issue_comments

I think this feature was removed in Datasette 0.61 and moved to a plugin. People who want hashed URLs can use the datasette-hashed-urls plugin to achieve the same affect.

It looks like you're trying to disable hashed urls, so I think you can just remove that config setting and things will work.

Ha, yes, I might end up making something very niche. That's OK.

I'm building a UI for Datasette that lets users make schema changes, so it's important to me that the tool work in a non-surprising way -- if you ask for a column of type X, you should get type X. If the column or table previously had CHECK constraints, they shouldn't be silently removed. And so on. I had hoped that I could just lean on sqlite-utils, but I think it's a little too surprising.

I think the bigger issue is that sqlite-utils mixes mechanism (it implements the 12-step way to alter SQLite tables) and policy (it has an opinionated stance on what column types should be used).

That might be a design choice to make it accessible to users by providing a reasonable set of defaults, but it doesn't quite fit my use case.

It might make sense to extract a separate library that provides just the mechanisms, and then sqlite-utils would sit on top of that library with its opinionated set of policies.

That would be a very big change, though.

I might take a stab at extracting the library, but just for the table schema migration piece, not all the other features that sqlite-utils supports. I wouldn't expect sqlite-utils to depend on it.

Part of my motivation is that I want to provide some other abilities, too, like support for CHECK constraints. I see that the issue in this repo (https://github.com/simonw/sqlite-utils/issues/358) proposes a bunch of short-hand constraints, which I wouldn't want to accidentally expose to people -- I want a layer that is a 1:1 mapping to SQLite.

I'd support that, but I'm not the author of this library.

One challenge is that would be a breaking change. Do you see a way to enable it without affecting existing users or bumping the major version number?

Ah, it looks like that is controlled by this dict: https://github.com/simonw/sqlite-utils/blob/main/sqlite_utils/db.py#L178

I suspect you could overwrite the datetime entry to achieve what you want

I think it's not currently possible: sqlite-utils requires that it be one of integer, text, float, blob (see code)

IMO, this is a bit of friction and it would be nice if it was more permissive. SQLite permits developers to use any data type when creating a table. For example, this is a perfectly cromulent sqlite session that creates a table with columns of type baz and bar:

``` sqlite> create table foo(column1 baz, column2 bar); sqlite> .schema foo CREATE TABLE foo(column1 baz, column2 bar); sqlite> select * from pragma_table_info('foo'); cid name type notnull dflt_value pk

0 column1 baz 0 0
1 column2 bar 0 0
```

The idea is that the application developer will know what meaning to ascribe to those types. For example, I'm working on a plugin to Datasette. Dates are tricky to handle. If you have some existing rows, you can look at the values in them to know how a user is serializing the dates -- as an ISO 8601 string? An RFC 3339 string? With millisecond precision? With timezone offset? But if you don't yet have any rows, you have to guess. If the column is of type TEXT, you don't even know that it's meant to hold a date! In this case, my plugin will look to see if the column is of type DATE or DATETIME, and assume a certain representation when writing.

Perhaps there is an argument that sqlite-utils is trying to conform to SQLite's strict mode, and that is why it limits the choices. In strict mode, SQLite requires that the data type be one of INT, INTEGER, REAL, TEXT, BLOB, ANY. But that can't be the case -- sqlite-utils supports FLOAT, which is not one of the valid types in strict mode, and it rejects INT, REAL and ANY, which are valid.

That said, it looks like the check is only enforced at the CLI level. If you use the API directly, I think it'll work.

I did some horrible things in https://github.com/cldellow/datasette-ui-extras/issues/2 to enable this in my plugin -- example here: https://dux-demo.fly.dev/cooking/posts?_facet=owner_user_id&owner_user_id=67

The implementation relies on two things:

• a filters_from_request hook that adds a good human description (unfortunately, without the benefit of the CSS styling you mention)
• doing something evil to hijack the exact and not operators in the Filters class. We can't leave them as is, or we'll get 2 human descriptions -- the built-in Datasette one and the one from my plugin. We can't remove them, or the filters UI will stop supporting the = and != operators

This got me thinking: it'd be neat if the list of operators that the filters UI supported wasn't a closed set.

A motivating example: adding a geospatial NEAR operator. Ideally it'd take two arguments - a target point and a radius, so you could express a filter like find me all rows whose lat/lng are within 10km of 43.4516° N, 80.4925° W. (Optionally, the UI could be enhanced if the geonames database was loaded and queried, so a user could say find me all rows whose lat/lng are within 10km of Kitchener, ON, and the city gets translated to a lat/lng for them)

Performance tests - I think most places don't have them as a formal gate enforced by CI. TypeScript and scalac seem to have tests that run to capture timings. The timings are included by a bot as a comment or build check, and also stored in a database so you can graph changes over time to spot regressions. Probably overkill for Datasette!

Window functions - oh, good point. Looks like Ubuntu shipped JSON1 support as far back as sqlite 3.11. I'll let this PR linger until there's a way to run against different SQLite versions. For now, I'm shipping this with datasette-ui-extras, since I think it's OK for a plugin to enforce a higher minimum requirement.

Tests - there actually did end up being test changes to capture the undercount bug of the current implementation, so the current implementation would fail against the new tests.

Perhaps a non-window function version could be written that uses random() instead of row_number() over () in order to get a unique key. It's technically not unique, but in practice, I imagine it'll work well.

I pushed a revised version which ends up being faster -- the example which currently takes 4 seconds now runs in 500ms.

Ack, this PR is broken. I see now that the inner.* is necessary for ensuring the correct count in the face of rows having duplicate values in views.

