In Postgres there is a family of INSTEAD OF triggers that allow you to do all this lensing within the database so that the application can have multiple views of the data that all resolve to a single normalized structure internally but present in a variety of ways. To my googling knowledge SQLite also supports INSTEAD OF triggers - is there a benefit to using the predefined JSON columns instead of writing the data into an actual normalized structure and then zipping it up at read time or during the regeneration of a materialized view?

Using `INSTEAD OF` or writing to a view implies that the SQL schema was designed with foreknowledge of the data that will be stored in the schema. These tools let you lens from one known, totally-specified schema (a "view") to another known schema (a "table") at write time. In contrast, the catch-all JSON column strategy allows us to store data the schema has no idea about -- or may never need to understand.

You're right that once we move a property from the "don't query, just store" category to the "query and/or update locally" category, we could use views/INSTEAD OF trigger to keep the JSON column and the normalized view in sync.

In the example given in the article foreknowledge of the property does exist - there's a specific DDL statement to run before the column-y thing is usable.

With the advent of schemaless I think there's been far too strong of a reactionary movement against schema adjustment. With some trivial investment in migration management infrastructure you can arrive at a point where changing the DB (even in drastic ways) is as easy as committing a new file to the code base. At my company we do make use of postgres JSON types for some limited fields related to logging and auditing where live queries never need to be executed - where the data is essentially serving as WOM - we've also got once instance where our DB is essentially a pass-thru between two services that don't talk to each other directly, for this usage JSON works great since those two services just coordinate on the expected format. In most other cases we'll just write quick migrations to expand our data definition as needed to fit whatever we need to put into it.

Relational databases are both powerful and quite flexible - so long as you don't enshrine the schema to an extent where it feels immutable.

It’s easier to evolve a schema to be strictly normalized when you have only a few (<1000) SQL databases running in an environment you control, especially if you also control the deployment of code that talks to those databases. In that kind of setting you can easily manage any kind of migration, and you know when a migration is fully rolled out and you can drop backwards compatibility.

In Notion’s SQLite use-case, we have many millions of instances of this schema running on hardware owned by our users, and those users ultimately control when our app updates & runs migrations on those devices. If I commit a new migration today, it’ll start rolling out tomorrow, but it will take weeks to saturate the install base of our app, and we can never be 100% sure all existing clients have upgraded.

Having JSON catch-all columns means we can evolve our API data types in parallel with that slow SQL schema rollout, instead of needing a strict order of “migrate first, then start writing” for any new bit of data we want to render & cache on the client.

I think that for Notion (neat company btw) it might make a lot of sense to keep the schema pretty loose but I have managed to keep an agile DB alive and well with a lot of flexibility. A lot of migration infrastructure packages and tools will store the migration state in the DB itself - we're using Phinx[1] internally which creates a `phinxlog` table with a record of migrations that have been run - there is tooling to migrate only up to a specific version and since the record of all executed versions is stored in the DB the tool can easily figure out what works needs to be done in which circumstances. The result is that we can easily roll different environments onto different versions.

Neither schema driven nor unstructured is always the right call - they both have their strengths and weaknesses - but I think that trust in data integrity is pretty important when writing flexible code on top of a data layer. Knowing that expected keys can't be omitted and that so-and-so column must conform to a given data domain can really alleviate defensive coding costs.

If you'd like to talk some more I can shoot you an email and we can sit down sometime?

1. https://phinx.org/

If it is a cache, is it really beneficial to store that data (wasting space) rather than query it again after migration (wasting bandwidth)?

If a lot of my cached data is never used again and waiting to be replaced by e.g. LRU strategy, storing all those fields that can't even be used in the current version seems like a huge waste of space.

It seems that it would only be a good idea if most of the fields of the JSON objects are extracted and used, and most of the cached objects are used again. Which might be the case at Notion (where the cached data is probably pages) but probably not universally.

There are some object properties the clients need to query, and other object properties the clients don’t care to query, but are needed to render the data in the UI, so it still needs to be stored locally.

For example, today we run queries on "parent page ID", so that is a normalized parent_id column in SQLite, instead of being stored in the JSON column.

We don't currently run queries on "text color" or "image width" object properties, but we still need to store that data in the cache to render a block. So, it's fine to leave those things in a JSON column. If we need to run a query like "find all images that are full-width" in the future, we can add a new image_width column/virtual column/index at that time.

Ah gotcha -- thanks!

You can choose whether to store the values or to compute them every time in SQLite

But in terms of the overall performance budget, for a mobile app, I don't think the choice matters much. JSON in SQLite is fast and even in the case of "storing every generated column with a lot of user data", I imagine the app binary is going to be larger than the database it produces.

I'm talking about a storage space/bandwidth tradeoff here, I don't know what you are saying about compute.

+1 for lenses. A FP concept.