#33699: Read ASGI request body from asyncio queue on-demand
-------------------------------+------------------------------------
Reporter: Noxx | Owner: nobody
Type: New feature | Status: new
Component: HTTP handling | Version: dev
Severity: Normal | Resolution:
Keywords: ASGI, async | Triage Stage: Accepted
Has patch: 1 | Needs documentation: 0
Needs tests: 0 | Patch needs improvement: 0
Easy pickings: 0 | UI/UX: 0
-------------------------------+------------------------------------
Changes (by Carlton Gibson):
* status: closed => new
* cc: Andrew Godwin, Carlton Gibson (added)
* type: Uncategorized => New feature
* component: File uploads/storage => HTTP handling
* version: 4.0 => dev
* keywords: => ASGI, async
* has_patch: 0 => 1
* resolution: invalid =>
* stage: Unreviewed => Accepted
Old description:
> == Story
> Our application supports many file uploads.
> In normal circumstances, those files are only a few megabytes to of
> around 100MB.
> However, some uploads can be several GB in size (10-15GB and up).
>
> I know, that file uploads with such a size can be handled differently,
> e.g. using a seperate endpoint, allowing slow requests to be served
> seperately from the application etc.
> and only to reference those files in a final request.
> This solution would introduce some other problems regarding house-
> keeping, non-atomic requests etc, thus I want to ignore this workaround
> for now.
>
> == Performance impact:
> The impact can be best observed when uploading a file which is bigger in
> it's size, e.g. 1GB+.
> On my maschine it takes the client around 700ms to send the request to
> the application, but than waits around 1.5s for the final response.
> Of course, those numbers are dramatically influenced additionally in
> production by storage speed, file size, webserver (uvicorn/daphne/..),
> load-balancers etc.
> But the final take here is, that the server does some additional work
> after the client finished its request.
> In a production-like environment the numbers peaks to 4s (send request)
> and 16s (waiting for response) for the same file size. Uploading a 3GB
> file, the numbers are 11s and 44s.
> As you can see, the 44s are very near the default gateway timeout of
> Nginx and the hard-coded one of AWS load-balancers.
> As soon as the server needs more than 60s to create the response, the
> client will get a gateway timeout error.
>
> == Workflow of file uploads:
>
> I'm not a Django-dev, so please correct me if I'm wrong. As far as I
> understand, the uploaded file is processed as the following:
>
> Given, that the application is running within an ASGI server, the whole
> request is received by the ASGIHandler.
> It's method #read_body creates a SpooledTemporaryFile.
> This temporary file contains the whole request body, including POST
> parameters and files.
>
> As soon as the request has been received (this is the point the client
> finished uploading the file) it is than transformed into an ASGIRequest.
> The ASGIRequest has #_load_post_and_files and #parse_file_upload as
> methods which parses the body into seperate files.
> This is done by reading the body (the temporary file) and iterating over
> them seperated by the POST seperator done by MultiPartParser.
> The generated chunks are than sent to upload handlers which process those
> files further on using #receive_data_chunk.
> The default upload handlers provided by Django will write those files to
> disc again, depending on their size.
>
> == Problem
> The problem here is, that the uploaded file(s) are transformed and
> written as well as read multiple times.
> First the whole body is written into a SpooledTemporaryFile which is re-
> read using streams (LazyStream) just to be written once more by an upload
> handler.
>
> The impact is low if the uploaded file is small, but increases
> dramatically if the size is increased, the file hits the disc and/or the
> storage is slow.
>
> == Optimization / Brainstorming
> Would it be possible to reduce the workflow to a single write call?
> E.g. if the ASGIHandler already splits the request body into seperate
> files, it would be possible to just forward the file pointers until the
> upload handlers needs to be called.
> Those handlers would be able to either use those files as-is or to re-
> read them if pre-processing is needed.
>
> In a best-case scenario, an user uploads a file whichis created as a
> temporary file in parallel.
> As soon as the request has been finished, the file is than moved to its
> final location (as already implemented by upload handlers by providing
> #temporary_file_path)
> The server would not need any time processing the request further on and
> would be able to sent the response within some milliseconds independent
> of the file size.
> The roundtrip time would be reduced by 2/3 and also the gateway timeout
> would be fixed.
>
> == Environment
> We're using Django 4.0.4 with Gunicorn 20.1.0 and Uvicorn 0.17.6.
>
> == Attachments
> I've attached two flame graphs of a file upload which hopfully
> illustrates this issue.
