issue_comments

I like the .add_memory_database() option. I also like that it makes it more obvious that this is a capability of Datasette, since I'm excited to see more plugins, features and tests that take advantage of it.

A couple of options:

datasette.add_memory_database("test_json_array")
# or make that first argument to add_database() optional and support:
datasette.add_database(memory_name="test_json_array")

I'm going to try a new pattern for testing this, enabled by #1151 - the test will create a new named in-memory database, write some records to it and then run some test facets against that. This will save me from having to add yet another fixtures table for this.

It's out! https://github.com/encode/httpx/releases/tag/0.17.0

Microsoft's playwright Python library solves this problem by code generating both their sync AND their async libraries https://github.com/microsoft/playwright-python/tree/master/scripts

Hah or use this trick, which genuinely rewrites the code at runtime using a class decorator! https://github.com/python-happybase/aiohappybase/blob/0990ef45cfdb720dc987afdb4957a0fac591cb99/aiohappybase/sync/_util.py#L19-L32

Some other interesting background reading: https://docs.sqlalchemy.org/en/14/orm/extensions/asyncio.html - in particular see how SQLALchemy has a await conn.run_sync(meta.drop_all) mechanism for running methods that haven't themselves been provided in an async version

From the poster here: https://github.com/sethmlarson/pycon-async-sync-poster/blob/master/poster.pdf

https://user-images.githubusercontent.com/9599/109401634-9f0a1400-7904-11eb-8b3a-37df0678b8dc.png">

unasync is an implementation of the exact pattern I was talking about above - it uses the tokenize module from the Python standard library to apply some clever rules to transform an async codebase into a sync one. https://unasync.readthedocs.io/en/latest/ - implementation here: https://github.com/python-trio/unasync/blob/v0.5.0/src/unasync/__init__.py

"how to use some new tools to more easily maintain a codebase that supports both async and synchronous I/O and multiple async libraries" - yeah that's exactly what I need, thank you!

Here's a really wild idea: I wonder if it would be possible to run a source transformation against either the sync or the async versions of the code to produce the equivalent for the other paradigm?

Could that even be as simple as a set of regular expressions against the await ... version that strips out or replaces the await and async def and async for statements?

If so... I could maintain just the async version, generate the sync version with a script and rely on robust unit testing to guarantee that this actually works.

I have a hunch this is actually going to be quite difficult, due to the internal complexity of some of the sqlite-utils API methods.

Consider db[table].extract(...) for example. It does a whole bunch of extra queries inside the method - each of those would need to be turned into an await call for the async version. Here's the method body today:

https://github.com/simonw/sqlite-utils/blob/09c3386f55f766b135b6a1c00295646c4ae29bec/sqlite_utils/db.py#L1060-L1152

Writing this method twice - looking similar but with await ... tucked in before every internal method it calls that needs to execute SQL - is going to be pretty messy.

One thing that would help a LOT is figuring out how to share the majority of the test code. If the exact same tests could run against both the sync and async versions with a bit of test trickery, maintaining parallel implementations would at least be a bit more feasible.

Another option here would be to add https://github.com/omnilib/aiosqlite/blob/main/aiosqlite/core.py as a dependency - it's four years old now and actively marinated, and the code is pretty small so it looks like a solid, stable, reliable dependency.

Datasette has its own implementation of a write queue for exactly this purpose - and there's no reason at all that should stay in Datasette rather than being extracted out and moved over here to sqlite-utils.

One small concern I have is around the API design. I'd want to keep supporting the existing synchronous API while also providing a similar API with await-based methods.

What are some good examples of libraries that do this? I like how https://www.python-httpx.org/ handles it, maybe that's a good example to imitate?

Thanks!

I requested my Gmail export from takeout - once that arrives I'll test it against this and then merge the PR.

