TCompactProtocol协议作为TBinaryProtocol协议的升级强化版,都作为二进制编码传输方式,采用了一种乐器MIDI文件的编码方法(wiki,百度下),简单介绍下两种思想:
1: ZigZag有符号数编码,如表格所示:
编码前 | 编码后 |
0 | 0 |
-1 | 1 |
1 | 2 |
-2 | 3 |
2 | 4 |
-3 | 5 |
其效果等效于正数等于原先 * 2,负数变正数。
32bits int = (i << 1) ^ (i >> 31), 64bits long = (l << 1) ^ (l >> 63)
2:VLQ(variable-length quantity)编码:
即一字节的最高位(MHB)为标志位,不参与具体的内容,意思数值的大小仅仅有其它七位来表示。当最高位bit为1时,表示下一个byte也是该数值的内容(下一个byte的低七位bits);当最高位bit为0时,下一个byte不参与其中。通过这样的方式,而不是int固定的4个bytes,long 8个bytes来讲,对于小数,能节约不少的空间大小;但凡事有利有弊,当数值比较大时,就要占用更多的空间,例如较大的int ,需要5bytes,较大的long需要10bytes.
两者的结合 :
当VLQ编码遇到负数时,例如:long -1; 0XFFFFFFFFFFFFFFFF,就需要10bytes了,通过和ZigZag的结合,吧负数转变相应的正数。当正数,负数的 |数值|较小时,都可以通过两者的结合,有效的压缩占用的空间大小。但同上,数值较大不可避免的占用比平常正常编码更多的空间。
源码分析:
首先来看一下int32,long64的ZigZag编码:
private long longToZigzag(long l) { return (l << 1) ^ (l >> 63); } /** * Convert n into a zigzag int. This allows negative numbers to be * represented compactly as a varint. */ private int intToZigZag(int n) { return (n << 1) ^ (n >> 31);//正数 n << 1 扩大两倍 , n >> 31 = 0 , ^ 0 不变 ,2 * n ; }
再看看int32,long64的varint写法:
byte[] i32buf = new byte[5]; //int32 最大需要5个字节 private void writeVarint32(int n) throws TException { int idx = 0; //index flag while (true) { if ((n & ~0x7F) == 0) { // if (n <= 2^7) 1byte i32buf[idx++] = (byte)n; // writeByteDirect((byte)n); break; // return; } else { i32buf[idx++] = (byte)((n & 0x7F) | 0x80); 、//else if(n > 2^ 7) 按小端方式给byte第八位贴上1标签,存放在buf。 // writeByteDirect((byte)((n & 0x7F) | 0x80)); n >>>= 7; //逻辑右移7bit,再次判断,loop } } trans_.write(i32buf, 0, idx); //吧buf写入传输层 } /** * Write an i64 as a varint. Results in 1-10 bytes on the wire. */ byte[] varint64out = new byte[10];//最大需要10bytes private void writeVarint64(long n) throws TException { int idx = 0; while (true) { if ((n & ~0x7FL) == 0) { //注意这边的 ~0x7FL(不能写成0x7F) varint64out[idx++] = (byte)n; break; } else { varint64out[idx++] = ((byte)((n & 0x7F) | 0x80)); n >>>= 7; } } trans_.write(varint64out, 0, idx); }
上面注解说明了varint的系统操作,预分配最大字节buffer,然后按照小端方式写入VLQ编码后实际内容。再来看看系统是怎么结合两者的:
public void writeI32(int i32) throws TException { writeVarint32(intToZigZag(i32)); //先调intToZigZag转换,在write VLQ。 } /** * Write an i64 as a zigzag varint. */ public void writeI64(long i64) throws TException { writeVarint64(longToZigzag(i64)); } public void writeI16(short i16) throws TException { //i16先按int32 zigzag编码转换 然后按VLQ转换 writeVarint32(intToZigZag(i16)); }
我们先系统的看一下TCompactProtocol按什么方法写入Thrift内部数据类型的,然后再看message的写法,一下是thrift内部数据类型,i16,i32,i64已经看完,在来看看别的:
private static class Types { public static final byte BOOLEAN_TRUE = 0x01; public static final byte BOOLEAN_FALSE = 0x02; public static final byte BYTE = 0x03; public static final byte I16 = 0x04; public static final byte I32 = 0x05; public static final byte I64 = 0x06; public static final byte DOUBLE = 0x07; public static final byte BINARY = 0x08; public static final byte LIST = 0x09; public static final byte SET = 0x0A; public static final byte MAP = 0x0B; public static final byte STRUCT = 0x0C; }
