HTTP 1.0 协议不支持分块输出，当响应体不为空时，需要在响应头中明确指定 Content-Length，以支持 HTTP 1.0 长连接。所以当把 lua_http10_buffering 设置为 on 输出 HTTP 1.0 响应时，ngx_lua 将缓存 ngx.say 和 ngx.print 的所有输出，同时延迟发送响应头直到接收到所有输出内容。这时 ngx_lua 可以计算响应体的总长度，并为 HTTP 1.0 客户端创建一个正确的 Content-Length 响应头。如果在正在执行的 Lua 代码中设置 Content-Length 响应头，这种缓冲模式将被禁用，即使已经将 lua_http10_buffering 指令设置为 on。
对于大型流式响应输出，禁用 lua_http10_buffering 以最小化内存占用非常重要。
请注意，一些常见的 HTTP 性能测试工具，例如 ab 和 http_load 默认发送 HTTP 1.0 请求。要强制 curl 发送 HTTP 1.0 请求，使用 -0 选项。
Engslish Source
The HTTP 1.0 protocol does not support chunked output and requires an explicit Content-Length header when the response body is not empty in order to support the HTTP 1.0 keep-alive.
So when a HTTP 1.0 request is made and the lua_http10_buffering directive is turned on, ngx_lua will buffer the
output of ngx.say and ngx.print calls and also postpone sending response headers until all the response body output is received.
At that time ngx_lua can calculate the total length of the body and construct a proper Content-Length header to return to the HTTP 1.0 client.
If the Content-Length response header is set in the running Lua code, however, this buffering will be disabled even if the lua_http10_buffering directive is turned on.
For large streaming output responses, it is important to disable the lua_http10_buffering directive to minimise memory usage.
Note that common HTTP benchmark tools such as ab and http_load issue HTTP 1.0 requests by default.
To force curl to send HTTP 1.0 requests, use the -0 option.