issues

Your code is full of open files that are never closed, especially when you deal with reading/writing json/yaml files.

If you run python with warnings enabled this problem becomes evident. This probably contributes to some memory leaks in long running datasettes if the GC will not 'collect' those resources properly.

This is easily fixed by using a context manager instead of just using open: python with open('some_file', 'w') as opened_file: opened_file.write('string')

In some newer parts of the code you use Path objects 'read_text' and 'write_text' functions which close the file properly and are prefered in some cases.

If you want I can create a PR for all places i found this pattern in.

Bellow is a fraction of places where i found a ResourceWarning: ```python

update-docs-help.py: 20 actual = actual.replace("Usage: cli ", "Usage: datasette ") 21: open(docs_path / filename, "w").write(actual) 22

datasette\app.py: 210 ): 211: inspect_data = json.load((config_dir / "inspect-data.json").open()) 212 if immutables is None:

266 if config_dir and (config_dir / "settings.json").exists() and not config: 267: config = json.load((config_dir / "settings.json").open()) 268 self._settings = dict(DEFAULT_SETTINGS, **(config or {}))

445 self._app_css_hash = hashlib.sha1( 446: open(os.path.join(str(app_root), "datasette/static/app.css")) 447 .read()

datasette\cli.py: 130 else: 131: out = open(inspect_file, "w") 132 loop = asyncio.get_event_loop()

459 if inspect_file: 460: inspect_data = json.load(open(inspect_file)) 461

```

Working on my PR i noticed that tests are very slow.

The plain pytest run took about 37 minutes for me. However i could shave of about 10 minutes from that if i used pytest-xdist to parallelize execution. pytest -n 8 is run only in 28 minutes on my machine.

I can create a PR to mention that in your documentation. This will be a simple change to add pytest-xdist to requirements and change a command to run pytest in documentation.

Does that make sense to you?

After a bit more investigation it looks like python-xdist is not an answer. It creates a race condition for tests that try to clead temp dir before run.

Profiling shows that most time is spent on conn.executescript(TABLES) in make_app_client function. Which makes sense.

Perhaps the better approach would be look at the app_client fixture which is already session scoped, but not used by all test cases. And/or use conn = sqlite3.connect(":memory:") which is much faster. And/or truncate tables after each TC instead of deleting the file and re-creating them.

I can take a look which is the best approach if you give the go-ahead.