> One is using the internal runserver (wsgi) and one of our (stripped)
> application using gunicorn+uvicorn (asgi)
New description:
Was "Performance does not scale when increasing file upload sizes"
== Story
Our application supports many file uploads.
In normal circumstances, those files are only a few megabytes to of around
100MB.
However, some uploads can be several GB in size (10-15GB and up).
I know, that file uploads with such a size can be handled differently,
e.g. using a seperate endpoint, allowing slow requests to be served
seperately from the application etc.
and only to reference those files in a final request.
This solution would introduce some other problems regarding house-keeping,
non-atomic requests etc, thus I want to ignore this workaround for now.
== Performance impact:
The impact can be best observed when uploading a file which is bigger in
it's size, e.g. 1GB+.
On my maschine it takes the client around 700ms to send the request to the
application, but than waits around 1.5s for the final response.
Of course, those numbers are dramatically influenced additionally in
production by storage speed, file size, webserver (uvicorn/daphne/..),
load-balancers etc.
But the final take here is, that the server does some additional work
after the client finished its request.
In a production-like environment the numbers peaks to 4s (send request)
and 16s (waiting for response) for the same file size. Uploading a 3GB
file, the numbers are 11s and 44s.
As you can see, the 44s are very near the default gateway timeout of Nginx
and the hard-coded one of AWS load-balancers.
As soon as the server needs more than 60s to create the response, the
client will get a gateway timeout error.
== Workflow of file uploads:
I'm not a Django-dev, so please correct me if I'm wrong. As far as I
understand, the uploaded file is processed as the following:
Given, that the application is running within an ASGI server, the whole
request is received by the ASGIHandler.
It's method #read_body creates a SpooledTemporaryFile.
This temporary file contains the whole request body, including POST
parameters and files.
As soon as the request has been received (this is the point the client
finished uploading the file) it is than transformed into an ASGIRequest.
The ASGIRequest has #_load_post_and_files and #parse_file_upload as
methods which parses the body into seperate files.
This is done by reading the body (the temporary file) and iterating over
them seperated by the POST seperator done by MultiPartParser.
The generated chunks are than sent to upload handlers which process those
files further on using #receive_data_chunk.
The default upload handlers provided by Django will write those files to
disc again, depending on their size.
== Problem
The problem here is, that the uploaded file(s) are transformed and written
as well as read multiple times.
First the whole body is written into a SpooledTemporaryFile which is re-
read using streams (LazyStream) just to be written once more by an upload
handler.
The impact is low if the uploaded file is small, but increases
dramatically if the size is increased, the file hits the disc and/or the
storage is slow.
== Optimization / Brainstorming
Would it be possible to reduce the workflow to a single write call?
E.g. if the ASGIHandler already splits the request body into seperate
files, it would be possible to just forward the file pointers until the
upload handlers needs to be called.
Those handlers would be able to either use those files as-is or to re-read
them if pre-processing is needed.
In a best-case scenario, an user uploads a file whichis created as a
temporary file in parallel.
As soon as the request has been finished, the file is than moved to its
final location (as already implemented by upload handlers by providing
#temporary_file_path)
The server would not need any time processing the request further on and
would be able to sent the response within some milliseconds independent of
the file size.
The roundtrip time would be reduced by 2/3 and also the gateway timeout
would be fixed.
== Environment
We're using Django 4.0.4 with Gunicorn 20.1.0 and Uvicorn 0.17.6.
== Attachments
I've attached two flame graphs of a file upload which hopfully illustrates
this issue.
One is using the internal runserver (wsgi) and one of our (stripped)
application using gunicorn+uvicorn (asgi)
--
Comment:
Thanks Noxx, re-opening, and accepting for review after discussion on
mailing list.
* https://groups.google.com/g/django-developers/c/fu6ZSmu-YJE
* https://github.com/django/django/pull/15704
The basic idea here is to read the request body as-needed, rather than
pulling it into the spooled temporary file before instantiating the ASGI
request.
Perhaps there will be gotchas, but that feels like the right thing to do,
so... 👍
--
Ticket URL: <https://code.djangoproject.com/ticket/33699#comment:2>
Django <https://code.djangoproject.com/>
The Web framework for perfectionists with deadlines.
--
You received this message because you are subscribed to the Google Groups
"Django updates" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
To view this discussion on the web visit
https://groups.google.com/d/msgid/django-updates/01070180d6300db3-51298621-1a75-4198-bc00-fd8db7a1a8e1-000000%40eu-central-1.amazonses.com.
