Port Datasette to ASGI

[simonw]

Datasette doesn't take much advantage of Sanic, and I'm increasingly having to work around parts of it because of idiosyncrasies that are specific to Datasette - caring about the exact order of querystring arguments for example.

Since Datasette is GET-only our needs from a web framework are actually pretty slim.

This becomes more important as I expand the plugins #14 framework. Am I sure I want the plugin ecosystem to depend on a Sanic if I might move away from it in the future?

If Datasette wasn't all about async/await I would use WSGI, but today it makes more sense to use ASGI. I'd like to be confident that switching to ASGI would still give me the excellent performance that Sanic provides.

https://github.com/django/asgiref/blob/master/specs/asgi.rst

[simonw]

I think I can do this almost entirely within my existing BaseView class structure.

First, decouple the async data() methods by teaching them to take a querystring object as an argument instead of a Sanic request object. The get() method can then send that new object instead of a request.

Next teach the base class how to obey the ASGI protocol.

I should be able to get support for both Sanic and uvicorn/daphne working in the same codebase, which will make it easy to compare their performance.

[simonw]

Thinking about this further, maybe I should embrace ASGI turtles-all-the-way-down and teach each datasette view class to take a scope to the constructor and act entirely as an ASGI component. Would be a nice way of diving deep into ASGI and I can add utility helpers for things like querystring evaluation as I need them.

[simonw]

I built this ASGI debugging tool to help with this migration: https://asgi-scope.now.sh/fivethirtyeight-34d6604/most-common-name%2Fsurnames.json?foo=bar&bazoeuto=onetuh&a=.

[simonw]

Results of an extremely simple micro-benchmark comparing the two shows that uvicorn is at least as fast as Sanic (benchmarks a little faster with a very simple payload): https://gist.github.com/simonw/418950af178c01c416363cc057420851

[simonw]

https://github.com/encode/uvicorn/blob/572b5fe6c811b63298d5350a06b664839624c860/uvicorn/run.py#L63 is how you start a Uvicorn server from code as opposed to the uvicorn CLI

from uvicorn.run import UvicornServer
UvicornServer().run(app, host=host, port=port)

[simonw]

This looks VERY relevant: https://github.com/encode/starlette

[tomchristie]

I’m up for helping with this.

Looks like you’d need static files support, which I’m planning on adding a component for. Anything else obviously missing?

For a quick overview it looks very doable - the test client ought to me your test cases stay roughly the same.

Are you using any middleware or other components for the Sanic ecosystem? Do you use cookies or sessions at all?

[simonw]

No cookies or sessions - no POST requests in fact, Datasette just cares about GET (path and querystring) and being able to return custom HTTP headers.

[tomchristie]

Okay. I reckon the latest version should have all the kinds of components you'd need:

Recently added ASGI components for Routing and Static Files support, as well as making few tweaks to make sure requests and responses are instantiated efficiently.

Don't have any redirect-to-slash / redirect-to-non-slash stuff out of the box yet, which it looks like you might miss.

[simonw]

I'm now hacking around with an initial version of this in the starlette branch.

Here's my work in progress, deployed using datasette publish now fixtures.db -n datasette-starlette-demo --branch=starlette --extra-options="--asgi"

https://datasette-starlette-demo.now.sh/

Lots more work to do - the CSS isn't being served correctly for example, it's showing this error when I hit /-/static/app.css:

INFO: 127.0.0.1 - "GET /-/static/app.css HTTP/1.1" 200
ERROR: Exception in ASGI application
Traceback (most recent call last):
  File "/Users/simonw/Dropbox/Development/datasette/venv/lib/python3.6/site-packages/uvicorn/protocols/http/httptools_impl.py", line 363, in run_asgi
    result = await asgi(self.receive, self.send)
  File "/Users/simonw/Dropbox/Development/datasette/venv/lib/python3.6/site-packages/starlette/staticfiles.py", line 91, in __call__
    await response(receive, send)
  File "/Users/simonw/Dropbox/Development/datasette/venv/lib/python3.6/site-packages/starlette/response.py", line 180, in __call__
    {"type": "http.response.body", "body": chunk, "more_body": False}
  File "/Users/simonw/Dropbox/Development/datasette/venv/lib/python3.6/site-packages/uvicorn/protocols/http/httptools_impl.py", line 483, in send
    raise RuntimeError("Response content shorter than Content-Length")
RuntimeError: Response content shorter than Content-Length

