {
// Create a lazy mutable copy (other fields are so lazy they start out as nulls).
this.stripeFiles = new ArrayList<>(StripeStoreFileManager.this.state.stripeFiles);
this.isFlush = isFlush;
}
private void mergeResults(Collection<StoreFile> compactedFiles, Collection<StoreFile> results)
throws IOException {
assert this.compactedFiles == null && this.results == null;
this.compactedFiles = compactedFiles;
this.results = results;
// Do logical processing.
if (!isFlush) removeCompactedFiles();
TreeMap<byte[], StoreFile> newStripes = processResults();
if (newStripes != null) {
processNewCandidateStripes(newStripes);
}
// Create new state and update parent.
State state = createNewState(false);
StripeStoreFileManager.this.state = state;
updateMetadataMaps();
}
private void deleteResults(Collection<StoreFile> compactedFiles) throws IOException {
this.compactedFiles = compactedFiles;
// Create new state and update parent.
State state = createNewState(true);
StripeStoreFileManager.this.state = state;
updateMetadataMaps();
}
private State createNewState(boolean delCompactedFiles) {
State oldState = StripeStoreFileManager.this.state;
// Stripe count should be the same unless the end rows changed.
assert oldState.stripeFiles.size() == this.stripeFiles.size() || this.stripeEndRows != null;
State newState = new State();
newState.level0Files = (this.level0Files == null) ? oldState.level0Files
: ImmutableList.copyOf(this.level0Files);
newState.stripeEndRows = (this.stripeEndRows == null) ? oldState.stripeEndRows
: this.stripeEndRows.toArray(new byte[this.stripeEndRows.size()][]);
newState.stripeFiles = new ArrayList<>(this.stripeFiles.size());
for (List<StoreFile> newStripe : this.stripeFiles) {
newState.stripeFiles.add(newStripe instanceof ImmutableList<?>
? (ImmutableList<StoreFile>)newStripe : ImmutableList.copyOf(newStripe));
}
List<StoreFile> newAllFiles = new ArrayList<>(oldState.allFilesCached);
List<StoreFile> newAllCompactedFiles = new ArrayList<>(oldState.allCompactedFilesCached);
if (!isFlush) {
newAllFiles.removeAll(compactedFiles);
if (delCompactedFiles) {
newAllCompactedFiles.removeAll(compactedFiles);
} else {
newAllCompactedFiles.addAll(compactedFiles);
}
}
if (results != null) {
newAllFiles.addAll(results);
}
newState.allFilesCached = ImmutableList.copyOf(newAllFiles);
newState.allCompactedFilesCached = ImmutableList.copyOf(newAllCompactedFiles);
return newState;
}
private void updateMetadataMaps() {
StripeStoreFileManager parent = StripeStoreFileManager.this;
if (!isFlush) {
for (StoreFile sf : this.compactedFiles) {
parent.fileStarts.remove(sf);
parent.fileEnds.remove(sf);
}
}
if (this.l0Results != null) {
for (StoreFile sf : this.l0Results) {
parent.ensureLevel0Metadata(sf);
}
}
}
/**
* @param index Index of the stripe we need.
* @return A lazy stripe copy from current stripes.
*/
private final ArrayList<StoreFile> getStripeCopy(int index) {
List<StoreFile> stripeCopy = this.stripeFiles.get(index);
ArrayList<StoreFile> result = null;
if (stripeCopy instanceof ImmutableList<?>) {
result = new ArrayList<>(stripeCopy);
this.stripeFiles.set(index, result);
} else {
result = (ArrayList<StoreFile>)stripeCopy;
}
return result;
}
/**
* @return A lazy L0 copy from current state.
*/
private final ArrayList<StoreFile> getLevel0Copy() {
if (this.level0Files == null) {
this.level0Files = new ArrayList<>(StripeStoreFileManager.this.state.level0Files);
}
return this.level0Files;
}
/**
* Process new files, and add them either to the structure of existing stripes,
* or to the list of new candidate stripes.
* @return New candidate stripes.
