issue_comments

Most of this is shipped in https://docs.datasette.io/en/1.0a3/changelog.html#a3-2023-08-09

New thinking on the trimmed-down default. Previously I was going to use "row" and "next_url" - I now want to do this instead: json { "ok": true, "rows": [ { "pk1": "a", "pk2": "a", "pk3": "a", "content": "a-a-a" }, { "pk1": "a", "pk2": "a", "pk3": "b", "content": "a-a-b" } ], "next": "a,a,b" } If there isn't a next page it will return "next": null.

This is even more succinct. I'm OK with people having to request next_url if they don't want to construct the new URL themselves.

The "ok": true is there so it can be false for errors, consistently.

https://latest.datasette.io/fixtures/compound_three_primary_keys.json?_size=2 now returns count:

json { "database": "fixtures", "table": "compound_three_primary_keys", "is_view": false, "human_description_en": "", "rows": [ { "pk1": "a", "pk2": "a", "pk3": "a", "content": "a-a-a" }, { "pk1": "a", "pk2": "a", "pk3": "b", "content": "a-a-b" } ], "truncated": false, "count": 1001,

I'm going to rename filtered_table_rows_count to count - to match the SQL count(*) function.

Here's the so-far updated documentation for this change: https://github.com/simonw/datasette/blob/a2dca62360ad4a961d4c46f68eae41b7d5c7b2c9/docs/json_api.rst#different-shapes

https://latest.datasette.io/fixtures/compound_three_primary_keys.json?_size=2 now shows the new default: json { "database": "fixtures", "table": "compound_three_primary_keys", "is_view": false, "human_description_en": "", "rows": [ { "pk1": "a", "pk2": "a", "pk3": "a", "content": "a-a-a" }, { "pk1": "a", "pk2": "a", "pk3": "b", "content": "a-a-b" } ], The old format can be had like this: https://latest.datasette.io/fixtures/compound_three_primary_keys.json?_size=2&_shape=arrays

It turned out the most significant part of this change - switching from an array of arrays to an array of objects for the "rows" key - was really easy: Datasette already had a ?_shape=arrays v.s. ?_shape=objects mechanism, so I switched which one was the default in https://github.com/simonw/datasette/commit/234230e59574ccb8d8a24c45ccd325f725812377

I'm going to take another look at this: - https://github.com/simonw/datasette/issues/878

OK, I'm ready to start working on this today.

I'm going to go with a default representation that looks like this:

json { "rows": [ {"id": 1, "name": "One"}, {"id": 2, "name": "Two"} ], "next_url": null } Note that there's no count - all it provides is the current selection of results and an indication as to how the next can be retrieved (null if there are no more results).

I'll implement ?_extra= to provide everything else.

I'm going to keep ?_shape=array working on the assumption that many existing uses of the Datasette API are already using that option, so it would be nice not to break them.

@simonw:

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

Interesting! Although I don't think it matters too much what the underlying implementation is - I more meant that limit is familiar to developers conceptually as "up to and including this number, if they exist", whereas "size" is potentially more ambiguous. However, it's probably no big deal either way.

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.

I think it’s a good idea if the top level item of the response JSON is always an object, rather than an array, at least as the default.

I agree it is more predictable if the top level item is an object with a rows or data object that contains an array of data, which then allows for other top-level meta data.

I can see the argument for removing this and just using an array for convenience - but I think that's OK as an option (as you have now).

Rather than have lots of top-level keys you could have a "meta" object to contain non-data stuff. You could use something like "links" for API endpoint URLs (or use a standard like HAL). Which would then leave the top level a bit cleaner - if that's what you what.

Have you had much feedback from users who use the Datasette API a lot?

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)

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?

I like the idea of specifying a limit of 0 if you don’t want any rows data - and returning an empty array under the rows key seems fine.

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.

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 think it’s a good idea if the top level item of the response JSON is always an object, rather than an array, at least as the default. Mainly because it allows you to add extra keys in a backwards-compatible way. Also just seems more expected somehow.