[simonw]

It looks like that's a bug in Starlette - filed here: encode/starlette#32

[simonw]

Tom shipped my fix for that bug already, so https://datasette-starlette-demo.now.sh/ is now serving CSS!

[simonw]

Refs
encode/uvicorn#168

[tomchristie]

Some notes:

• Starlette just got a bump to 0.3.0 - there's some renamings in there. It's got enough functionality now that you can treat it either as a framework or as a toolkit. Either way the component design is all just here's an ASGI app all the way through.
• Uvicorn got a bump to 0.3.3 - Removed some cyclical references that were causing garbage collection to impact performance. Ought to be a decent speed bump.
• Wrt. passing config - Either use a single envvar that points to a config, or use multiple envvars for the config. Uvicorn could get a flag to read a .env file, but I don't see ASGI itself having a specific interface there.

[simonw]

Wow, this issue has been open for a full year now!

I've been thinking about this a lot. I've decided I want Datasette to use ASGI 3.0 internally with no dependencies on anything else - then I want the option to run Datasette under both daphne and uvicorn - because uvicorn doesn't support Python 3.5 but Datasette still needs to (primarily for Glitch), and daphne works with 3.5.

So I'm going to try to go the following route:

• Every Datasette view becomes an ASGI app
• The Datasette application itself is an ASGI app that routes to those views
• When you pip install datasette you get Daphne as a dependency (I'd like you to be able to opt-out of installing Daphne, I'm not yet sure how that would work)
• A new asgi_serve plugin hook allows a plugin to serve Datasette using uvicorn (or hypercorn) instead

[simonw]

I said earlier that I only need to support GET - I actually need to be able to support POST too, mainly to support plugins (e.g. a plugin that allows authenticated login before you can view Datasette, but potentially also plugins that let you write data directly to SQLite as well).

[simonw]

All of the tests are now passing!

I still need a solution for this:

datasette/datasette/app.py

Lines 706 to 714 in 5bd510b

	app = DatasetteRouter(routes)
	# First time server starts up, calculate table counts for immutable databases
	# TODO: re-enable this mechanism
	# @app.listener("before_server_start")
	# async def setup_db(app, loop):
	# for dbname, database in self.databases.items():
	# if not database.is_mutable:
	# await database.table_counts(limit=60 * 60 * 1000)

I think the answer is ASGI lifespan, which is supported by Uvicorn. https://asgi.readthedocs.io/en/latest/specs/lifespan.html#startup

[simonw]

I also need to actually take advantage of raw_path such that pages like https://fivethirtyeight.datasettes.com/fivethirtyeight/twitter-ratio%2Fsenators can be correctly served.

[simonw]

Tests are failing on Python 3.5: https://travis-ci.org/simonw/datasette/jobs/549380098 - error is TypeError: the JSON object must be str, not 'bytes'

[simonw]

And now the tests are all passing!

Still to do:

• Use raw_path so table names containing / can work correctly
• Get ?_trace=1 working again
• Replacement for @app.listener("before_server_start")
• Replace Sanic request object with my own request class, so I can remove Sanic dependency

[simonw]

I'm going to move the remaining work into a pull request.

[simonw]

It's alive! Here's the first deployed version: https://a559123.datasette.io/

You can confirm it's running under ASGI by viewing https://a559123.datasette.io/-/versions and looking for the "asgi" key.

Compare to the last version of master running on Sanic here: http://aa91112.datasette.io/

[simonw]

I wrote about this on my blog: https://simonwillison.net/2019/Jun/23/datasette-asgi/