Here's the test I wrote:

git diff tests/test_custom_pages.py
diff --git a/tests/test_custom_pages.py b/tests/test_custom_pages.py
index 6a23192..5a71f56 100644
--- a/tests/test_custom_pages.py
+++ b/tests/test_custom_pages.py
@@ -2,11 +2,19 @@ import pathlib
 import pytest
 from .fixtures import make_app_client

+TEST_TEMPLATE_DIRS = str(pathlib.Path(__file__).parent / "test_templates")
+

 @pytest.fixture(scope="session")
 def custom_pages_client():
+    with make_app_client(template_dir=TEST_TEMPLATE_DIRS) as client:
+        yield client
+
+
+@pytest.fixture(scope="session")
+def custom_pages_client_with_base_url():
     with make_app_client(
-        template_dir=str(pathlib.Path(__file__).parent / "test_templates")
+        template_dir=TEST_TEMPLATE_DIRS, config={"base_url": "/prefix/"}
     ) as client:
         yield client

@@ -23,6 +31,12 @@ def test_request_is_available(custom_pages_client):
     assert "path:/request" == response.text


+def test_custom_pages_with_base_url(custom_pages_client_with_base_url):
+    response = custom_pages_client_with_base_url.get("/prefix/request")
+    assert 200 == response.status
+    assert "path:/prefix/request" == response.text
+
+
 def test_custom_pages_nested(custom_pages_client):
     response = custom_pages_client.get("/nested/nest")
     assert 200 == response.status

I added a debug line just before for regex, wildcard_template here:

https://github.com/simonw/datasette/blob/afed51b1e36cf275c39e71c7cb262d6c5bdbaa31/datasette/app.py#L1148-L1155

And it showed that for some reason request.path is /prefix/prefix/request here - the prefix got doubled somehow.

Sounds like a bug - thanks for reporting this.

Could I take this part:

             suggested_facet_sql = """ 
                 select distinct json_type({column}) 
                 from ({sql}) 
             """.format( 
                 column=escape_sqlite(column), sql=self.sql 
             ) 

And add where {column} is not null and {column} != '' perhaps?

Relevant code: https://github.com/simonw/datasette/blob/afed51b1e36cf275c39e71c7cb262d6c5bdbaa31/datasette/facets.py#L271-L295

Could this handle lists of objects too? That would be pretty amazing - if the column has a [{...}, {...}] list in it could turn that into a many-to-many.

Linking to rows from custom queries is a lot harder - because given an arbitrary string of SQL it's difficult to analyze it and figure out which (if any) of the returned columns represent a primary key.

It's possible to manually write a SQL query that returns a column that will be treated as a link to another page using this plugin, but it's not particularly straight-forward: https://datasette.io/plugins/datasette-json-html

Agreed, this would be extremely useful. I'd love to be able to facet against custom queries. It's a fair bit of work to implement but it's not impossible. Closing this as a duplicate of #972.

If there's no primary key in the JSON could use the hash_id mechanism.

This came up in office hours!

New paragraph at the top of https://docs.datasette.io/en/latest/writing_plugins.html

Want to start by looking at an example? The Datasette plugins directory lists more than 50 open source plugins with code you can explore. The plugin hooks page includes links to example plugins for each of the documented hooks.

sqlite-utils create-view currently has a --ignore option, so adding that to sqlite-utils drop-view and sqlite-utils drop-table makes sense as well.

So:

    db["my_table"].drop(ignore=True)

Other methods (db.create_view() for example) have ignore=True to mean "don't throw an error if this causes a problem", so I'm good with adding that to .drop_view().

I don't like using it as the default partly because that would be a very minor breaking API change, but mainly because I don't want to hide mistakes people make - e.g. if you mistype the name of the table you are trying to drop.

Example from the docs:

>>> db = sqlite_utils.Database(memory=True)
>>> db["dogs"].insert({"name": "Cleo"})
>>> for pk, row in db["dogs"].pks_and_rows_where():
...     print(pk, row)
1 {'rowid': 1, 'name': 'Cleo'}

>>> db["dogs_with_pk"].insert({"id": 5, "name": "Cleo"}, pk="id")
>>> for pk, row in db["dogs_with_pk"].pks_and_rows_where():
...     print(pk, row)
5 {'id': 5, 'name': 'Cleo'}

>>> db["dogs_with_compound_pk"].insert(
...     {"species": "dog", "id": 3, "name": "Cleo"},
...     pk=("species", "id")
... )
>>> for pk, row in db["dogs_with_compound_pk"].pks_and_rows_where():
...     print(pk, row)
('dog', 3) {'species': 'dog', 'id': 3, 'name': 'Cleo'}

Documentation: https://sqlite-utils.datasette.io/en/latest/python-api.html#listing-rows-with-their-primary-keys

WIP in a pull request.

