There's a ton of tools/languages that compile to SQL, which may be nice in Datasette. Some examples:

• Logica https://logica.dev
• PRQL https://prql-lang.org
• Malloy, but not sure if it works with SQLite? https://github.com/malloydata/malloy

It would be cool if plugins could extend Datasette to use these languages, in both the code editor and API usage.

A few things I'd imagine a datasette-prql or datasette-logica plugin would do:

• prql= instead of sql=
• Code editor support (syntax highlighting, autocomplete)
• Hide/show SQL

Very often I get tripped up when trying to configure my Datasette instances. For example: if I want to change the port my app listen too, do I do that with a CLI flag, a --setting flag, inside metadata.json, or an env var? If I want to up the time limit of SQL statements, is that under metadata.json or a setting? Where does my plugin configuration go?

Normally I need to look it up in Datasette docs, and I quickly find my answer, but the number of places where "config" goes it overwhelming.

• Flat CLI flags like --port, --host, --cors, etc.
• --setting, like default_page_size, sql_time_limit_ms etc
• Inside metadata.json, including plugin configuration

Typically my Datasette deploys are extremely long shell commands, with multiple --setting and other CLI flags.

Proposal: Consolidate all "config" into datasette.toml

I propose that we add a new datasette.toml that combines "settings", "metadata", and other common CLI flags like --port and --cors into a single file. It would be similar to "Cargo.toml" in Rust projects, "package.json" in Node projects, and "pyproject.toml" in Python, etc.