*/
private TreeMap<byte[], StoreFile> processResults() throws IOException {
TreeMap<byte[], StoreFile> newStripes = null;
for (StoreFile sf : this.results) {
byte[] startRow = startOf(sf), endRow = endOf(sf);
if (isInvalid(endRow) || isInvalid(startRow)) {
if (!isFlush) {
LOG.warn("The newly compacted file doesn't have stripes set: " + sf.getPath());
}
insertFileIntoStripe(getLevel0Copy(), sf);
this.l0Results.add(sf);
continue;
}
if (!this.stripeFiles.isEmpty()) {
int stripeIndex = findStripeIndexByEndRow(endRow);
if ((stripeIndex >= 0) && rowEquals(getStartRow(stripeIndex), startRow)) {
// Simple/common case - add file to an existing stripe.
insertFileIntoStripe(getStripeCopy(stripeIndex), sf);
continue;
}
}
// Make a new candidate stripe.
if (newStripes == null) {
newStripes = new TreeMap<>(MAP_COMPARATOR);
}
StoreFile oldSf = newStripes.put(endRow, sf);
if (oldSf != null) {
throw new IOException("Compactor has produced multiple files for the stripe ending in ["
+ Bytes.toString(endRow) + "], found " + sf.getPath() + " and " + oldSf.getPath());
}
}
return newStripes;
}
/**
* Remove compacted files.
* @param compactedFiles Compacted files.
*/
private void removeCompactedFiles() throws IOException {
for (StoreFile oldFile : this.compactedFiles) {
byte[] oldEndRow = endOf(oldFile);
List<StoreFile> source = null;
if (isInvalid(oldEndRow)) {
source = getLevel0Copy();
} else {
int stripeIndex = findStripeIndexByEndRow(oldEndRow);
if (stripeIndex < 0) {
throw new IOException("An allegedly compacted file [" + oldFile + "] does not belong"
+ " to a known stripe (end row - [" + Bytes.toString(oldEndRow) + "])");
}
source = getStripeCopy(stripeIndex);
}
if (!source.remove(oldFile)) {
throw new IOException("An allegedly compacted file [" + oldFile + "] was not found");
}
}
}
/**
* See {@link #addCompactionResults(Collection, Collection)} - updates the stripe list with
* new candidate stripes/removes old stripes; produces new set of stripe end rows.
* @param newStripes New stripes - files by end row.
*/
private void processNewCandidateStripes(
TreeMap<byte[], StoreFile> newStripes) throws IOException {
// Validate that the removed and added aggregate ranges still make for a full key space.
boolean hasStripes = !this.stripeFiles.isEmpty();
this.stripeEndRows = new ArrayList<>(Arrays.asList(StripeStoreFileManager.this.state.stripeEndRows));
int removeFrom = 0;
byte[] firstStartRow = startOf(newStripes.firstEntry().getValue());
byte[] lastEndRow = newStripes.lastKey();
if (!hasStripes && (!isOpen(firstStartRow) || !isOpen(lastEndRow))) {
throw new IOException("Newly created stripes do not cover the entire key space.");
}
boolean canAddNewStripes = true;
Collection<StoreFile> filesForL0 = null;
if (hasStripes) {
// Determine which stripes will need to be removed because they conflict with new stripes.
// The new boundaries should match old stripe boundaries, so we should get exact matches.
if (isOpen(firstStartRow)) {
removeFrom = 0;
} else {
removeFrom = findStripeIndexByEndRow(firstStartRow);
if (removeFrom < 0) throw new IOException("Compaction is trying to add a bad range.");
++removeFrom;
}
int removeTo = findStripeIndexByEndRow(lastEndRow);
if (removeTo < 0) throw new IOException("Compaction is trying to add a bad range.");
// See if there are files in the stripes we are trying to replace.
ArrayList<StoreFile> conflictingFiles = new ArrayList<>();
for (int removeIndex = removeTo; removeIndex >= removeFrom; --removeIndex) {
conflictingFiles.addAll(this.stripeFiles.get(removeIndex));
}
if (!conflictingFiles.isEmpty()) {
// This can be caused by two things - concurrent flush into stripes, or a bug.
// Unfortunately, we cannot tell them apart without looking at timing or something
// like that. We will assume we are dealing with a flush and dump it into L0.
if (isFlush) {
long newSize = StripeCompactionPolicy.getTotalFileSize(newStripes.values());
LOG.warn("Stripes were created by a flush, but results of size " + newSize
+ " cannot be added because the stripes have changed");
canAddNewStripes = false;
filesForL0 = newStripes.values();
} else {
long oldSize = StripeCompactionPolicy.getTotalFileSize(conflictingFiles);
LOG.info(conflictingFiles.size() + " conflicting files (likely created by a flush) "
+ " of size " + oldSize + " are moved to L0 due to concurrent stripe change");
filesForL0 = conflictingFiles;
}
if (filesForL0 != null) {
for (StoreFile sf : filesForL0) {
insertFileIntoStripe(getLevel0Copy(), sf);
}
l0Results.addAll(filesForL0);
}
}
if (canAddNewStripes) {
// Remove old empty stripes.
int originalCount = this.stripeFiles.size();
for (int removeIndex = removeTo; removeIndex >= removeFrom; --removeIndex) {
if (removeIndex != originalCount - 1) {
this.stripeEndRows.remove(removeIndex);
}
this.stripeFiles.remove(removeIndex);
}
}
}
if (!canAddNewStripes) return; // Files were already put into L0.
