When you read from the input filter chain as $r->read does, the http input filter automatically handles the protocol and passes the dechunked data up to the caller. It does not spool the stream at all.
You'd have to look at how mod perl implements read to see if it loops its ap_get_brigade calls on the input filter chain to fill the passed buffer to the desired length or not. But under no circumstances should you have to deal with chunked data directly.
Sent from my iPhone
This is the Transfer-Encoding: chunked I was writing about:
On Wed, Jul 3, 2013 at 11:34 AM, Jim Schueler <email@example.com>
I played around with chunking recently in the context of media streaming: The client is only requesting a "chunk" of data. "Chunking" is how media players perform a "seek". It was originally implemented for FTP transfers: E.g, to transfer a large file in (say 10K) chunks. In the case that you describe below, if no Content-Length is specified, that indicates "send the remainder".
From what I know, a "chunk" request header is used this way to specify the server response. It does not reflect anything about the data included in the body of the request. So first, I would ask if you're confused about this request information.
Hypothetically, some browsers might try to upload large files in small chunks and the "chunk" header might reflect a push transfer. I don't know if "chunk" is ever used for this purpose. But it would require the following characteristics:
1. The browser would need to originally inquire if the server is
capable of this type of request.
2. Each chunk of data will arrive in a separate and independent HTTP
request. Not necessarily in the order they were sent.
3. Two or more requests may be handled by separate processes
simultaneously that can't be written into a single destination.
4. Somehow the server needs to request a resend if a chunk is missing.
Solving this problem requires an imaginitive use of HTTP.
Sounds messy. But might be appropriate for 100M+ sized uploads. This *may* reflect your situation. Can you please confirm?
For a single process, the incoming content-length is unnecessary. Buffered I/O automatically knows when transmission is complete. The read() argument is the buffer size, not the content length. Whether you spool the buffer to disk or simply enlarge the buffer should be determined by your hardware capabilities. This is standard IO behavior that has nothing to do with HTTP chunk. Without a "Content-Length" header, after looping your read() operation, determine the length of the aggregate data and pass that to Catalyst.
But if you're confident that the complete request spans several smaller (chunked) HTTP requests, you'll need to address all the problems I've described above, plus the problem of re-assembling the whole thing for Catalyst. I don't know anything about Plack, maybe it can perform all this required magic.
Otherwise, if the whole purpose of the Plack temporary file is to pass a file handle, you can pass a buffer as a file handle. Used to be IO::String, but now that functionality is built into the core.
By your last paragraph, I'm really lost. Since you're already passing the request as a file handle, I'm guessing that Catalyst creates the tempororary file for the *response* body. Can you please clarify? Also, what do you mean by "de-chunking"? Is that the same think as re-assembling?
Wish I could give a better answer. Let me know if this helps.
On Tue, 2 Jul 2013, Bill Moseley wrote:
For requests that are chunked (Transfer-Encoding: chunked and no
Content-Length header) calling $r->read returns unchunked data from the
That's indeed handy. Is that mod_perl doing that un-chunking or is it
But, it leads to some questions.
First, if $r->read reads unchunked data then why is there a
Transfer-Encoding header saying that the content is chunked? Shouldn't
that header be removed? How does one know if the content is chunked or
Second, if there's no Content-Length header then how does one know how much
data to read using $r->read?
One answer is until $r->read returns zero bytes, of course. But, is
that guaranteed to always be the case, even for, say, pipelined requests?
My guess is yes because whatever is de-chunking the request knows to stop
after reading the last chunk, trailer and empty line. Can anyone elaborate
on how Apache/mod_perl is doing this?
Perhaps I'm approaching this incorrectly, but this is all a bit untidy.
I'm using Catalyst and Catalyst needs a Content-Length. So, I have a Plack
Middleware component that creates a temporary file writing the buffer from
$r->read( my $buffer, 64 * 1024 ) until that returns zero bytes. I pass
this file handle onto Catalyst.
Then, for some content-types, Catalyst (via HTTP::Body) writes the body to
another temp file. I don't know how Apache/mod_perl does its de-chunking,
but I can call $r->read with a huge buffer length and Apache returns that.
So, maybe Apache is buffering to disk, too.
In other words, for each tiny chunked JSON POST or PUT I'm creating two (or
three?) temp files which doesn't seem ideal.