public Entry<InternalKey, Slice> next(int bucket, long minSeq) {
if (hasNext()) {
// parse the key in the block
Entry<InternalKey, Slice> entry = iterator.next();
InternalKey internalKey = entry.getKey();
Preconditions.checkState(
internalKey != null, "Corrupt key for %s", internalKey.getUserKey().toString(UTF_8));
if (internalKey.bucket() == bucket && internalKey.getSequenceNumber() > minSeq) {
return entry;
}
}
return null;
}
private int readWriteBatch(int bucket, SliceInput record, int updateSize, Entrys entries)
throws IOException {
List<Entry<Slice, Slice>> list = Lists.newArrayListWithCapacity(updateSize);
int count = 0;
while (record.isReadable()) {
count++;
ValueType valueType = ValueType.getValueTypeByPersistentId(record.readByte());
if (valueType == VALUE) {
Slice key = readLengthPrefixedBytes(record);
Slice value = readLengthPrefixedBytes(record);
list.add(new BlockEntry(key, value));
} else if (valueType == DELETION) {
Slice key = readLengthPrefixedBytes(record);
list.add(new BlockEntry(key, Slices.EMPTY_SLICE));
} else {
throw new IllegalStateException("Unexpected value type " + valueType);
}
}
if (count != updateSize) {
throw new IOException(
String.format(
"Expected %d entries in log record but found %s entries", updateSize, entries));
}
int available = 0;
for (Entry<Slice, Slice> e : list) {
InternalKey iKey =
new InternalKey(e.getKey(), 0, e.getValue().length() != 0 ? VALUE : DELETION);
if (bucket == iKey.bucket()) {
entries.add(e);
available++;
}
}
return available;
}
private void readSst(
String standby,
int bucket,
Position pos,
int batchSize,
Version v,
FileQueue queue,
Entrys entrys,
boolean availableRollback,
int deep) {
// 防止调用超时，和栈溢出异常
if (deep > 50) {
return;
}
FileMetaData file = null;
if (pos.fileNumber() < 0) {
file = queue.poll();
} else {
file = queue.poll(pos.fileNumber());
}
// 当file为null时，意味着当前的version所有sst文件已经迁移完成，
// 如果存在新的version，就切换到新的version，否则切换到log模式，开始去拉去log文件的数据
if (file == null) {
v.release();
pos.switchToNew();
Version snapshot = versions.getCurrent();
if (v != snapshot) {
v = snapshot;
v.retain();
pos.version(v.hashCode());
pos.fileNumber(-1);
readSst(
standby,
bucket,
pos,
batchSize,
v,
new FileQueue(bucket, v, pos.maxFileNumber()),
entrys,
availableRollback,
++deep);
} else {
// 日志模式不锁定version,只锁定log文件
// log文件的锁定逻辑见DBImpl的logQ属性
pos.fileNumber(log.getFileNumber());
pos.pointer(0);
pos.fileType(LOG);
}
return;
} else {
pos.fileNumber(file.getNumber());
}
// 读取大于maxSeq的记录
String key = generateKey(standby, bucket);
SlaveIterator si = v.iterators.get(key);
if (si == null) {
si = new SlaveIterator(file);
} else {
if (si.getFileNumber() != file.getNumber()) {
si = new SlaveIterator(file);
}
}
v.iterators.put(key, si);
if (availableRollback) {
if (pos.pointer() > 0) {
logger.warn("{}-bucket sst happen rollback, position={}.", bucket, pos);
si.seekToFirst();
} else {
pos.pointer(file.getNumber());
}
}
long maxSeq = pos.maxSeq();
while (si.hasNext()) {
Entry<InternalKey, Slice> e = si.next(bucket, maxSeq);
if (e != null) {
InternalKey ikey = e.getKey();
pos.curMaxSeq(ikey.getSequenceNumber());
entrys.add(
new BlockEntry(
ikey.getUserKey(),
ValueType.VALUE.equals(ikey.getValueType()) ? e.getValue() : Slices.EMPTY_SLICE));
}
if (entrys.size() > batchSize) {
break;
}
}
// 判断文件是否已经读取完成
if (!si.hasNext()) {
v.iterators.remove(key);
FileMetaData fmd = queue.peek();
if (fmd != null) {
pos.fileNumber(fmd.getNumber());
} else {
pos.fileNumber(pos.curMaxFileNumber + 1); // +1意味着该version不存在该文件
}
}
// 继续读取下一个文件的数据
if (entrys.size() < batchSize) {
readSst(standby, bucket, pos, batchSize, v, queue, entrys, false, ++deep);
}
}