The API design guidance for the UK government also recommends this: https://www.gov.uk/guidance/gds-api-technical-and-data-standards#use-json

I also strongly dislike having versioned APIs (eg with a /v1/ path prefix, as it invariably means that old versions stop working at some point, even though the bit of the API you’re using might not have changed at all.

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:

json [ { "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

json { "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?

In building https://latest-with-plugins.datasette.io/github/issue_comments.Notebook?_labels=on I discovered the following patterns for importing data into both Pandas and Observable/d3: python import pandas df = pandas.read_json( "https://latest-with-plugins.datasette.io/github/issue_comments.json?_shape=array" ) And: javascript d3 = require("d3@5") rows = d3.json( "https://latest-with-plugins.datasette.io/github/issue_comments.json?_shape=array" ) Once again I find myself torn on the best possible default. A list of JSON objects is instantly compatible with both pandas.read_json() and d3.json() - but it leaves nowhere to put the extra information like pagination and suchlike!

Even given this I still think the correct default is an object with "rows", "total" and "next_url" keys. I should commit to that and implement it - this thought exercise has been running for far too long.

https://latest-with-plugins.datasette.io/fixtures/roadside_attraction_characteristics/1.json-preview returns a 500 error at the moment - a KeyError on 'filtered_table_rows_count'.

This should be the key focus for Datasette 0.52.

I think this all comes down to how the ?_extras= mechanism works (see #262), as first hinted at in a30c5b220c15360d575e94b0e67f3255e120b916 (see commit message) when I added this long-forgotten undocumented feature: https://latest.datasette.io/fixtures/attraction_characteristic/2.json?_extras=foreign_key_tables

Extras need to be able to execute additional SQL, since that would solve the problem we have now where the expensive "suggested facets" code runs on all .json output even when its results are not being shown.

OK, https://latest-with-plugins.datasette.io/ is running that now - e.g. https://latest-with-plugins.datasette.io/fixtures/roadside_attractions.json-preview or https://latest-with-plugins.datasette.io/fixtures/compound_three_primary_keys.json-preview

json { "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 }

I'll update datasette-json-preview with that now.

Great point about CORS, I hadn't considered that.

I think I'm going to keep the Link: header (added in #1014) because I quite enjoy using it with GitHub and WordPress, but I'm not going to have it be the default way of doing pagination. For the default shape I'm now leaning towards this:

json { "total": 36, "rows": [{"id": 1, "name": "Cleo"}], "next_url": "https://latest-with-plugins.datasette.io/fixtures/facetable.json?_next=5" }

So three keys: total, rows and next_url. Then extra keys can be added using ?_extra= with various named bundles.

I vote against headers. It has a lot of strikes against it: poor discoverability, new developers often don’t know how to use them, makes CORS harder, makes it hard to use eg with JQ, needs ad hoc specification for each bit of metadata, etc.

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.

Here's the datasette-json-preview plugin I'll be using to experiment with different formats: https://github.com/simonw/datasette-json-preview

Building this plugin reminded me of an oddity of the register_output_renderer() plugin hook: one of the arguments that can be passed to it is data, which is the default internal data structure created by Datasette - but I deliberately avoided documenting that on https://docs.datasette.io/en/stable/plugin_hooks.html#register-output-renderer-datasette because it's not a stable interface.

That's not ideal. I'd like custom renderers to be able to access this data to get at things like suggested facets, on an opt-in basis.

So maybe that kind of stuff is re-implemented as "extras" which are awaitable callables - then renderer plugins can call the extras that they need to as part of their execution.

To illustrate the problem (in this case the need to access next_url) here's my first prototype of the plugin: ```python from datasette import hookimpl from datasette.utils.asgi import Response

@hookimpl def register_output_renderer(datasette): return { "extension": "json-preview", "render": json_preview, }

def json_preview(data, columns, rows): next_url = data.get("next_url") headers = {} if next_url: headers["link"] = '<{}>; rel="next"'.format(next_url) return Response.json([dict(zip(columns, row)) for row in rows], headers=headers) ```

I'm going to prototype what it would look like if the default shape was a list of objects and ?_extra= turns that into an object with a rows key, in a plugin. As a separate extension (maybe .json-preview).

