Hacker News Reader

Comment by SPBS | Hacker News Reader

STRICT tables forbid the use of any other types besides INT/INTEGER, REAL, TEXT, BLOB (and ANY).

What about TIMESTAMP? DATETIME? You can't use that as a type name anymore. Either pick INT, REAL or TEXT for that. What about JSON? You can't use it as a type name anymore. Pick TEXT. Basically type names, even if SQLite didn't have explicit support for that type, were helpful documentation hints in what the column contained. Now everything devolves to either INT or TEXT. I have UUID fields, I can't even call them UUID or BINARY(16), just BLOB.

View on HN

I think a logical future step would be to add the SQL-99 CREATE DOMAIN statement. That allows you to define a custom type alias, for example:

CREATE DOMAIN UUID AS TEXT;

Or if you prefer:

CREATE DOMAIN UUID AS BINARY;

Without CREATE DOMAIN, for STRICT tables they don't know which basic type to use for an unknown type name. They could use something like "unknown types are equivalent to ANY" but that would mean they couldn't implement CREATE DOMAIN in the future without breaking backward compatibility. If they did implement CREATE DOMAIN I expect it would be supported for STRICT tables only.

What makes it really powerful is you can also define a CHECK constraint, NOT NULL constraint, and DEFAULTs to be automatically applied to table columns of this custom type. Examples:

CREATE DOMAIN UUID AS TEXT CHECK (value REGEXP '[a-f0-9]{8}(-[a-f0-9]{4}){4}[a-f0-9]{8}');

CREATE DOMAIN UUID AS BINARY CHECK (LENGTH(value) = 16);

I think Postgres is the only major relational database to implement CREATE DOMAIN, although there are some more obscure RDBMSes which support it too (such as Firebird, Interbase, Sybase/SAP SQL Anywhere).

The SQLite developers haven't said they plan to do this, but I imagine they are thinking about it and may do it at some point. I think it is the kind of feature likely to appeal to them, it should be relatively simple to implement (especially if they don't implement ALTER DOMAIN, only CREATE and DROP), but provide a lot of power in exchange.

came here to say this (approximately)

Presumably the point of the strict typing forbidding aliases here, is to force the developer to realize/acknowledge that SQLite does not in fact have separate internal native types with these names.

It's like the modern C convention of avoiding "int" in favor of u8/u16/u32/u64. Rather than pretending "int" is a separate thing, you have to explicitly choose and acknowledge the concrete primitive type you're really getting, and explicitly opt into the semantics that use of that concrete primitive type implies.

Not a great analogy because the semantics of the classic C types still leak through. For example, ((uint16_t)0xFFFF) * ((uint16_t)0xFFFF) would cause undefined behavior if int is 32 bits wide, due to implicit arithmetic promotions.

Also, many APIs of the C standard library are defined to take or give char/int/etc., and it takes effort to translate between those types and u8/u32/etc. in a safe way that works on all platforms.

That's not because "int is 32 bits wide", though; that's because you're on a machine whose native register size is more than 16 bits, and so your processor's MUL instruction likely expects to always write the product to a 32/64-bit destination register, even if the source registers are both 16-bit.

IMHO C is far too lax with things like this; an ideal systems language would force you to annotate maths operations with the "working precision" for each operation, making explicit any promotion done to the inputs before the math happens. And then, if your "working precision" exceeded the destination variable's size, you'd also be forced to do an explicit demoting cast, just as if you were assigning from a declared larger-sized variable.

Yes, it'd be mostly irrelevant for everyday int math stuff; but it'd make things so much clearer as soon as you started working with floats (where you'd have to explicitly opt into Intel's 80-bit extended "working precision.")

Rust requires explicit conversions for numeric types for exactly the reasons you mention.

Most systems languages (that aren't C) force you to be explicit about demoting casts between formal variables; that's the easy part. But no language I know of (not even Rust) gets rid of implicit promotion during math operations, in favor of 1. forcing you to specify the size of the destination register you want the instruction to use (i.e. the working precision), and then 2. requiring an explicit demoting cast if the destination formal variable is smaller than the working precision you specified.

I mean, C's inline assembler, and LLVM IR, both require this; but they also require you to specify which work register you're talking about, rather than just the size, which breaks the conceit of a "portable systems language."

Yup. Most newcomers to SQLite are bitten hard by the fake datetime support.

  > What about TIMESTAMP? DATETIME?

The same thing as we've always done: proper column names and strong datatypes in the application layer.

Yeah, I'm starting to warm to the idea of `CREATE TABLE tbl ( created_at INT ) STRICT;`. The name still hints at what domain type it contains, and enforcing INT timestamps should make any indexes on that column considerably smaller.

> What about TIMESTAMP? DATETIME?

I usually use TEXT and store ISO 8601.

Datetime and timestamp is fine if you need date functions ("- interval '7' day", etc.) but most of the time you can just convert the text to datetime when running the query without much performance penalty in other databases.

What is the usecase for TIMESTAMP and DATETIME that cannot be solved today?

If you’re going to convert it anyway, why use a human readable format in the first place? If you store it as unix epoch, not only do you never have to worry about time zones, it’s also a sortable integer instead of text.