A sample of what it could look like:

```toml

"top level" configuration that are currently CLI flags on datasette serve

[config] port = 8020 host = "0.0.0.0" cors = true

replaces multiple --setting flags

[settings] base_url = "/app/datasette/" default_allow_sql = true sql_time_limit_ms = 3500

replaces metadata.json.

The contents of datasette-metadata.json could be defined in this file instead, but supporting separate files is nice (since those are easy to machine-generate)

[metadata] include="./datasette-metadata.json"

plugin-specific

[plugins] [plugins.datasette-auth-github] client_id = {env = "DATASETTE_AUTH_GITHUB_CLIENT_ID"} client_secret = {env = "GITHUB_CLIENT_SECRET"}

[plugins.datasette-cluster-map]

latitude_column = "lat" longitude_column = "lon" ```

Pros

• Instead of multiple files and CLI flags, everything could be in one tidy file
• Editing config in a separate file is easier than editing CLI flags, since you don't have to kill a process + edit a command every time
• New users will know "just edit my datasette.toml instead of needing to learn metadata + settings + CLI flags
• Better dev experience for multiple environment. For example, could have datasette -c datasette-dev.toml for local dev environments (enables SQL, debug plugins, long timeouts, etc.), and a datasette -c datasette-prod.toml for "production" (lower timeouts, less plugins, monitoring plugins, etc.)

Cons

• Yet another config-management system. Now Datasette users will need to know about metadata, settings, CLI flags, and datasette.toml. However with enough documentation + announcements + examples, I think we can get ahead of it.
• If toml is chosen, would need to add a toml parser for Python version <3.11
• Multiple sources of config require priority. For example: Would --setting default_allow_sql off override the value inside [settings]? What about --port?

Other Notes

Toml

I chose toml over json because toml supports comments. I chose toml over yaml because Python 3.11 has builtin support for it. I also find toml easier to work with since it doesn't have the odd "gotchas" that YAML has ("ex 3.10 resolving to 3.1, Norway NO resolving to false, etc.). It also mimics pyproject.toml which is nice. Happy to change my mind about this however

Plugin config will be difficult

Plugin config is currently in metadata.json in two places:

1. Top level, under "plugins.[plugin-name]". This fits well into datasette.toml as [plugins.plugin-name]
2. Table level, under "databases.[db-name].tables.[table-name].plugins.[plugin-name]. This doesn't fit that well into datasette.toml, unless it's nested under [metadata]?

Extensions, static, one-off plugins?

We could also include equivalents of --plugins-dir, --static, and --load-extension into datasette.toml, but I'd imagine there's a few security concerns there to think through.

Explicitly list with plugins to use?

I believe Datasette by default will load all install plugins on startup, but maybe datasette.toml can specify a list of plugins to use? For example, a dev version of datasette.toml can specify datasette-pretty-traces, but the prod version can leave it out

The current _internal database is used by Datasette core to cache info about databases/tables/columns/foreign keys of databases in a Datasette instance. It's a temporary database created at startup, that can only be seen by the root user. See an example _internal DB here, after logging in as root.

The current _internal database has a few rough edges:

• It's part of datasette.databases, so many plugins have to specifically exclude _internal from their queries examples here
• It's only used by Datasette core and can't be used by plugins or 3rd parties
• It's created from scratch at startup and stored in memory. Why is fine, the performance is great, but persistent storage would be nice.

Additionally, it would be really nice if plugins could use this _internal database to store their own configuration, secrets, and settings. For example:

• datasette-auth-tokens creates a _datasette_auth_tokens table to store auth token metadata. This could be moved into the _internal database to avoid writing to the gues database
• datasette-socrata creates a socrata_imports table, which also can be in _internal
• datasette-upload-csvs creates a _csv_progress_ table, which can be in _internal
• datasette-write-ui wants to have the ability for users to toggle whether a table appears editable, which can be either in datasette.yaml or on-the-fly by storing config in _internal

In general, these are specific features that Datasette plugins would have access to if there was a central internal database they could read/write to:

• Dynamic configuration. Changing the datasette.yaml file works, but can be tedious to restart the server every time. Plugins can define their own configuration table in _internal, and could read/write to it to store configuration based on user actions (cell menu click, API access, etc.)
• Caching. If a plugin or Datasette Core needs to cache some expensive computation, they can store it inside _internal (possibly as a temporary table) instead of managing their own caching solution.
• Audit logs. If a plugin performs some sensitive operations, they can log usage info to _internal for others to audit later.