boolean:
public void writeBool(boolean b) throws TException { if (booleanField_ != null) { // we haven‘t written the field header yet writeFieldBeginInternal(booleanField_, b ? Types.BOOLEAN_TRUE : Types.BOOLEAN_FALSE); booleanField_ = null; } else { // we‘re not part of a field, so just write the value. writeByteDirect(b ? Types.BOOLEAN_TRUE : Types.BOOLEAN_FALSE);//按照上面对应的boolean_yes,boolean_no字节值写入。 } }
TCompactProtocol写入Boolean分两种情况,1:该boolean值为TStruct中的内部成员时TField时,得写入header数据(即内容和数据类型压缩在一起写);2 :如果不为TField内部类型的话,直接按byte写入。关于TStruct和TField的细节请参照上篇。
具体tstruct写入,稍后分析。
byte:
public void writeByte(byte b) throws TException { writeByteDirect(b);//one byte 直接写入。 }
private byte[] byteDirectBuffer = new byte[1]; private void writeByteDirect(byte b) throws TException { byteDirectBuffer[0] = b; trans_.write(byteDirectBuffer); }
double:
public void writeDouble(double dub) throws TException { byte[] data = new byte[]{0, 0, 0, 0, 0, 0, 0, 0}; //8个字节 fixedLongToBytes(Double.doubleToLongBits(dub), data, 0); //double 转long bit 分布,然后按照fix64编码传输。 trans_.write(data); }
private void fixedLongToBytes(long n, byte[] buf, int off) { buf[off+0] = (byte)( n & 0xff); buf[off+1] = (byte)((n >> 8 ) & 0xff); buf[off+2] = (byte)((n >> 16) & 0xff); buf[off+3] = (byte)((n >> 24) & 0xff); buf[off+4] = (byte)((n >> 32) & 0xff); buf[off+5] = (byte)((n >> 40) & 0xff); buf[off+6] = (byte)((n >> 48) & 0xff); buf[off+7] = (byte)((n >> 56) & 0xff); }
可以看出double类型,先按Double.doubletoLongBits()转换后,按照fixed64编码写入(8字节小端写入),如上。
bytearray:
public void writeBinary(ByteBuffer bin) throws TException { int length = bin.limit() - bin.position();//计算数据len writeBinary(bin.array(), bin.position() + bin.arrayOffset(), length); }
private void writeBinary(byte[] buf, int offset, int length) throws TException { writeVarint32(length); //按VLQ编码写入len值,这里没有使用zigzag编码(zigzag编码主要解决负数VLQ编码占用大空间的情况,这里len不为负,直接VLQ写入) trans_.write(buf, offset, length);//写入实际内buff中内容 }
string:
public void writeString(String str) throws TException { try { byte[] bytes = str.getBytes("UTF-8");//utf-8编码,得到字节数组 writeBinary(bytes, 0, bytes.length);//抵用writeBinary,see 上面 } catch (UnsupportedEncodingException e) { throw new TException("UTF-8 not supported!"); } }
容器类型:
SetTag:
public void writeSetBegin(TSet set) throws TException { writeCollectionBegin(set.elemType, set.size);//set类型,长度值 }
type byte:
public final class TType { public static final byte STOP = 0; public static final byte VOID = 1;//java中没有这种类型,这里存在只是为了别的语言,可能 public static final byte BOOL = 2; public static final byte BYTE = 3; public static final byte DOUBLE = 4; public static final byte I16 = 6; public static final byte I32 = 8; public static final byte I64 = 10; public static final byte STRING = 11; public static final byte STRUCT = 12; public static final byte MAP = 13; public static final byte SET = 14; public static final byte LIST = 15; public static final byte ENUM = 16;//低下static {}中,该类型也没用到。 所以4bits 够用了 }