That fixes the overcounting, but I think can undercount when the rows have the same data, eg a view like:

sql SELECT '["bar"]' tags UNION ALL SELECT '["bar"]'

will produce a count of {"bar": 1 }, when it should be {"bar": 2}. In fact, this could apply in tables without primary keys, too.

If inner came from a base table that had a primary key or a rowid, we could use those column(s) to solve that case.

I guess a general solution would be to compute a window function so we have a distinct ID for each row. Will fiddle to see if I can get that working.

A +1 for switching to CustomRow: I think you currently only get a CustomRow if the result set had a column that was an fkey (this code)

Otherwise you get vanilla sqlite3.Rows, which will fail if you try to access .columns or lookup the cell by name, which surprised me recently

I don't know how/if you do automated tests for performance, so I haven't changed any of the tests.

Aha, it's user error on my part.

Adding

sqlite3_db_config.argtypes = [ctypes.c_void_p, ctypes.c_int, ctypes.c_int, ctypes.c_int]

makes it work reliably both on the CLI and from datasette, and now I can reproduce the errors you mentioned in the issue description.

I'm on Ubuntu, unfortunately. :( Would it still be relevant?

I think I've narrowed things down a bit more.

Even sqlite3_free(sqlite3_malloc(128)) segfaults -- this suggests to me that it's something about the sqlite3 library that was loaded, vs, say, getting the wrong db handle when I go spelunking in the Connection object.

Your comment introduced me to this issue in sqlite and to the ctypes module - thanks!

I also hope that the datasette developers will enable this mode in a test environment [...] perhaps we could figure out how to invoke it using ctypes

I'm not a Datasette developer, but I am curious to learn more about getting unholy access to the sqlite C APIs inside of Datasette. (Such access could also help #1293, and if done without grovelling inside of pysqlite's Connection object for the db handle, could even be relatively safe.)

I experimented a bit. I came up with https://gist.github.com/cldellow/85bba507c314b127f85563869cd94820

If you run python3 enable-strict-quoting-sqlite3.py, it seems to set those flags correctly -- SELECT "foo" fails where it would normally succeed.

But if you put it in a plugins/ dir and run datasette --plugins-dir plugins/, it segfaults when it tries to call sqlite3_db_config on the connections created by Datasette.

I am... confused. I'm pretty sure I'm using the same python and the same libsqlite3 in both scenarios, so I would expect it to work.

@gwk do you know anything that might help me debug the segfault? I gather that my approach of going grovelling inside of a PyObject is particularly dangerous, but I was thinking (a) it's necessary in order to test Datasette's use of the sqlite3 library and (b) even if it's not portable, it'd be good enough for running the tests on a single machine.

Actually, I discovered your post showing how a plugin can add a Datasette hook. That's wild! I've released datasette-rewrite-sql that adds this ability, albeit via monkey patching.

I had hoped to be able to expose request to the hook (or, even better actor) when the SQL was being run as a result of a user's HTTP request.

But some spelunking in the code makes me suspect that would actually require co-operation from Datasette itself. I'd be happy to be wrong and pointed in the right direction, though!

Hey @lovasoa, I hope you don't mind - I pulled this PR into datasette-ui-extras, a plugin I'm making that collects UI tweaks to Datasette.

You can apply it to your own Datasette instance by running datasette install datasette-ui-extras

Thanks for the response!

Yes, it does seem like a pretty nice developer experience--both the automagical labelling of fkeys, and the ability to index the row by column name in addition to column index.

I made a repo at https://github.com/code402/datasette-lambda to demonstrate the idea, and scratch my personal itch for this.

The demo relies on some central authority having already published a public, reusable Lambda layer with Datasette & its dependencies. I think that differs from the other publish plugins which seem to mainly publish Dockerfiles that the host will interpret to install deps from a requirements.txt file.

I chose that approach because uvloop appears to be a dependency with native code that needs to be compiled for the target runtime environment. In this case, that's Amazon Linux 2. I'm not 100% clear on whether that's still required, because:

• maybe uvloop is only needed for uvicorn, which the demo doesn't actually use since HTTP routing is handled by API Gateway
• it seems like uvloop may be an optional, drop-in optimization for asyncio in any case (but I may be misreading this; I'm very much a Python noob)

If it's the case that uvloop is truly optional, then I think the publish plugin could do the packaging on the user's machine, regardless of what flavour of operating system they're on. That'd be a bit slower for the user, but would provide the most long-term flexibility in terms of supporting plugins.

Hi Simon,

I'm thinking of attempting this. Can you clarify some questions I have?

1) I assume the goal is to have a CORS-friendly HTTPS endpoint that hosts the datasette service + user's db.

2) If that's the goal, I think Lambda alone is insufficient. Lambda provides the compute fabric, but not the HTTP routing. You'd also need to add Application Load Balancer or API Gateway to provide an HTTP endpoint that routes to the lambda function.

Do you have a preference between ALB or API GW? ALB has better economics at scale, but has a minimum monthly cost. API GW has worse per-request economics, but scales to zero when no requests are happening.

3) Does Datasette have any native components, or is it all pure python? If it has native bits, they'll likely need to be recompiled to work on Amazon Linux 2.

4) There are a few disparate services that need to be wired together to expose a Python service securely to the web. If I was doing this outside of the datasette publish system, I'd use an AWS CloudFormation template. Even within datasette, I think it still makes sense to use a CloudFormation template and just have the publish plugin invoke it (via the standard aws cli) with user-specified parameters. Does that sound reasonable to you?

Thanks for your help!