I prototyped this and I like it:

In [1]: import sqlite_utils
In [2]: db = sqlite_utils.Database("/Users/simon/Dropbox/Development/datasette/fixtures.db")
In [3]: list(db["compound_primary_key"].pks_and_rows_where())
Out[3]: [(('a', 'b'), {'pk1': 'a', 'pk2': 'b', 'content': 'c'})]

table.pks_and_rows_where(...) is explicit and I think less ambiguous than the other options.

I quite like pks_with_rows_where(...) - but grammatically it suggests it will return the primary keys that exist where their rows match the criteria - "pks with rows" can be interpreted as "pks for the rows that..." as opposed to "pks accompanied by rows"

Is pk_rows_where() a good name? It sounds like it returns "primary key rows" which isn't a thing. It actually returns rows along with their primary key.

Other options:

• table.rows_with_pk_where(...) - should this return (row, pk) rather than (pk, row)?
• table.rows_where_pk(...)
• table.pk_and_rows_where(...)
• table.pk_with_rows_where(...)
• table.pks_with_rows_where(...) - because rows is pluralized, so pks should be pluralized too?
• table.pks_rows_where(...)

One option: .rows_where() could grow a ensure_pk=True option which checks to see if the table is a rowid table and, if it is, includes that in the select.

Or... how about you can call .rows_where(..., pks=True) and it will yield (pk, rowdict) tuple pairs instead of just returning the sequence of dictionaries?

I'm always a little bit nervous of methods that vary their return type based on their arguments. Maybe this would be a separate method instead?

    for pk, row in table.pk_rows_where(...):
        # ...

Here's the current implementation of .extract(): https://github.com/simonw/sqlite-utils/blob/806c21044ac8d31da35f4c90600e98115aade7c6/sqlite_utils/db.py#L1049-L1074

Tricky detail here: I create the lookup table first, based on the types of the columns that are being extracted.

I need to do this because extraction currently uses unique tuples of values, so the table has to be created in advance.

But if I'm using these new expand functions to figure out what's going to be extracted, I don't know the names of the columns and their types in advance. I'm only going to find those out during the transformation.

This may turn out to be incompatible with how .extract() works at the moment. I may need a new method, .extract_expand() perhaps? It could be simpler - work only against a single column for example.

I can still use the existing sqlite-utils extract CLI command though, with a --json flag and a rule that you can't run it against multiple columns.

Problem with calling this argument transform= is that the term "transform" already means something else in this library.

I could use convert= instead.

... but that doesn't instantly make me think of turning a value into multiple columns.

How about expand=? I've not used that term anywhere yet.

db["Reports"].extract(["Reported by"], expand={"Reported by": json.loads})

I think that works. You're expanding a single value into several columns of information.

I'm going to go with last-wins - so if multiple transform functions return the same key the last one will over-write the others.

Maybe the Python version takes an optional dictionary mapping column names to transformation functions? It could then merge all of those results together - and maybe throw an error if the same key is produced by more than one column.

    db["Reports"].extract(["Reported by"], transform={"Reported by": json.loads})

Or it could have an option for different strategies if keys collide: first wins, last wins, throw exception, add a prefix to the new column name. That feels a bit too complex for an edge-case though.

It would be OK if the CLI version only allows you to specify a single column if you are using the --json option.

For the Python version I'd like to be able to provide a transformation callback function - which can be json.loads but could also be anything else which accepts the value of the current column and returns a Python dictionary of columns and their values to use in the new table.

Likewise the sqlite-utils extract command takes one or more columns:

Usage: sqlite-utils extract [OPTIONS] PATH TABLE COLUMNS...

  Extract one or more columns into a separate table

Options:
  --table TEXT             Name of the other table to extract columns to
  --fk-column TEXT         Name of the foreign key column to add to the table
  --rename <TEXT TEXT>...  Rename this column in extracted table

The Python .extract() method currently starts like this:

def extract(self, columns, table=None, fk_column=None, rename=None):
        rename = rename or {}
        if isinstance(columns, str):
            columns = [columns]
        if not set(columns).issubset(self.columns_dict.keys()):
            raise InvalidColumns(
                "Invalid columns {} for table with columns {}".format(
                    columns, list(self.columns_dict.keys())
                )
            )
        ...

Note that it takes a list of columns (and treats a string as a single item list). That's because it can be called with a list of columns and it will use them to populate another table of unique tuples of those column values.