protected void writeCollectionBegin(byte elemType, int size) throws TException { if (size <= 14) { // 1110 writeByteDirect(size << 4 | getCompactType(elemType));//size <= 14时,size << 4 | 对应的TTyte,压缩从一个byte写入。 } else { writeByteDirect(0xf0 | getCompactType(elemType));// 1111 0000| ttype ,按one byte写入 writeVarint32(size);// VLQ编码写入len } }
getCompactType(xx):
private byte getCompactType(byte ttype) { return ttypeToCompactType[ttype]; }
static { ttypeToCompactType[TType.STOP] = TType.STOP; ttypeToCompactType[TType.BOOL] = Types.BOOLEAN_TRUE; ttypeToCompactType[TType.BYTE] = Types.BYTE; ttypeToCompactType[TType.I16] = Types.I16; ttypeToCompactType[TType.I32] = Types.I32; ttypeToCompactType[TType.I64] = Types.I64; ttypeToCompactType[TType.DOUBLE] = Types.DOUBLE; ttypeToCompactType[TType.STRING] = Types.BINARY; ttypeToCompactType[TType.LIST] = Types.LIST; ttypeToCompactType[TType.SET] = Types.SET; ttypeToCompactType[TType.MAP] = Types.MAP; ttypeToCompactType[TType.STRUCT] = Types.STRUCT; }
public void writeListEnd() throws TException {} //no-op 空操作,走个形式而已
list tag:
public void writeListBegin(TList list) throws TException { writeCollectionBegin(list.elemType, list.size); }
public void writeListEnd() throws TException {}
同上,就不重复了。
map tag:
public void writeMapBegin(TMap map) throws TException { if (map.size == 0) {//size == 0 writeByteDirect(0); //直接写入one byte 0完事。 } else { writeVarint32(map.size); //VLQ写入长度 writeByteDirect(getCompactType(map.keyType) << 4 | getCompactType(map.valueType)); //one byte 写入 keyType(TType),valueType(TType) (keyType << 4 | valueType) 与avro的map不同,其key } //type只能为string类型。 }
wirteMapEnd()也是no-op操作就不贴了。
介绍完内置类型的写入方式,可以介绍写message了。
public void writeMessageBegin(TMessage message) throws TException { writeByteDirect(PROTOCOL_ID); // 1000 0010 one byte protocol_id writeByteDirect((VERSION & VERSION_MASK) | ((message.type << TYPE_SHIFT_AMOUNT) & TYPE_MASK));// ((0000 0001 & 0001 1111) | (type << 5)) & 1110 0000); one byte高三位messageType | writeVarint32(message.seqid); //低五位version bits, VLQ编码写入message 的sequence increment id. writeString(message.name); //消息名,即方法名。 }
private static final byte PROTOCOL_ID = (byte)0x82;//1000 0010 private static final byte VERSION = 1; private static final byte VERSION_MASK = 0x1f; // 0001 1111 private static final byte TYPE_MASK = (byte)0xE0; // 1110 0000 private static final byte TYPE_BITS = 0x07; // 0000 0111 private static final int TYPE_SHIFT_AMOUNT = 5;
这里的version应该为了以后的version更新。byte类型的messageType(call, execption, oneway,reply)具体请见上篇TBinaryProtocol分析。为了发消息的完整性,还是贴出TServiceClient的sendBase()步骤:
protected void sendBase(String methodName, TBase args) throws TException { oprot_.writeMessageBegin(new TMessage(methodName, TMessageType.CALL, ++seqid_)); args.write(oprot_); oprot_.writeMessageEnd(); oprot_.getTransport().flush(); }
现在该进行TBASE的write()了,即方法参数和返回值的封装类写,还是以hello.thrift为例:
hellostring_args的write():
public void write(org.apache.thrift.protocol.TProtocol oprot) throws org.apache.thrift.TException { schemes.get(oprot.getScheme()).getScheme().write(oprot, this); }
