Modern SQLite: Features You Didn't Know It Had
slicker.me
slicker.me
I don't think I've ever done much with SQLite's JSON functions, but I have on one or two occasions used a constraint to enforce a TEXT column contains valid JSON, which would have been very tedious to do otherwise.
I very, very much did. I was using a Python package that used a lot of NumPy internally, and sometimes its return values would be Python integers, and sometimes they'd be NumPy integers.
The Python integers would get written to SQLite as SQLite integers. The NumPy integers would get written to SQLite as SQLite binary blobs. Preventing you from doing simple things like even comparing for equal values.
Setting to STRICT caused an error whenever my code tried to insert a binary blob into an integer column, so I knew where in the code I needed to explicitly convert the values to Python integers when necessary.
On the STRICT mode, I've asked this elsewhere and never gotten an answer: does anyone have a loose-typing example application where SQLite's non-strict, different-type-allowed-for-each-row has been a big benefit? I love the simplicity of SQLite's small number of column types, but the any-type-allowed-anywhere design always seemed a little strange.
This flexibility is convenient when only one application reads and writes to the table. But if multiple applications access the same tables, the lack of a strictly enforced schema becomes a liability. The same is true when using generic tools to process data in SQLite tables, because such tools don't know what type of data to expect. The column type may be X but the actual data may be of type Y.
Sqlite came from TCL which is all strings. https://www.tcl-lang.org/
An example of where this would be a benefit is if you stored date/times in different formats (changing as an app evolved.)
I copied the idea for the ->> operator from PostgreSQL. But in PostgreSQL, the ->> operator always returns a text rendering of the value from the JSON, even if the value is really an integer or floating point number. PG is rigidly typed, so that's all it can do. But SQLite is flexibly typed, so the ->> operator can return anything - text, integer, floating-point, NULL - whatever value if finds in the JSON.
Ideally, I'd like to be able to specify the stored type (or at least, side step numeric affinity), and give the type a name (for introspection, documentation).
Specifying that a column is a DATETIME, a JSON, or a DECIMAL is useful, IMO.
Alas, neither STRICT nor non-STRICT tables allow this.
My preference would be for SQLite to actually support commonplace data types. But as long as it doesn't, I can see the appeal in using the schema to specify what data you're storing in your database.
I think the naive approach is to tokenize the input and append "*" for prefix matching. I'm not too experienced and this can probably be improved a lot. There are many settings like different tokenizers, stemming, etc. Additionally, a lot can be built on top like weighting, boosting exact matches, etc.
Does anyone know good resources for this to learn and draw inspiration from?
O'Reilly's "Relevant Search" isn't the worst here, but you'll be porting/writing a bit yourself.
Is there a reason why something more custom built, like ParadeDB Community edition won't meet your needs?
I understand you're speaking about SQLite, while ParadeDB is PostgreSQL but as you know, it's non-trivial to get good search quality, so I'm trying to understand your situation and needs.
They recently landed multi-writer support for their rust SQLite re-implementation, which is personally the biggest issue I've had with using SQLite for high concurrency applications.
`PRAGMA journal_mode = 'mvcc';`
https://docs.turso.tech/tursodb/concurrent-writes
Very excited to see if SQLite responds by adding native support, I'm hoping competition here will spur improvements on both sides.
Curious if others here are running SQLite in production and if you would be interested in something like this.
SELECT
settings -> '$.languages' languages
FROM
user_settings
WHERE
settings ->> '$.languages' LIKE '%"en"%';
Granted it was small DB (few megabytes), I also wanted to avoid collecting changes one by one, I simply wanted a diff over last time.
And ON CONFLICT which can help dedupe among other things in a simple and performant way.
with thing_key as (
insert into item(key, description) values('thing', 'a thing') on conflict do nothing )
insert into user_note(uid, key, note) values (123, 'thing', 'I like this thing') on conflict (uid, thing) do update set note = 'I like this thing');Does anyone have any other favorite modern bloom-filter-based search solutions that dont need to store copies of all the documents in the search db? Ideally something that can run in WASM too so we can ship a tiny search index to the browser. I found https://github.com/tinysearch/tinysearch but haven't tried it yet.
Typically one would design their "DTO"s to catch such errors right in the application layer, way before it even made it into the DB.
Different people call this serialization-deserialization layer different names (DTO being the most ubiquitous I think) but in general, one programs the serialization-deserialization layer to catch structural issues (age is "O" instead of 0).
The DB is then delegated to catch unchanging ground truths and domain consistency issues (people who have a DOB into the future are signing up or humans who don't have any email or address at all when the business needs it).
In your case, it's great the DB caught gaps in the application layer but one would handle it way before it even made it there.
The way I think about DB types are:
1. unexpected gaps/bugs: "What did I miss in my application layer?"