• Long running process status. Many plugins (datasette-upload-csvs, datasette-litestream, datasette-socrata) perform tasks that run for a really long time, and want to give continue status updates to the user. They can store this info inside_internal
• Safer authentication. Passwords and authentication plugins usually store credentials/hashed secrets in configuration files or environment variables, which can be difficult to handle. Now, they can store them in _internal

Proposal

• We remove _internal from datasette.databases property.
• We add new datasette.get_internal_db() method that returns the _internal database, for plugins to use
• We add a new --internal internal.db flag. If provided, then the _internal DB will be sourced from that file, and further updates will be persisted to that file (instead of an in-memory database)
• When creating internal.db, create a new _datasette_internal table to mark it a an "datasette internal database"
• In datasette serve, we check for the existence of the _datasette_internal table. If it exists, we assume the user provided that file in error and raise an error. This is to limit the chance that someone accidentally publishes their internal database to the internet. We could optionally add a --unsafe-allow-internal flag (or database plugin) that allows someone to do this if they really want to.

New features unlocked with this

These features don't really need a standardized _internal table per-say (plugins could currently configure their own long-time storage features if they really wanted to), but it would make it much simpler to create these kinds of features with a persistent application database.

• datasette-comments : A plugin for commenting on rows or specific values in a database. Comment contents + threads + email notification info can be stored in _internal
• Bookmarks: "Bookmarking" an SQL query could be stored in _internal, or a URL link shortener
• Webhooks: If a plugin wants to either consume a webhook or create a new one, they can store hashed credentials/API endpoints in _internal

Metadata in Datasette is a really powerful feature, but is a bit difficult to work with. It was initially a way to add "metadata" about your "data" in Datasette instances, like descriptions for databases/tables/columns, titles, source URLs, licenses, etc. But it later became the go-to spot for other Datasette features that have nothing to do with metadata, like permissions/plugins/canned queries.

Specifically, I've found the following problems when working with Datasette metadata:

1. Metadata cannot be updated without re-starting the entire Datasette instance.
2. The metadata.json/metadata.yaml has become a kitchen sink of unrelated (imo) features like plugin config, authentication config, canned queries
3. The Python APIs for defining extra metadata are a bit awkward (the datasette.metadata() class, get_metadata() hook, etc.)

Possible solutions

Here's a few ideas of Datasette core changes we can make to address these problems.

Re-vamp the Datasette Python metadata APIs

The Datasette object has a single datasette.metadata() method that's a bit difficult to work with. There's also no Python API for inserted new metadata, so plugins have to rely on the get_metadata() hook.

The get_metadata() hook can also be improved - it doesn't work with async functions yet, so you're quite limited to what you can do.

(I'm a bit fuzzy on what to actually do here, but I imagine it'll be very small breaking changes to a few Python methods)

Add an optional datasette_metadata table

Datasette should detect and use metadata stored in a new special table called datasette_metadata. This would be a regular table that a user can edit on their own, and would serve as a "live updating" source of metadata, than can be changed while the Datasette instance is running.

Not too sure what the schema would look like, but I'd imagine:

sql CREATE TABLE datasette_metadata( level text, target any, key text, value any, primary key (level, target) )

Every row in this table would map to a single metadata "entry".

• level would be one of "datasette", "database", "table", "column", which is the "level" the entry describes. For example, level="table" means it is metadata about a specific table, level="database" for a specific database, or level="datasette" for the entire Datasette instance.
• target would "point" to the specific object the entry metadata is about, and would depend on what level is specific.
• level="database": target would be the string name of the database that the metadata entry is about. ex "fixtures"
• level="table": target would be a JSON array of two strings. The first element would be the database name, and the second would be the table name. ex ["fixtures", "students"]
• level="column": target would be a JSON array of 3 strings: The database name, table name, and column name. Ex ["fixtures", "students", "student_id"]
• key would be the type of metadata entry the row has, similar to the current "keys" that exist in metadata.json. Ex "about_url", "source", "description", etc
• value would be the text value of be metadata entry. The literal text value of a description, about_url, column_label, etc

A quick sample:

level | target | key | value -- | -- | -- | -- datasette | NULL | title | my datasette title... db | fixtures | source | <description of my database source> table | ["fixtures", "students"] | label_column | student_name column | ["fixtures", "students", "birthdate"] | description | <description of the fixtures.students.birthdate column>

This datasette_metadata would be configured with other tools, and hopefully not manually by end users. Datasette Core could also offer a UI for editing entries in datasette_metadata, to update descriptions/columns on the fly.

Re-vamp metadata.json and move non-metadata config to another place

The motivation behind this is that it's awkward that metadata.json contains config about things that are not strictly metadata, including:

• Plugin configuration
• Authentication/permissions (ex the allow key on datasettes/databases/tables
• Canned queries. might be controversial, but in my mind, canned queries are application-specific code and configuration, and don't describe the data that exists in SQLite databases.

I think we should move these outside of metadata.json and into a different file. The datasette.json idea in #2093 may be a good solution here: plugin/permissions/canned queries can be defined in datasette.json, while metadata.json/datasette_metadata will strictly be about documenting databases/tables/columns.

sqlite-utils should detect and have specially handling for column with a JSON column. For example:

sql CREATE TABLE "dogs" ( id INTEGER PRIMARY KEY, name TEXT, friends JSON );

Automatic Nesting

According to "Nested JSON Values", sqlite-utils will only expand JSON if the --json-cols flag is passed. It looks like it'll try to json.load all text column to test if its JSON, which can get expensive on non-json columns.

Instead, sqlite-utils should be default (ie without the --json-cols flags) do the maybe_json() operation on columns with a declared JSON type. So the above table would expand the "friends" column as expected, withoutthe --json-cols flag:

bash sqlite-utils dogs.db "select * from dogs" | python -mjson.tool

[ { "id": 1, "name": "Cleo", "friends": [ { "name": "Pancakes" }, { "name": "Bailey" } ] } ]

I'm sure there's other ways sqlite-utils can specially handle JSON columns, so keeping this open while I think of more

I'd like there to be a plugin system for sqlite-utils, similar to the datasette/llm plugins. I'd like to make plugins that would do things like:

• Register SQLite extensions for more SQL functions + virtual tables
• Register new subcommands
• Different input file formats for sqlite-utils memory
• Different output file formats (in addition to --csv --tsv --nl etc.

A few real-world use-cases of plugins I'd like to see in sqlite-utils:

• Register many of my sqlite extensions in sqlite-utils (sqlite-http, sqlite-lines, sqlite-regex, etc.)
• New subcommands to work with sqlite-vss vector tables
• Input/ouput Parquet/Avro/Arrow IPC files with sqlite-arrow

When #2052 merges, I'd like to add support to add extensions/functions to the Datasette code editor.

I'd eventually like to build a JS plugin for sqlite-docs, to add things like:

• Inline documentation for tables/columns on hover
• Inline docs for custom functions that are loaded in
• More detailed autocomplete for tables/columns/functions

I did some hacking to see what this would look like, see here:

There can be a new hook that allows JS plugins to add new "extension" in the CodeMirror editorview here:

https://github.com/simonw/datasette/blob/8cd60fd1d899952f1153460469b3175465f33f80/datasette/static/cm-editor-6.0.1.js#L25

Will need some more planning. For example, the Codemirror bundle in Datasette has functions that we could re-export for plugins to use (so we don't load 2 version of "@codemirror/autocomplete", for example.

From the Datasette discord: A user tried running the following command on windows:

datasette --load-extension="C:\spatialite\mod_spatialite-5.0.1-win-x86\mod_spatialite.dll" This failed with "The specified module could not be found", because the entrypoint option introduced in #1789 splits the input differently. Instead of loading the extension found at "C:\spatialite\mod_spatialite-5.0.1-win-x86\mod_spatialite.dll", it instead tried to load the extension at "C" with entrypoint `"\spatialite\mod_spatialite-5.0.1-win-x86\mod_spatialite.dll".

This is hard because most absolute windows paths have a colon in them, like C:\foo.txt or D:\bar.txt. I'd image the --static flag is also vulnerable to this type of bug.

The "solution" is to use a relative path instead, but that doesn't feel that great.

Problem

SQLite extensions have the option to define multiple "entrypoints" in each loadable extension. For example, the upcoming version of sqlite-lines will have 2 entrypoints: the default sqlite3_lines_init (which SQLite will automatically guess for) and sqlite3_lines_noread_init. The sqlite3_lines_noread_init version omits functions that read from the filesystem, which is necessary for security purposes when running untrusted SQL (which Datasette does).

(Similar multiple entrypoints will also be added for sqlite-http).

The --load-extension flag, however, doesn't give the option to specify a different entrypoint, so the default one is always used.

Proposal