Maybe .jsonfull becomes a new renderer that returns ALL of the defined ?_extra= blocks.

Or... ?_extra=all turns on ALL of the available information blocks (some of which can come from plugins).

Since the total count can be expensive to calculate, I'm inclined to make that an opt-in extra - maybe ?_extra=count.

Based on that, the default JSON shape could look something like this:

json { "rows": [{"id": 1}, {"id": 2}], "next": "2", "next_url": "/db/table?_next=2" } And with ?_extra=count: json { "rows": [{"id": 1}, {"id": 2}], "next": "2", "next_url": "/db/table?_next=2", "count": 31 }

The core issue that I keep reconsidering is whether the default .json representation should be an object or a list.

Arguments in favour of a list:

• It's what I always want. Almost all of the code that I've written against the API myself uses ?_shape=array.
• It's really easy to use. You can pipe it to e.g. sqlite-utils insert, you can load it into JavaScript without thinking about it.

Arguments against:

• Nowhere to put pagination or total counts. I added pagination to the link: HTTP header in #1014 (inspired by the WordPress and GitHub APIs) but I haven't solved for total count, and there's other stuff that's useful like "truncated": true to indicate that more than 1000 results were returned and they were truncated.
• An array is inherently non-extensible: if the root item is an object it's easy to add new features to it in a backwards-compatible way in the future. An array is a fixed format.

But maybe that last point is a positive? It ensures the default .json format remains completely predictable forever.

If .json DID default to an array of objects, the ?_shape= argument could still be used to get back alternative formats.

Maybe .json?_extra=total changes the shape of that default to be this instead:

json { "rows": [{"id": 1}, {"id": 2}], "total": 104 }

The thing I care about most though is next_url. That could be provided like so:

.json?_extra=total&_extra=next - alternative syntax .json?_extra=total,next:

json { "rows": [{"id": 1}, {"id": 2}], "total": 104, "next": "2", "next_url": "/db/table.json?_extra=total&_extra=next&_next=2" } This is feeling a bit verbose for a common combination though.

Should that default also include "columns" as a list of strings? That would be duplicate data of the keys in the "rows" list of objects, and I've never found myself wanting it in my own code - so I'm going to say no.

Plan: release a new release of Datasette (probably 0.49) with the new JSON API design, but provide a plugin called something like datasette-api-0-48 which runs as ASGI wrapping middleware and internally rewrites incoming requests to e.g. /db/table.json to behave if they have the ?_extra= params on them necessary to produce the 0.48 version of the JSON.

Anyone who has built applications against 0.48 can install that plugin.

Would it be so bad if the default format had a "rows" key containing the array of rows? Maybe it wouldn't. The reason I always use ?_shape=array is because I want an array of objects, rather than an array of arrays that I have to match up again with their columns.

A default format that's an object rather than array also gives something for the ?_extra= parameter to add its extras to.

I'd like to revisit the idea of using ?_extra=x to opt-in to extra blocks of JSON, from #262

(I think I may have been over-thinking the details of this is for a couple of years now.)

I'm going to hack together a preview of this in a branch and deploy it somewhere so people can see what I've got planned. Much easier to evaluate a working prototype than static examples.