So a new mechanism that can instead read JSON values from a single column needs to be compatible with that existing design.

I bet the bug is in here: https://github.com/simonw/sqlite-utils/blob/806c21044ac8d31da35f4c90600e98115aade7c6/sqlite_utils/db.py#L593-L602

I really like the way the Share feature on Stack Overflow works: https://stackoverflow.com/questions/18934149/how-can-i-use-postgresqls-text-column-type-in-django

A "Share link" button would only be needed on the table page and the arbitrary query page I think - and maybe on the row page, especially as that page starts to grow more features in the future.

This can definitely be done with a plugin.

Adding to Datasette itself is an interesting idea. I think it's possible that many users these days no longer assume they can paste a URL from the browser address bar (if they ever understood that at all) because to many apps are SPAs with broken URLs.

The shareable URLs are actually a key feature of Datasette - so maybe they should be highlighted in the default UI?

I built a "copy to clipboard" feature for datasette-copyable and wrote up how that works here: https://til.simonwillison.net/javascript/copy-button

Bug is here: https://github.com/simonw/datasette/blob/42caabf7e9e6e4d69ef6dd7de16f2cd96bc79d5b/datasette/filters.py#L149-L165

Those json_each lines should be:

select {t}.rowid from {t}, json_each([{t}].[{c}]) j

This is another change which is a little bit hard to figure out because I haven't solved #878 yet.

I'm not going to work on this for a while, but if anyone has needs or ideas around that they can add them to this issue.

It's good on mobile - iOS at least. Going to close this open new issues if anyone reports bugs.

Another benefit to default:object - I could include a key that shows a list of available extras. I could then use that to power an interactive API explorer.

OK, I'm back to the "top level object as the default" side of things now - it's pretty much unanimous at this point, and it's certainly true that it's not a decision you'll even regret.

Yet another option: support a ?_path=x option which returns a nested path from the result. So you could do this:

/github/commits.json?_path=rows - to get back a top-level array pulled from the "rows" key.

Have you given any thought as to whether to pretty print (format with spaces) the output or not? Can be useful for debugging/exploring in a browser or other basic tools which don’t parse the JSON. Could be default (can’t be much bigger with gzip?) or opt-in.

Adding a ?_pretty=1 option that does that is a great idea, I'm filing a ticket for it: #1237

... though thinking about this further, I could re-implement the select * from commits (but only return a max of 10 results) feature using a nested select * from (select * from commits) limit 10 query.

Minor suggestion: rename size query param to limit, to better reflect that it’s a maximum number of rows returned rather than a guarantee of getting that number, and also for consistency with the SQL keyword?

The problem there is that ?_size=x isn't actually doing the same thing as the SQL limit keyword. Consider this query:

https://latest-with-plugins.datasette.io/github?sql=select+*+from+commits - select * from commits

Datasette returns 1,000 results, and shows a "Custom SQL query returning more than 1,000 rows" message at the top. That's the size kicking in - I only fetch the first 1,000 results from the cursor to avoid exhausting resources. In the JSON version of that at https://latest-with-plugins.datasette.io/github.json?sql=select+*+from+commits there's a "truncated": true key to let you know what happened.

I find myself using ?_size=2 against Datasette occasionally if I know the rows being returned are really big and I don't want to load 10+MB of HTML.

This is only really a concern for arbitrary SQL queries though - for table pages such as https://latest-with-plugins.datasette.io/github/commits?_size=10 adding ?_size=10 actually puts a limit 10 on the underlying SQL query.

(I started a thread on Twitter about this: https://twitter.com/simonw/status/1363220355318358016)

Another option: .json always returns an object with a list of keys that gets increased through adding ?_extra= parameters.

.jsona always returns a JSON array of objects

I had something similar to this in Datasette a few years ago - a .jsono extension, which still redirects to the shape=array version.

I just noticed that https://latest-with-plugins.datasette.io/github/commits.json-preview?_extra=total&_size=0&_trace=1 executes 35 SQL queries at the moment! A great reminder that a big improvement from this change will be a reduction in queries through not calculating things like suggested facets unless they are explicitly requested.

The only advantage of headers is that you don’t need to do .rows, but that’s actually good as a data validation step anyway—if .rows is missing assume there’s an error and do your error handling path instead of parsing the rest.

This is something I've not thought very hard about. If there's an error, I need to return a top-level object, not a top-level array, so I can provide details of the error.