schema的write():
public void write(org.apache.thrift.protocol.TProtocol oprot, helloString_args struct) throws org.apache.thrift.TException { struct.validate(); oprot.writeStructBegin(STRUCT_DESC); if (struct.para != null) { oprot.writeFieldBegin(PARA_FIELD_DESC); oprot.writeString(struct.para); oprot.writeFieldEnd(); } oprot.writeFieldStop(); oprot.writeStructEnd(); }
private static final org.apache.thrift.protocol.TStruct STRUCT_DESC = new org.apache.thrift.protocol.TStruct("helloString_args");// 方法参数封装类的TStruct表示。 private static final org.apache.thrift.protocol.TField PARA_FIELD_DESC = new org.apache.thrift.protocol.TField("para", org.apache.thrift.protocol.TType.STRING, (short)1);
ok,此处的oprot为TCompactProtocol,看看他的writeStructBegin():
public void writeStructBegin(TStruct struct) throws TException { lastField_.push(lastFieldId_);记住上次write struct 最后的field id. lastFieldId_ = 0; //从本次参数写开始。 }
private ShortStack lastField_ = new ShortStack(15); //用于存放Tstructs中的field id(也就是thrift定义文件中service方法参数的标号 1:,2:);用于跟踪当前struct或者之前struct的field id
接下来,写writeFieldBegin()吧:
public void writeFieldBegin(TField field) throws TException { if (field.type == TType.BOOL) { //如果该方法参数为boolean类型, // we want to possibly include the value, so we‘ll wait. booleanField_ = field; //这里先做下标记,等会和具体boolean值一块写,压缩嘛!一开始介绍些基本数据类型(上面)的boolean的两种情况,第一种指当boolean值为Tfield的话,压缩一下,跟这里相结合, } else { //这里先记录下header metadata,等写实际内容时,即writeBoolean在一块写。 writeFieldBeginInternal(field, (byte)-1); } }
private void writeFieldBeginInternal(TField field, byte typeOverride) throws TException { // short lastField = lastField_.pop(); // if there‘s a type override, use that. // -1获得其内置数据类型,如果非-1情况,(指的是boolean)直接写入其byte值 ,true 0x01,false 0x02 byte typeToWrite = typeOverride == -1 ? getCompactType(field.type) : typeOverride; // typeOverride为写Boolean值,特设的,对其优化,one byte写入 // check if we can use delta encoding for the field id 增量式编码前提,用one byte 4MSB来做增量式编码,所有field id之间的差不能大于15.每次写Tstruct(即一个方法参数的封装类,其中可能含有很多参数) if (field.id > lastFieldId_ && field.id - lastFieldId_ <= 15) { // 因为每次写struct时,都会设置last_fieldid_ = 0,所以都是一次方法RPC调用参数表示ID之间的比较。不会出现上次RPC方法调用的参数id和 // write them together //本次RPC方法调用参数id的比较。 writeByteDirect((field.id - lastFieldId_) << 4 | typeToWrite); //本次field id和上次field id做增量 << 4和复写标志做 |,用一个byte传输,压缩空间。 } else { // write them separate writeByteDirect(typeToWrite); //分开写 one byte 复写标志。 writeI16(field.id); //i16 (zigzag + vlq编码)写入,参数个数最大2^16个。 } lastFieldId_ = field.id; //重新复制lastfield_id // lastField_.push(field.id); }
然后就是写具体的参数值内容了,写完后写上writeFieldEnd()操作;
structs所有的参数都写完后,调用writeFieldStop():
public void writeFieldStop() throws TException { writeByteDirect(TType.STOP);// one byte value 0,占位符吧,标志读完了。 }
writeStructEnd():
public void writeStructEnd() throws TException { lastFieldId_ = lastField_.pop();//重新写structs时,会吧这值压入stack,并重新附上0. }
public void writeMessageEnd() throws TException {}
读操作就不分析了,朋友们可以参照了去看看。