2. expected and unchanging constraints: "What are some unchanging ground truth constraints in my business?", "What are the absolute physical limits of my data?" - while I check for a negative age in my DTOs to provide fast errors to the user, I put these constraints in the DB because it's an unchanging rule of reality.
Crucially, by keeping volatile business rules out of the database and restricting it only to these ground truths, I avoid being dragged down by constant DB migrations in a fast-evolving business
https://sqlite.org/json1.html#table_valued_functions_for_par...
I did this recently for a fork of the main MiSTer executable because of a few disagreements with how Sorg runs the project, and it was actually pretty easy to change out the save file features to use SQLite, and now things are a little more resistant to crashes and sudden power loss than you'd get with the terrible raw-dogged writing that it was doing before.
It's fast, well documented, and easy to use, and yeah it has a lot more features than people realize.
I did not know SQLite allows writing data that does not match the column type. Yuck. Now I need to review anything I built and fix it.
I understand why they wouldn’t, but STRICT should be the default.
Why is it a problem that it allows data that does not match the column type? SQLite is intended for embedded databases, where only your application reads and writes from the tables. In this scenario, as long as you write data that matches the column's data type, data in the table does match the column type.
I'm afraid people forget that SQLite is (or was?) designed to be a superior `open()` replacement.
It's great that modern SQLite has all these nice features, but if Dr. Hipp was reading this thread, I would assume he would be having very mixed feelings about the ways people mention using SQLite here.
I do remain perplexed, though, about how people continue to think that rigid typing helps reliability in a scripting language (like SQL or JSON) where all values are subclasses of a single superclass. I have never seen that in my own practice. I don't know of any objective research that supports the idea that rigid typing is helpful in that context. Maybe I missed something...
I don't understand this. By values do you mean a row (in database terms)? I don't understand what that has to do with rigid typing.
Lack of rigid typing has two issues, in my opinion: First, when two or more applications have to read data from a single database, lack of an agreed-upon-and-enforced schema is a limitation. Second, when you use generic tools to process data, the tools have no idea what type of data to expect in a column, if they can't rely on the table schema.
So, in case you missed it, you're responding to Dr. Hipp himself :)
> I don't understand what that has to do with rigid typing.
Now I would like to learn a bit from Dr. Hipp himself, so here's my take on it:
Scripting languages (like my fav, Python) have duck or dynamic typing (a variation of what I believe Dr. Hipp, you specifically call manifest typing). Dr. Hipp's take is that the datatype of a value is associated with the value itself, not with the container that holds it (the "column"). (I must say I chose the word "container" here to jive with Dr. Hipp's manifest. Curious whether he chose that word for typing for the same reason! )
- In Python, everything is fundamentally a `PyObject`.
- In SQLite, every piece of data is (or was?) stored internally as a `sqlite3_value` struct.
As a result, a stack that uses Python and SQLite is extremely dynamic and if implemented correctly, is agnostic of a strict type - it doesn't actually care. The only time it blows up is if the consumer has a bug and fails to account for it.
Hence, because this possibility exists, and that no objective research has proven strict typing improves reliability in scripting environments, it's entirely possible our love for strict types is just mental gymnastics that could also have been addressed, equally well, without strict typing.
I can reattempt the "HN conjure" on Wes McKinney and see if this was a similar reason he had to compromise on dynamic typing (NumPy enforces static typing) to Pandas 1.x df because, as both of them are likely to say, real datasets of significant size rarely have all "valid" data. This allows Pandas to handle invalid and missing fields precisely because of this design (even if it affects performance)
A good dynamic design should work with both ("valid" and "invalid") present. For example: layer additional "views" on top of the "real life" database that enforce your business rules while you still get to keep all the real world, messy data.
OTOH, if you dont like that design but must absolutely need strict types, use Rust/C++/PostgreSQL/Arrow, etc. They are built from the ground up on strict types.
With this in mind, if you still want to delve into the "Lack of rigid typing has two issues" portion, I am very happy to engage (and hope Dr. Hipp addresses it so I learn and improve!)
The real world is noisy, has surprises in store for us and as much as engineers like us would like to say we understand it, we don't! So instead of being so cocksure about things, we should instead be humble, acknowledge our ignorance and build resilient, well engineered software.
Again, Dr. Hipp, Thank you for chiming in and I would be much obliged to learn more from you.
Relaxing that guarantee has benefits. For example, it can make application evolution easier--being able to store strings in a column originally intended for numbers is convenient. But that convenience can become a liability when multiple applications read from and write to the same database. In those cases, you want applications to adhere to a shared schema contract, and the RDBMS is typically expected to enforce that contract.
It also creates problems for generic tools such as reporting systems, which rely on stable data types--for example, to determine whether a column can be aggregated or how it should be formatted for display.
He is.
“Developers should program it right” is less effective than a system that ensures it must be done right.
Read the comments in this thread for examples of subtle bugs described by developers.
