issue_comments

I'd like to start working on this. I've made a custom template for index.html that contains a form that contains a search input. But I'm not sure where to go from here. When user enters a search term, I'd like for that term to go into a function I'll make that will search all tables with full text search enabled.

Can I make additional custom Python scripts for this or must I edit datasette's files directly?

So maybe something like this: curl https://api.github.com/repos/simonw/datasette/pulls?state=all | \ sqlite-utils insert git.db pulls - \ --flatten=base \ --flatten=head \ --extract=user:users:id \ --extract=head_repo.license:licenses:key \ --extract=head_repo.owner:users \ --extract=head_repo --extract=base_repo.license:licenses:key \ --extract=base_repo.owner:users \ --extract=base_repo Is the order of those nested --extract lines significant I wonder? It would be nice if the order didn't matter and the code figured out the right execution plan on its own.

@IgnoredAmbience this is now shipped in sqlite-utils 1.2 - documentation here:

• https://sqlite-utils.readthedocs.io/en/latest/python-api.html#python-api-defaults-not-null
• https://sqlite-utils.readthedocs.io/en/latest/cli.html#cli-defaults-not-null

I'm going to reuse the ForeignKey named tuple here:

https://github.com/simonw/sqlite-utils/blob/d645032cfa4edbccd0542eecdddca29edf9f7b07/sqlite_utils/db.py#L17-L19

Still need to add this mechanism to .create_table() - this code here is all that needs to be modified - it needs to learn to deal with the alternative syntax for foreign keys and guess the missing data if necessary:

https://github.com/simonw/sqlite-utils/blob/d645032cfa4edbccd0542eecdddca29edf9f7b07/sqlite_utils/db.py#L115-L119

I may ignore --flatten for the moment - users can do their own flattening using jq if they need that.

curl https://api.github.com/repos/simonw/datasette/pulls?state=all | jq " [.[] | . + { base_label: .base.label, base_ref: .base.ref, base_sha: .base.sha, base_user: .base.user, base_repo: .base.repo, head_label: .head.label, head_ref: .head.ref, head_sha: .head.sha, head_user: .head.user, head_repo: .head.repo } | del(.base, .head, ._links)] " Output: https://gist.github.com/simonw/2703ed43fcfe96eb8cfeee7b558b61e1

So for --extract the format is path-to-property:table-to-extract-to:primary-key

If we find an array (as opposed to a direct nested object) at the end of the dotted path we do a m2m table.

And if primary-key is omitted maybe we do the rowid thing with a foreign key back to ourselves.

Another complexity from the https://api.github.com/repos/simonw/datasette/pulls example:

We don't actually want head and base to be pulled out into a separate table. Our ideal table design would probably look something like this:

• url: ...
• id: 285698310
• ...
• user_id: 9599 -> refs users
• head_label: simonw:travis-38dev
• head_ref: travis-38dev
• head_sha: f274f9004302c5ca75ce89d0abfd648457957e31
• head_user_id: 9599 -> refs users
• head_repo_id: 107914493 -> refs repos
• base_label: simonw:master
• base_ref: master
• base_sha: 5e8fbf7f6fbc0b63d0479da3806dd9ccd6aaa945
• base_user_id: 9599 -> refs users
• base_repo_id: 107914493 -> refs repos

So the nested head and base sections here, instead of being extracted into another table, were flattened into their own columns.

So perhaps we need a flatten-nested-into-columns mechanism which can be used in conjunction with a extract-to-tables mechanism.

I like the term "extract" for what we are doing here, partly because that's the terminology I used in csvs-to-sqlite.

There's an interesting difference here between nested objects with a primary-key style ID and nested objects without.

If a nested object does not have a primary key, we could still shift it out to another table but it would need to be in a context where it has an automatic foreign key back to our current record.

A good example of something where that would be useful is the outageDevices key in https://github.com/simonw/pge-outages/blob/d890d09ff6e2997948028528e06c82e1efe30365/pge-outages.json#L13-L25

json { "outageNumber": "407367", "outageStartTime": "1560355216", "crewCurrentStatus": "PG&E repair crew is on-site working to restore power.", "currentEtor": "1560376800", "cause": "Our preliminary determination is that your outage was caused by scheduled maintenance work.", "estCustAffected": "3", "lastUpdateTime": "1560355709", "hazardFlag": "0", "latitude": "37.35629", "longitude": "-119.70469", "outageDevices": [ { "latitude": "37.35409", "longitude": "-119.70575" }, { "latitude": "37.35463", "longitude": "-119.70525" }, { "latitude": "37.35562", "longitude": "-119.70467" } ], "regionName": "Ahwahnee" }

These could either be inserted into an outageDevices table that uses rowid... or we could have a mechanism where we automatically derive a primary key for them based on a hash of their data, hence avoiding creating duplicates even though we don't have a provided primary key.

I think I can do something here with a very simple head.repo.owner path syntax. Normally this kind of syntax would have to take the difference between dictionaries and lists into account but I don't think that matters here.

If I'm going to do this then I should make the other_table and other_column arguments optional here too:

https://github.com/simonw/sqlite-utils/blob/2fed87da6ea990d295672e4db2c8ae97b787913e/sqlite_utils/cli.py#L201-L215

Maybe foreign_keys could even optionally just be a list of columns - it could then attempt to detect the related tables based on some rules-of-thumb and raise an error if there's no obvious candidate.

Rules: * If the column name ends in _id, remove that suffix and look for a matching table. * Try for a table which is the column name without the _id suffix with an s appended to it * Try for a table that's the exact match for the column name

If none of these rules match, raise an error.

So the above example could be further simplified to: python db["usages"].insert_all( usages_to_insert, foreign_keys=["line_id", "definition_id"] )

Maybe it's time to create a sqlite-utils create-table command here too, rather than forcing people to create tables only by inserting example data.

But what to do for creating a table?

For the Python function I could do this: python db["cats"].create({ "id": int, "name": str, "score": int, "weight": float, }, pk="id", not_null={"weight"}, defaults={"score": 1})

The CLI tool only every creates tables as a side-effect of a sqlite-utils insert or sqlite-utils upsert. I can have them accept optional arguments, --not-null colname and --default colname value: echo '{"name": "Cleo", "age": 4, "score": 2}' | \ sqlite-utils insert dogs.db dogs - \ --not-null age \ --not-null name \ --default score 1

Since you can't have one without the other, I'm going with --not-null-default= and not_null_default= for the add column versions of this.

Hmm... we need the ability to pass --not-null when we are creating a table as well.

If you attempt to add NOT NULL to a column after a table has first been created you get this error:

sqlite3.OperationalError: Cannot add a NOT NULL column with default value NULL

Two challenges here:

1. We need a way to specify which tables should be used - e.g. "put records from the "user" key in a users table, put multiple records from the "labels" key in a table called labels" (we can pick an automatic name for the m2m table, though it might be nice to have an option to customize it)

2. Should we deal with nested objects? Consider https://api.github.com/repos/simonw/datasette/pulls for example:

Here we have head.user as a user, head.repo as a repo, and head.repo.owner as another user.

Ideally our mechanism for specifying which table things should be pulled out into would handle this, but it's getting a bit complicated.