But this means that client code will have to handle this difference - it will have to know that the returned data can be array-shaped if nothing went wrong, and object-shaped if there's an error.

The HTTP status code helps here - calling client code can know that a 200 status code means there will be an array, but an error status code means an object.

If developers really hate that the shape could be different, they can always use ?_extra=next to ensure that the top level item is an object whether or not an error occurred. So I think this is OK.

A really exciting opportunity this opens up is for parallel execution - the facets() and suggested_facets() and total() async functions could be called in parallel, which could speed things up if I'm confident the SQLite thread pool can execute on multiple CPU cores (it should be able to because the Python sqlite3 module releases the GIL while it's executing C code).

This design should be influenced by how it's implemented.

One implementation that could be nice is that each of the keys that can be requested - next_url, total etc - maps to an async def function which can do the work. So that expensive count(*) will only be executed by the async def total function if it is requested.

This raises more questions: Both next and next_url work off the same underlying data, so if they are both requested can we re-use the work that next does somehow? Maybe by letting these functions depend on each other (so next_url() knows to first call next(), but only if it hasn't been called already.

I think I need to flesh out the full default collection of ?_extra= parameters in order to design how they will work under the hood.

I think I want to support ?_extra=next_url,total in addition to ?_extra=next_url&_extra=total - partly because it's less characters to type, and also because I know there exist URL handling library that don't know how to handle the same parameter multiple times (though they're going to break against Datasette already, so it's not a big deal).

Am I saying you won't get back a key in the response unless you explicitly request it, either by name or by specifying a bundle of extras (e.g. all or paginated)?

The "truncated": true key that tells you that your arbitrary query returned more than X results but was truncated is pretty important, do I really want people to have to opt-in to that one?

Also: having bundles like all or paginated live in the same namespace as single keys like next_url or total is a little odd - you can't tell by looking at them if they'll add a key called all or if they'll add a bunch of other stuff.

Maybe bundles could be prefixed with something, perhaps an underscore? ?_extra=_all and ?_extra=_paginated for example.

Demos:

• https://latest-with-plugins.datasette.io/github/commits.json-preview
• https://latest-with-plugins.datasette.io/github/commits.json-preview?_extra=next_url
• https://latest-with-plugins.datasette.io/github/commits.json-preview?_extra=total
• https://latest-with-plugins.datasette.io/github/commits.json-preview?_extra=next_url&_extra=total
• https://latest-with-plugins.datasette.io/github/commits.json-preview?_extra=total&_size=0

Having shortcuts means I could support ?_extra=all for returning ALL possible keys.

Also how would you opt out of returning the "rows" key? I sometimes want to do this - if I want to get back just the count or just the facets for example.

Some options:

• /fixtures/roadside_attractions.json?_extra=total&_extra=-rows
• /fixtures/roadside_attractions.json?_extra=total&_skip=rows
• /fixtures/roadside_attractions.json?_extra=total&_size=0

I quite like that last one with ?_size=0. I think it would still return "rows": [] but that's OK.

I'm leaning back in the direction of a flat JSON array of objects as the default - this:

/fixtures/roadside_attractions.json

Would return:

[
    {
      "pk": 1,
      "name": "The Mystery Spot",
      "address": "465 Mystery Spot Road, Santa Cruz, CA 95065",
      "latitude": 37.0167,
      "longitude": -122.0024
    },
    {
      "pk": 2,
      "name": "Winchester Mystery House",
      "address": "525 South Winchester Boulevard, San Jose, CA 95128",
      "latitude": 37.3184,
      "longitude": -121.9511
    },
    {
      "pk": 3,
      "name": "Burlingame Museum of PEZ Memorabilia",
      "address": "214 California Drive, Burlingame, CA 94010",
      "latitude": 37.5793,
      "longitude": -122.3442
    },
    {
      "pk": 4,
      "name": "Bigfoot Discovery Museum",
      "address": "5497 Highway 9, Felton, CA 95018",
      "latitude": 37.0414,
      "longitude": -122.0725
    }
]

To get the version that includes pagination information you would use the ?_extra= parameter. For example:

/fixtures/roadside_attractions.json?_extra=total&_extra=next_url

{
  "rows": [
    {
      "pk": 1,
      "name": "The Mystery Spot",
      "address": "465 Mystery Spot Road, Santa Cruz, CA 95065",
      "latitude": 37.0167,
      "longitude": -122.0024
    },
    {
      "pk": 2,
      "name": "Winchester Mystery House",
      "address": "525 South Winchester Boulevard, San Jose, CA 95128",
      "latitude": 37.3184,
      "longitude": -121.9511
    },
    {
      "pk": 3,
      "name": "Burlingame Museum of PEZ Memorabilia",
      "address": "214 California Drive, Burlingame, CA 94010",
      "latitude": 37.5793,
      "longitude": -122.3442
    },
    {
      "pk": 4,
      "name": "Bigfoot Discovery Museum",
      "address": "5497 Highway 9, Felton, CA 95018",
      "latitude": 37.0414,
      "longitude": -122.0725
    }
  ],
  "total": 4,
  "next_url": null
}

ANY usage of the ?_extra= parameter would turn the list into an object with a "rows" key.

Opting in to the total is nice because it's actually expensive to run a count, so only doing a count if the user requests it feels good.

But... having to add ?_extra=total&_extra=next_url for the common case of wanting both the total count and the URL to get the next page of results is a bit verbose. So maybe support aliases, like ?_extra=paginated which is a shortcut for ?_extra=total&_extra=next_url?

Need to test this on mobile.

Now live on https://latest.datasette.io/_memory

Encoded using https://meyerweb.com/eric/tools/dencoder/

%3Csvg%20aria-labelledby%3D%22cm-drag-to-resize%22%20role%3D%22img%22%20fill%3D%22%23ccc%22%20stroke%3D%22%23ccc%22%20xmlns%3D%22http%3A%2F%2Fwww.w3.org%2F2000%2Fsvg%22%20viewBox%3D%220%200%2016%2016%22%20width%3D%2216%22%20height%3D%2216%22%3E%0A%20%20%3Ctitle%20id%3D%22cm-drag-to-resize%22%3EDrag%20to%20resize%3C%2Ftitle%3E%0A%20%20%3Cpath%20fill-rule%3D%22evenodd%22%20d%3D%22M1%202.75A.75.75%200%20011.75%202h12.5a.75.75%200%20110%201.5H1.75A.75.75%200%20011%202.75zm0%205A.75.75%200%20011.75%207h12.5a.75.75%200%20110%201.5H1.75A.75.75%200%20011%207.75zM1.75%2012a.75.75%200%20100%201.5h12.5a.75.75%200%20100-1.5H1.75z%22%3E%3C%2Fpath%3E%0A%3C%2Fsvg%3E

I'm going to use a variant of the Datasette menu icon. Here it is in #ccc with an ARIA label:

<svg aria-labelledby="cm-drag-to-resize" role="img" fill="#ccc" stroke="#ccc" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 16 16" width="16" height="16">
  <title id="cm-drag-to-resize">Drag to resize</title>
  <path fill-rule="evenodd" d="M1 2.75A.75.75 0 011.75 2h12.5a.75.75 0 110 1.5H1.75A.75.75 0 011 2.75zm0 5A.75.75 0 011.75 7h12.5a.75.75 0 110 1.5H1.75A.75.75 0 011 7.75zM1.75 12a.75.75 0 100 1.5h12.5a.75.75 0 100-1.5H1.75z"></path>
</svg>

I only want it to resize up and down, not left to right - so I'm not keen on the default resize handle:

https://user-images.githubusercontent.com/9599/108573363-364de680-72c9-11eb-8741-5112463ebfaa.png">

https://rawgit.com/Sphinxxxx/cm-resize/master/demo/index.html

I can use https://github.com/Sphinxxxx/cm-resize for this

Big usability improvement, see also #1236

Documentation: https://sqlite-utils.datasette.io/en/latest/cli.html#attaching-additional-databases

Documentation: https://sqlite-utils.datasette.io/en/latest/python-api.html#attaching-additional-databases

This feature is now released! https://docs.datasette.io/en/stable/changelog.html#v0-55

I bumped the two lines in the Dockerfile to FROM python:3.7.10-slim-stretch as build and ran this to build it:

docker build -f Dockerfile -t datasetteproject/datasette:python-3-7-10 .