I want there to be a new command line option of the --load-extension flag to specify a custom entrypoint like so: datasette my.db \ --load-extension ./lines0 sqlite3_lines0_noread_init

Then, under the hood, this line of code:

https://github.com/simonw/datasette/blob/7af67b54b7d9bca43e948510fc62f6db2b748fa8/datasette/app.py#L562

Would look something like this:

python conn.execute("SELECT load_extension(?, ?)", [extension, entrypoint])

One potential problem: For backward compatibility, I'm not sure if Click allows cli flags to have variable number of options ("arity"). So I guess it could also use a : delimiter like --static:

datasette my.db \ --load-extension ./lines0:sqlite3_lines0_noread_init

Or maybe even a new flag name?

datasette my.db \ --load-extension-entrypoint ./lines0 sqlite3_lines0_noread_init

Personally I prefer the : option... and maybe even --load-extension -> --load? Definitely out of scope for this issue tho

datasette my.db \ --load./lines0:sqlite3_lines0_noread_init

Currently for canned queries, you have to inline SQL in your metadata.yaml like so:

yaml databases: fixtures: queries: neighborhood_search: sql: |- select neighborhood, facet_cities.name, state from facetable join facet_cities on facetable.city_id = facet_cities.id where neighborhood like '%' || :text || '%' order by neighborhood title: Search neighborhoods

This works fine, but for a few reasons, I usually have my canned queries already written in separate .sql files. I'd like to instead re-use those instead of re-writing it.

So, I'd like to see a new "sql_file" key that works like so:

metadata.yaml:

yaml databases: fixtures: queries: neighborhood_search: sql_file: neighborhood_search.sql title: Search neighborhoods neighborhood_search.sql: sql select neighborhood, facet_cities.name, state from facetable join facet_cities on facetable.city_id = facet_cities.id where neighborhood like '%' || :text || '%' order by neighborhood

Both of these would work in the exact same way, where Datasette would instead open + include neighborhood_search.sql on startup.

A few reasons why I'd like to keep my canned queries SQL separate from metadata.yaml:

• Keeping SQL in standalone SQL files means syntax highlighting and other text editor integrations in my code
• Multiline strings in yaml, while functional, are a tad cumbersome and are hard to edit
• Works well with other tools (can pipe .sql files into the sqlite3 CLI, or use with other SQLite clients easier)
• Typically my canned queries are quite long compared to everything else in my metadata.yaml, so I'd love to separate it where possible

Let me know if this is a feature you'd like to see, I can try to send up a PR if this sounds right!

I have a SQLite DB that's pretty large, 23GB and something like 300 million rows. I expect that most queries I run on it will be slow, which is fine, but there are some things that Datasette does that makes working with the DB very slow. Specifically, when I access the .json metadata for a table (which I believe it comes from datasette/views/database.py, it takes 43 seconds for the request to come in:

bash $ time curl localhost:9999/out.json {"database": "out", "size": 24291454976, "tables": [{"name": "PageviewsHour", "columns": ["file", "code", "page", "pageviews"], "primary_keys": [], "count": null, "hidden": false, "fts_table": null, "foreign_keys": {"incoming": [], "outgoing": [{"other_table": "PageviewsHourFiles", "column": "file", "other_column": "file_id"}]}, "private": false}, {"name": "PageviewsHourFiles", "columns": ["file_id", "filename", "sha256", "size", "day", "hour"], "primary_keys": ["file_id"], "count": null, "hidden": false, "fts_table": null, "foreign_keys": {"incoming": [{"other_table": "PageviewsHour", "column": "file_id", "other_column": "file"}], "outgoing": []}, "private": false}, {"name": "sqlite_sequence", "columns": ["name", "seq"], "primary_keys": [], "count": 1, "hidden": false, "fts_table": null, "foreign_keys": {"incoming": [], "outgoing": []}, "private": false}], "hidden_count": 0, "views": [], "queries": [], "private": false, "allow_execute_sql": true, "query_ms": 43340.23213386536} real 0m43.417s user 0m0.006s sys 0m0.016s

I suspect this is because a COUNT(*) is happening under the hood, which, when I run it through sqlite directly, does take around the same time:

```bash $ time sqlite3 out.db < <(echo "select count(*) from PageviewsHour;") 362794272

real 0m44.523s user 0m2.497s sys 0m6.703s ```

I'm using the .json request in the Observable Datasette Client to 1) verify that a link passed in is a reachable Datasette instance, and 2) a quick way to look at metadata for a db. A few different solutions I can think of:

1. Have some other endpoint, like /-/datasette.json that the Observable Datasette client can fetch from to verify that the passed in URL is a valid Datasette (doesnt solve the slow problem, feel free to split this issue into 2)
2. Have a way to turn off table counts when accessing a database's .json view, like ?no_count=1 or something
3. Maybe have a timeout on the table_counts() function if it takes too long. which is odd, because it seems like it already does that (I think?), I can debug a little more if that's the case

More than happy to debug further, or send a PR if you like one of the proposals above!