HTTP/2：二进制分帧与多路复用

连接 (Connection)  一条 TCP 连接
 └── 流 (Stream)   一次请求-响应 = 一个流，有唯一 ID（客户端发起用奇数）
      └── 帧 (Frame) 流被切成的最小传输单位，带着流 ID
 
一条连接上的字节序列（多路复用）：
[流1-HEADERS][流3-HEADERS][流1-DATA][流5-HEADERS][流3-DATA][流1-DATA]...
       ▲ 不同流的帧交错发送，接收方按流 ID 重组，互不等待

+-----------------------------------------------+
| Length (24 bit)  载荷长度                       |
+---------------+---------------+---------------+
| Type (8)      | Flags (8)     |                |
+-+-------------+---------------+---------------+
|R|  Stream Identifier (31 bit) 流 ID            |
+=+=============================================+
|  Frame Payload 载荷 ...                        |
+-----------------------------------------------+

python3 -c '
import struct
 
FRAME_TYPES = {0: "DATA", 1: "HEADERS", 3: "RST_STREAM", 4: "SETTINGS",
               7: "GOAWAY", 8: "WINDOW_UPDATE"}
 
def build_frame(ftype, flags, stream_id, payload):
    length = len(payload)
    # 24bit 长度 + 8bit 类型 + 8bit 标志 + 32bit 流ID(最高位保留)
    header = struct.pack(">I", length)[1:] + bytes([ftype, flags]) + struct.pack(">I", stream_id)
    return header + payload
 
def parse_frame(raw):
    length = int.from_bytes(raw[0:3], "big")
    ftype, flags = raw[3], raw[4]
    stream_id = int.from_bytes(raw[5:9], "big") & 0x7FFFFFFF
    return length, FRAME_TYPES.get(ftype, ftype), flags, stream_id, raw[9:9+length]
 
# 模拟一条连接上交错的三个帧（两个流）
wire = b""
wire += build_frame(1, 0x4, 1, b"\x82\x86\x84")        # 流1 HEADERS(END_HEADERS)
wire += build_frame(1, 0x4, 3, b"\x82\x86\x84")        # 流3 HEADERS
wire += build_frame(0, 0x1, 1, b"hello stream 1")       # 流1 DATA(END_STREAM)
 
print("线路上的原始字节:", wire.hex())
print()
pos = 0
while pos < len(wire):
    length, ftype, flags, sid, payload = parse_frame(wire[pos:])
    print("帧:", ftype, "| 流ID:", sid, "| flags:", hex(flags),
          "| 载荷:", payload if ftype == "DATA" else payload.hex())
    pos += 9 + length
'

python3 -c '
# 简化模拟：首次全量 + 入表，之后只发 1 字节索引
headers = [
    ("user-agent", "Mozilla/5.0 (Macintosh; Intel Mac OS X) Chrome/126.0"),
    ("cookie", "sid=a1b2c3d4e5f6; theme=dark; lang=zh-CN; tracker=xyz789"),
    ("accept", "text/html,application/xhtml+xml,application/xml;q=0.9"),
]
 
dynamic_table = {}
def hpack_size(kv):
    if kv in dynamic_table:
        return 1                       # 命中动态表：一个索引字节
    dynamic_table[kv] = len(dynamic_table) + 62
    k, v = kv
    return 2 + len(k) + len(v)         # 首次：长度前缀+字面量（忽略哈夫曼）
 
raw_per_req = sum(len(k) + 2 + len(v) + 2 for k, v in headers)
 
total_raw, total_hpack = 0, 0
for i in range(1, 101):
    total_raw += raw_per_req
    total_hpack += sum(hpack_size(kv) for kv in headers)
    if i in (1, 2, 100):
        print("第", i, "个请求后  明文累计:", total_raw, "字节   HPACK 累计:", total_hpack, "字节")
 
print()
print("100 个请求节省:", round(100 - total_hpack * 100 / total_raw), "%")
'

弱网丢包时的讽刺场景：
HTTP/1.1: 6 条连接，丢包只堵其中 1 条 —— 另外 5 条照跑
HTTP/2:   1 条连接，丢包全体罚站     —— 高丢包率下反而更慢