// Now, insert new stripes. The total ranges match, so we can insert where we removed.
byte[] previousEndRow = null;
int insertAt = removeFrom;
for (Map.Entry<byte[], StoreFile> newStripe : newStripes.entrySet()) {
if (previousEndRow != null) {
// Validate that the ranges are contiguous.
assert !isOpen(previousEndRow);
byte[] startRow = startOf(newStripe.getValue());
if (!rowEquals(previousEndRow, startRow)) {
throw new IOException("The new stripes produced by "
+ (isFlush ? "flush" : "compaction") + " are not contiguous");
}
}
// Add the new stripe.
ArrayList<StoreFile> tmp = new ArrayList<>();
tmp.add(newStripe.getValue());
stripeFiles.add(insertAt, tmp);
previousEndRow = newStripe.getKey();
if (!isOpen(previousEndRow)) {
stripeEndRows.add(insertAt, previousEndRow);
}
++insertAt;
}
}
}
@Override
public List<StoreFile> getLevel0Files() {
return this.state.level0Files;
}
@Override
public List<byte[]> getStripeBoundaries() {
if (this.state.stripeFiles.isEmpty()) return new ArrayList<>();
ArrayList<byte[]> result = new ArrayList<>(this.state.stripeEndRows.length + 2);
result.add(OPEN_KEY);
Collections.addAll(result, this.state.stripeEndRows);
result.add(OPEN_KEY);
return result;
}
@Override
public ArrayList<ImmutableList<StoreFile>> getStripes() {
return this.state.stripeFiles;
}
@Override
public int getStripeCount() {
return this.state.stripeFiles.size();
}
@Override
public Collection<StoreFile> getUnneededFiles(long maxTs, List<StoreFile> filesCompacting) {
// 1) We can never get rid of the last file which has the maximum seqid in a stripe.
// 2) Files that are not the latest can't become one due to (1), so the rest are fair game.
State state = this.state;
Collection<StoreFile> expiredStoreFiles = null;
for (ImmutableList<StoreFile> stripe : state.stripeFiles) {
expiredStoreFiles = findExpiredFiles(stripe, maxTs, filesCompacting, expiredStoreFiles);
}
return findExpiredFiles(state.level0Files, maxTs, filesCompacting, expiredStoreFiles);
}
private Collection<StoreFile> findExpiredFiles(ImmutableList<StoreFile> stripe, long maxTs,
List<StoreFile> filesCompacting, Collection<StoreFile> expiredStoreFiles) {
// Order by seqnum is reversed.
for (int i = 1; i < stripe.size(); ++i) {
StoreFile sf = stripe.get(i);
synchronized (sf) {
long fileTs = sf.getReader().getMaxTimestamp();
if (fileTs < maxTs && !filesCompacting.contains(sf)) {
LOG.info("Found an expired store file: " + sf.getPath() + " whose maxTimeStamp is "
+ fileTs + ", which is below " + maxTs);
if (expiredStoreFiles == null) {
expiredStoreFiles = new ArrayList<>();
}
expiredStoreFiles.add(sf);
}
}
}
return expiredStoreFiles;
}
@Override
public double getCompactionPressure() {
State stateLocal = this.state;
if (stateLocal.allFilesCached.size() > blockingFileCount) {
// just a hit to tell others that we have reached the blocking file count.
return 2.0;
}
if (stateLocal.stripeFiles.isEmpty()) {
return 0.0;
}
int blockingFilePerStripe = blockingFileCount / stateLocal.stripeFiles.size();
// do not calculate L0 separately because data will be moved to stripe quickly and in most cases
// we flush data to stripe directly.
int delta = stateLocal.level0Files.isEmpty() ? 0 : 1;
double max = 0.0;
for (ImmutableList<StoreFile> stripeFile : stateLocal.stripeFiles) {
int stripeFileCount = stripeFile.size();
double normCount =
(double) (stripeFileCount + delta - config.getStripeCompactMinFiles())
/ (blockingFilePerStripe - config.getStripeCompactMinFiles());
if (normCount >= 1.0) {
// This could happen if stripe is not split evenly. Do not return values that larger than
// 1.0 because we have not reached the blocking file count actually.
return 1.0;
}
if (normCount > max) {
max = normCount;
}
}
return max;
}
@Override
public Comparator<StoreFile> getStoreFileComparator() {
return StoreFile.Comparators.SEQ_ID;
}
}