You're right, of course. But this must be balanced with the fact that applications evolve, and often need to change the type of data they store. How would you manage that if this is an iOS app? If SQLite didn't allow you to store a different type of value than the column type, you would have to create a new table and migrate data to a new table. Or create a new column and abandon the old column. Your app updates will appear to not be smooth to users. So it is a tradeoff. The choice SQLite made is pragmatic, even if it makes some of us that are used to the guarantees offered by traditional RDBMSs queasy.
No wait, what do you mean?
As I mentioned at https://news.ycombinator.com/item?id=47619982 - your application layer should be validating the data on its way in and out. I mention the two reasons I use for DB fall back
I dare say you will be hard pressed to find a dataset of significant size that doesn't have at least one invalid entry somewhere. Increasingly strict type rules will not fix that.
> I dare say you will be hard pressed to find a dataset of significant size that doesn't have at least one invalid entry somewhere
I agree. In my experience, and you've forgotten more than I have learned, the mark of a good data engineer is how they account for invalid entries or whether they get `/dev/null`ed.
> Checking the datatype is not the same as validating. There is lots of data out there that is invalid, and yet still has the correct type. In fact, that is the common case. Increasingly strict type rules will not fix that.
I am having a hard time letting go of this opportunity to learn from you, so in case you have time and read this again - When you say "There is lots of data out there that is invalid, and yet still has the correct type", I read "type" as "shape" or "memory layout" and "invalid" as "semantically wrong".
So, is a good example of this a value of `-1` for a person's age? The database sees a perfectly valid integer (the correct shape), but the business logic knows a person cannot be negative one years old (semantically invalid).
In that case, to be explicit, `0` or `14` is a "valid type" for an age (usually integer), but completely invalid data if it's sitting in an invoicing application for an adult-only business?
Again, thank you for your time and attention, these interactions are very valuable.
PS: I'm reminded of a friend complaining that their perfectly valid email would keep getting rejected by a certain bank. It was likely a regex they were using to validate email was incomplete and would refuse their perfectly valid email address
Somewhat ironically[^1], if STRICT were suddenly made the default, countless applications which work today would fail to work right after that update. SQLite is on billions upon billions of devices, frequently with many installations on any given device, and even a 0.001% regression rate adds up to many, many clients.
One of the reasons people trust SQLite is because, as a project policy, it does not pull the rug out from under them by changing long-held defaults.
[^1]: it's ironic because proponents of Strict tables frequently say that it improves app stability and robustness, whereas activation of Strict tables in apps not designed to handle it will, in fact, make them more fragile. Spoiler alert: most SQLite apps are not designed for Strict tables.
"C:\Windows\System32\winsqlite3.dll"
"C:\Program Files (x86)\Windows Kits\10\Include\10.0.27975.0\um\winsqlite\winsqlite3.h"
"C:\Program Files (x86)\Windows Kits\10\Include\10.0.27975.0\um\winsqlite\winsqlite3ext.h"
However, I will concede, and the article doesn't mention at all, far less are aware that you can build HA, cross region replicated SQLite using purely OSS software provided you architect your software around it. Now that would be a really good `Modern SQLite: Features You Didn't Know It Had` article!
Another interesting discussion point is how far self hosted PostgreSQL and pgBackRest can get you to a near-zero data loss high RPO, RTO setup. Its simply amazing we can self host all this.
All of the listed features except for strict tables and generated columns have been in SQLite for 10+ years, and those two are certainly not new. The JSON APIs were not made part of the standard distribution until 3.38 (2022-02) but were added in 3.9 (2015-10) and widely used long before they were upgraded from an optional extension to a core feature.
- Generated columns: 3.31 (2020-01)
- Strict tables: 3.37 (2021-11)
Correct.
As I mentioned in my GP, these features have been around a long, long, long time and in the current age of AI that would happily tell you these features exist if you remotely even hint at it, I would assume one would really have to go out of their way to be ignorant of them.
It doesn't hurt to have an article like this reiterate all that information though, I just would have loved the same level of effort put into something that's not as easily available yet.
A gap exists, and the article doesn't mention at all, about actually configuring solutions like litestream, rqlite, dqlite (not the same!) to build HA, cross region replicated SQLite using purely OSS software provided you architect your software around it. Now that would be a really good "Modern SQLite: Features You Didn't Know It Had" article!
Another interesting discussion point is how far self hosted PostgreSQL and pgBackRest can get you to a near-zero data loss high RPO, RTO setup. It just blows my mind that we can self host all this on consumer grade hardware without having to touch a DC or "the cloud" at all.
For bool, it could just be an alias of a numeric type. Something equivalent to number check(col = 0 or col = 1) would be perfectly fine.
Date/time handling is pretty weak. Having to store values as GMT text is just inconvenient.
When retrieving values in Node.js, I ended up using new Date(val), which caused the classic bug where a GMT-stored value gets interpreted in the local timezone.
The correct approach was new Date(val + ".Z"), but I really don’t want to deal with that kind of hassle anymore.