Another idea: the default output could be the list of dicts: json [ { "pk1": "a", "pk2": "a", "pk3": "a", "content": "a-a-a" }, ... ] BUT... I could include pagination information in the HTTP headers - as seen in the WordPress REST API or the GitHub API:

``` ~ % curl -s -i 'https://api.github.com/repos/simonw/datasette/commits' | head -n 40 HTTP/1.1 200 OK server: GitHub.com date: Fri, 11 Sep 2020 21:37:46 GMT content-type: application/json; charset=utf-8 status: 200 OK cache-control: public, max-age=60, s-maxage=60 vary: Accept, Accept-Encoding, Accept, X-Requested-With etag: W/"71c99379743513394e880c6306b66bf9" last-modified: Fri, 11 Sep 2020 21:32:54 GMT x-github-media-type: github.v3; format=json link: https://api.github.com/repositories/107914493/commits?page=2; rel="next", https://api.github.com/repositories/107914493/commits?page=44; rel="last" access-control-expose-headers: ETag, Link, Location, Retry-After, X-GitHub-OTP, X-RateLimit-Limit, X-RateLimit-Remaining, X-RateLimit-Used, X-RateLimit-Reset, X-OAuth-Scopes, X-Accepted-OAuth-Scopes, X-Poll-Interval, X-GitHub-Media-Type, Deprecation, Sunset access-control-allow-origin: * strict-transport-security: max-age=31536000; includeSubdomains; preload x-frame-options: deny x-content-type-options: nosniff x-xss-protection: 1; mode=block referrer-policy: origin-when-cross-origin, strict-origin-when-cross-origin content-security-policy: default-src 'none' X-Ratelimit-Limit: 60 X-Ratelimit-Remaining: 55 X-Ratelimit-Reset: 1599863850 X-Ratelimit-Used: 5 Accept-Ranges: bytes Content-Length: 118240 X-GitHub-Request-Id: EC76:0EAD:313F40:5291A4:5F5BEE37

[ { "sha": "d02f6151dae073135a22d0123e8abdc6cbef7c50", "node_id": "MDY6Q29tbWl0MTA3OTE0NDkzOmQwMmY2MTUxZGFlMDczMTM1YTIyZDAxMjNlOGFiZGM2Y2JlZjdjNTA=", "commit": { ``` Alternative shapes would provide the pagination information (and other extensions) in the JSON, e.g.:

/squirrels/squirrels.json?_shape=paginated json { "rows": [ { "pk1": "a", "pk2": "a", "pk3": "a", "content": "a-a-a" } ], "pagination": { "next": "234", "count": 442 } }

https://latest.datasette.io/fixtures/compound_three_primary_keys.json?_size=2&_shape=array returns this: json [ { "pk1": "a", "pk2": "a", "pk3": "a", "content": "a-a-a" }, { "pk1": "a", "pk2": "a", "pk3": "b", "content": "a-a-b" } ] There's one big problem with this format: it doesn't provide any space for pagination information.

Here's the default JSON at the moment: https://latest.datasette.io/fixtures/compound_three_primary_keys.json?_size=2

json { "database": "fixtures", "table": "compound_three_primary_keys", "is_view": false, "human_description_en": "", "rows": [ [ "a", "a", "a", "a-a-a" ], [ "a", "a", "b", "a-a-b" ] ], "truncated": false, "filtered_table_rows_count": 1001, "expanded_columns": [], "expandable_columns": [], "columns": [ "pk1", "pk2", "pk3", "content" ], "primary_keys": [ "pk1", "pk2", "pk3" ], "units": {}, "query": { "sql": "select pk1, pk2, pk3, content from compound_three_primary_keys order by pk1, pk2, pk3 limit 3", "params": {} }, "facet_results": {}, "suggested_facets": [ { "name": "pk1", "toggle_url": "http://latest.datasette.io/fixtures/compound_three_primary_keys.json?_size=2&_facet=pk1" }, { "name": "pk2", "toggle_url": "http://latest.datasette.io/fixtures/compound_three_primary_keys.json?_size=2&_facet=pk2" }, { "name": "pk3", "toggle_url": "http://latest.datasette.io/fixtures/compound_three_primary_keys.json?_size=2&_facet=pk3" } ], "next": "a,a,b", "next_url": "http://latest.datasette.io/fixtures/compound_three_primary_keys.json?_size=2&_next=a%2Ca%2Cb", "query_ms": 17.56119728088379, "source": "tests/fixtures.py", "source_url": "https://github.com/simonw/datasette/blob/master/tests/fixtures.py", "license": "Apache License 2.0", "license_url": "https://github.com/simonw/datasette/blob/master/LICENSE" } There's a lot of stuff in there. This increases the risk that future minor changes might break existing API consumers.

It returns rows as a list of lists of values, and expects you to correlate these with the list of columns. I originally designed it like this because I thought this was a more efficient representation than repeating the column names in a dictionary for every row. With hindsight this was a bad optimization - I always use ?shape=array because it's more convenient, and gzip encoding of the response means there's no bandwidth saving. Users who want that efficiency should request it using a custom ?_shape=.