public MemtableUnfilteredPartitionIterator makePartitionIterator(
final ColumnFilter columnFilter, final DataRange dataRange, final boolean isForThrift) {
AbstractBounds<PartitionPosition> keyRange = dataRange.keyRange();
boolean startIsMin = keyRange.left.isMinimum();
boolean stopIsMin = keyRange.right.isMinimum();
boolean isBound = keyRange instanceof Bounds;
boolean includeStart = isBound || keyRange instanceof IncludingExcludingBounds;
boolean includeStop = isBound || keyRange instanceof Range;
Map<PartitionPosition, AtomicBTreePartition> subMap;
if (startIsMin)
subMap = stopIsMin ? partitions : partitions.headMap(keyRange.right, includeStop);
else
subMap =
stopIsMin
? partitions.tailMap(keyRange.left, includeStart)
: partitions.subMap(keyRange.left, includeStart, keyRange.right, includeStop);
int minLocalDeletionTime = Integer.MAX_VALUE;
// avoid iterating over the memtable if we purge all tombstones
if (cfs.getCompactionStrategyManager().onlyPurgeRepairedTombstones())
minLocalDeletionTime = findMinLocalDeletionTime(subMap.entrySet().iterator());
final Iterator<Map.Entry<PartitionPosition, AtomicBTreePartition>> iter =
subMap.entrySet().iterator();
return new MemtableUnfilteredPartitionIterator(
cfs, iter, isForThrift, minLocalDeletionTime, columnFilter, dataRange);
}
/**
* Should only be called by ColumnFamilyStore.apply via Keyspace.apply, which supplies the
* appropriate OpOrdering.
*
* <p>replayPosition should only be null if this is a secondary index, in which case it is
* *expected* to be null
*/
long put(PartitionUpdate update, UpdateTransaction indexer, OpOrder.Group opGroup) {
AtomicBTreePartition previous = partitions.get(update.partitionKey());
long initialSize = 0;
if (previous == null) {
final DecoratedKey cloneKey = allocator.clone(update.partitionKey(), opGroup);
AtomicBTreePartition empty = new AtomicBTreePartition(cfs.metadata, cloneKey, allocator);
// We'll add the columns later. This avoids wasting works if we get beaten in the putIfAbsent
previous = partitions.putIfAbsent(cloneKey, empty);
if (previous == null) {
previous = empty;
// allocate the row overhead after the fact; this saves over allocating and having to free
// after, but
// means we can overshoot our declared limit.
int overhead = (int) (cloneKey.getToken().getHeapSize() + ROW_OVERHEAD_HEAP_SIZE);
allocator.onHeap().allocate(overhead, opGroup);
initialSize = 8;
} else {
allocator.reclaimer().reclaimImmediately(cloneKey);
}
}
long[] pair = previous.addAllWithSizeDelta(update, opGroup, indexer);
minTimestamp = Math.min(minTimestamp, previous.stats().minTimestamp);
liveDataSize.addAndGet(initialSize + pair[0]);
columnsCollector.update(update.columns());
statsCollector.update(update.stats());
currentOperations.addAndGet(update.operationCount());
return pair[1];
}
@SuppressWarnings("resource") // log and writer closed by SSTableTxnWriter
public SSTableTxnWriter createFlushWriter(
String filename, PartitionColumns columns, EncodingStats stats) {
// we operate "offline" here, as we expose the resulting reader consciously when done
// (although we may want to modify this behaviour in future, to encapsulate full flush behaviour
// in LifecycleTransaction)
LifecycleTransaction txn = null;
try {
txn = LifecycleTransaction.offline(OperationType.FLUSH);
MetadataCollector sstableMetadataCollector =
new MetadataCollector(cfs.metadata.comparator)
.commitLogIntervals(
new IntervalSet(commitLogLowerBound.get(), commitLogUpperBound.get()));
return new SSTableTxnWriter(
txn,
cfs.createSSTableMultiWriter(
Descriptor.fromFilename(filename),
(long) partitions.size(),
ActiveRepairService.UNREPAIRED_SSTABLE,
sstableMetadataCollector,