Then I ran it with:

docker run -p 8001:8001 -v `pwd`:/mnt datasetteproject/datasette:python-3-7-10 datasette -p 8001 -h 0.0.0.0 /mnt/fixtures.db

http://0.0.0.0:8001/-/versions confirmed that it was now running Python 3.7.10

I'll bump to 3.7.10 for the moment - the fix for 3.8 isn't out until March 1st according to https://news.ycombinator.com/item?id=26186434

https://www.python.org/downloads/release/python-3710/

Demo is now live here: https://latest.datasette.io/_memory

The documentation is at https://docs.datasette.io/en/latest/sql_queries.html#cross-database-queries - it links to this example query: https://latest.datasette.io/_memory?sql=select%0D%0A++%27fixtures%27+as+database%2C+*%0D%0Afrom%0D%0A++%5Bfixtures%5D.sqlite_master%0D%0Aunion%0D%0Aselect%0D%0A++%27extra_database%27+as+database%2C+*%0D%0Afrom%0D%0A++%5Bextra_database%5D.sqlite_master

The implementation in #1232 is ready to land. It's the simplest-thing-that-could-possibly-work: you can run datasette one.db two.db three.db --crossdb and then use the /_memory page to run joins across tables from multiple databases.

It only works on the first 10 databases that were passed to the command-line. This means that if you have a Datasette instance with hundreds of attached databases (see Datasette Library) this won't be particularly useful for you.

So... a better, future version of this feature would be one that lets you join across databases on command - maybe by hitting /_memory?attach=db1&attach=db2 to get a special connection.

Also worth noting: plugins that implement the prepare_connection() hook can attach additional databases - so if you need better, customized support for this one way to handle that would be with a custom plugin.

UI listing the attached tables:

https://user-images.githubusercontent.com/9599/108424809-8015ce80-71ee-11eb-8752-03c459f89320.png">

I tested the demo deployment like this:

datasette publish cloudrun fixtures.db extra_database.db \                                      
            -m fixtures.json \
            --plugins-dir=plugins \
            --branch=crossdb \
            --extra-options="--setting template_debug 1 --crossdb" \
            --install=pysqlite3-binary \
            --service=datasette-latest-crossdb

Due to bug #1233 I'm going to publish the additional database as extra_database.db rather than extra database.db as it is used in the tests.

I think the bug is here: https://github.com/simonw/datasette/blob/640ac7071b73111ba4423812cd683756e0e1936b/datasette/utils/__init__.py#L349-L373

I'll document this option on https://docs.datasette.io/en/stable/sql_queries.html

The new warning looks like this:

https://user-images.githubusercontent.com/9599/108420562-d41db480-71e8-11eb-87e3-1cddd65627c2.png">

It would also be neat if https://latest.datasette.io/ had multiple databases attached in order to demonstrate this feature.

It would be neat if the /_memory page showed a list of attached databases, to indicate that the --crossdb option is working and give people links to click to start running queries.

I have a working version now, moving development to a pull request.

For the moment I'm going to hard-code a SQLITE_LIMIT_ATTACHED=10 constant and only attach the first 10 databases to the _memory connection.

select * from pragma_database_list(); is useful - shows all attached databases for the current connection.

It turns out SQLite defaults to a maximum of 10 attached databases. This can be increased using a compile-time constant, but even with that it cannot be more than 62: https://stackoverflow.com/questions/9845448/attach-limit-10

Ideally I'd figure out a way to replicate this error in a concurrent unit test.

I started trying to use locks to resolve this but I've not figured out the right way to do that yet - here's my first experiment:

diff --git a/datasette/app.py b/datasette/app.py
index 9e15a16..1681c9d 100644
--- a/datasette/app.py
+++ b/datasette/app.py
@@ -217,6 +217,7 @@ class Datasette:
         self.inspect_data = inspect_data
         self.immutables = set(immutables or [])
         self.databases = collections.OrderedDict()
+        self._refresh_schemas_lock = threading.Lock()
         if memory or not self.files:
             self.add_database(Database(self, is_memory=True), name="_memory")
         # memory_name is a random string so that each Datasette instance gets its own
@@ -324,6 +325,13 @@ class Datasette:
         self.client = DatasetteClient(self)

     async def refresh_schemas(self):
+        return
+        if self._refresh_schemas_lock.locked():
+            return
+        with self._refresh_schemas_lock:
+            await self._refresh_schemas()
+
+    async def _refresh_schemas(self):
         internal_db = self.databases["_internal"]
         if not self.internal_db_created:
             await init_internal_db(internal_db)