new SerializationHeader(true, cfs.metadata, columns, stats),
txn));
} catch (Throwable t) {
if (txn != null) txn.close();
throw t;
}
}
private static int estimateRowOverhead(final int count) {
// calculate row overhead
try (final OpOrder.Group group = new OpOrder().start()) {
int rowOverhead;
MemtableAllocator allocator = MEMORY_POOL.newAllocator();
ConcurrentNavigableMap<PartitionPosition, Object> partitions = new ConcurrentSkipListMap<>();
final Object val = new Object();
for (int i = 0; i < count; i++)
partitions.put(
allocator.clone(
new BufferDecoratedKey(new LongToken(i), ByteBufferUtil.EMPTY_BYTE_BUFFER), group),
val);
double avgSize = ObjectSizes.measureDeep(partitions) / (double) count;
rowOverhead =
(int) ((avgSize - Math.floor(avgSize)) < 0.05 ? Math.floor(avgSize) : Math.ceil(avgSize));
rowOverhead -= ObjectSizes.measureDeep(new LongToken(0));
rowOverhead += AtomicBTreePartition.EMPTY_SIZE;
allocator.setDiscarding();
allocator.setDiscarded();
return rowOverhead;
}
}
private long estimatedSize() {
long keySize = 0;
for (PartitionPosition key : partitions.keySet()) {
// make sure we don't write non-sensical keys
assert key instanceof DecoratedKey;
keySize += ((DecoratedKey) key).getKey().remaining();
}
return (long)
((keySize // index entries
+ keySize // keys in data file
+ liveDataSize.get()) // data
* 1.2); // bloom filter and row index overhead
}
private Collection<SSTableReader> writeSortedContents(File sstableDirectory) {
boolean isBatchLogTable =
cfs.name.equals(SystemKeyspace.BATCHES)
&& cfs.keyspace.getName().equals(SystemKeyspace.NAME);
logger.debug("Writing {}", Memtable.this.toString());
Collection<SSTableReader> ssTables;
try (SSTableTxnWriter writer =
createFlushWriter(
cfs.getSSTablePath(sstableDirectory), columnsCollector.get(), statsCollector.get())) {
boolean trackContention = logger.isTraceEnabled();
int heavilyContendedRowCount = 0;
// (we can't clear out the map as-we-go to free up memory,
// since the memtable is being used for queries in the "pending flush" category)
for (AtomicBTreePartition partition : partitions.values()) {
// Each batchlog partition is a separate entry in the log. And for an entry, we only do 2
// operations: 1) we insert the entry and 2) we delete it. Further, BL data is strictly
// local,
// we don't need to preserve tombstones for repair. So if both operation are in this
// memtable (which will almost always be the case if there is no ongoing failure), we can
// just skip the entry (CASSANDRA-4667).
if (isBatchLogTable && !partition.partitionLevelDeletion().isLive() && partition.hasRows())
continue;
if (trackContention && partition.usePessimisticLocking()) heavilyContendedRowCount++;
if (!partition.isEmpty()) {
try (UnfilteredRowIterator iter = partition.unfilteredIterator()) {
writer.append(iter);
}
}
}
if (writer.getFilePointer() > 0) {
logger.debug(
String.format(
"Completed flushing %s (%s) for commitlog position %s",
writer.getFilename(),
FBUtilities.prettyPrintMemory(writer.getFilePointer()),
commitLogUpperBound));
// sstables should contain non-repaired data.
ssTables = writer.finish(true);
} else {
logger.debug(
"Completed flushing {}; nothing needed to be retained. Commitlog position was {}",
writer.getFilename(),
commitLogUpperBound);
writer.abort();
ssTables = Collections.emptyList();
}
if (heavilyContendedRowCount > 0)
logger.trace(
String.format(
"High update contention in %d/%d partitions of %s ",
heavilyContendedRowCount, partitions.size(), Memtable.this.toString()));
return ssTables;
}
}
public Partition getPartition(DecoratedKey key) {
return partitions.get(key);
}
public int partitionCount() {
return partitions.size();
}
public boolean isClean() {
return partitions.isEmpty();
}