Regarding time and timezones, it's best to not prematurely optimise. Ideally store both. It's much more valuable to track the timezone from your source then to have to worry about it later. ISO8601 is the universal standard.

A case where you see this issue play out is the horrible Strava-*device sync. You track in a different timezone and they store+visualise activities against some weird client-side profile setting, which causes morning runs to render at 11pm. Only adding timezone later on. They totally mess this up.

First of all, I don't see how using unix epoch timestamps can be called "premature optimization", it's a pretty widely used and standardized way of saving timestamps.

Secondly, if you're in "SQLite Strict Land" and you don't have access to abstractions like "tell the database this is a date", then the best way of storing timestamps is unix epoch, I would be extremely surprised if databases don't already do this behind the scenes when abstracting away things like dates for the user.

Thirdly, what is the value of tracking "source timezone"? This is solving a non-existent problem: if you are getting the timestamp in unix epoch from the source, and you're storing it in unix epoch, the "source timezone" is already known: it's UTC just like all unix timestamps are.

Timezones is fundamentally a data presentation concern, and I strongly believe they should not be a part of the source data.

> You track in a different timezone and they store+visualise activities against some weird client-side profile setting, which causes morning runs to render at 11pm. Only adding timezone later on. They totally mess this up.

This is exactly the type of issue that happens because you involve timezones in your source data. If all applications and databases only concern themselves with unix timestamps, and the conversion to a specific timezone only happens in the application layer upon display, this type of issue simply does not happen, because time "1638097466" when I'm writing this is the exact same time everywhere on the globe.

(Of course, similar issues can happen due to user error if two applications have different time zone settings and the user mistakenly enter a timestamp in the wrong TZ, but that's definitely not solved by making time zones a part of the data itself)

I agree that this is the preferred way of dealing with it. Unfortunately, it's not always possible. Some cases that come to mind:

- Importing event/action data that contains date/time values with a timezone but insufficient information on the place where the event occurred. Converting to UTC and throwing away the timezone means you're losing information. You can no longer recover the local time of the event. You can no longer answer questions like, did this happen in broad daylight or in the dark?

- Importing local date/time data without a timezone or place (I've seen lots of financial records like this). In this case, you simply don't know where on the timeline the event actually took place. The best you can do is store the local date/time info as is. You can't even sort it with date/time data from other sources. It's bad data but throwing it away or incorrectly converting it to UTC may be worse.

- Calendar/reminder/alarm clock kind of apps. You don't want to set your alarm to 7 am, travel to another timezone and have the alarm blare at you at 4 am. Sometimes you really do want local time.

- There are other cases where local times are not strictly necessary but so much more convenient. Think shop opening hours in a mapping app for instance. You don't want to update all opening hours every time something timezone related happens, such as the beginning or end of daylight saving time.

You are correct, there are many other reasons for saving dates and times in a database apart from recording "events", and for those it does make sense to use "relative" descriptions of time or date. I'd argue though, that if your data collection of events is imperfect, like if you have no idea which system an event came from or whether you can trust the syntax of the timestamps, those are primary problems that should be fixed and not "worked around" by changing how your timestamps are saved.

For example, if you don't know where a data point originated, that's already a pretty big issue regardless of whether the time is correct or not. If you have financial data with ambiguous timestamps, this is not only a problem but potentially a compliance problem, since banks are heavily regulated. I think it's unlikely that it's acceptable for a bank to be unable to answer the question "when did this event take place", so the fundamental issue should be fixed, not tolerated.

You're not always in a position to fix these issues, nor is it always worth the effort. It may be a complete side-issue that no one cares about. The information may exist somewhere but not in the dataset you're working with.

Having done quite a bit of data integration work in my life, I can tell you that fixing things can range from the unavoidable to the quixotic.

Another use case is the definition of a business day.

A business day is almost always in local time. Example: business is open from 9AM until 6PM. This is affected by day light saving and it moves back and forth in UTC.

Working with financial systems, payments are business day aligned and if you try to have a system with UTC time you might be surprised what sort of issues are surface. I learned these things the hard way. :)

That only applies if what you're storing identifies a point in time. In business code we often have to work with date times in a specific timezone. e.g. if you schedule a meeting on 2022-07-01 09:00 in Berlin time, throwing away the timezone would be a mistake which results in incorrect behaviour when the DST rules change.

> That only applies if what you're storing identifies a point in time.

A meeting is definitely "a point in time", and your example illustrates my point: the display of "2022-07-01 09:00 UTC+2" should be up to the client (which has a timezone setting) but the storage of this meeting's time should, in my opinion, not be "2022-07-01 09:00 UTC+2" but the timezone-neutral "1656658800".

Since there is technically no such thing as "Berlin time", it should be up to the client to decide whether the date you selected in the future (July 1, 2022) is UTC+1 (winter) or UTC+2 (summer) and then store the time in a neutral way to ensure it's then displayed correctly in other clients that have another timezone set (for example the crazy Indian UTC+5:30).

Meetings are a good example of when this matters because it's important that meetings happen at the same time for everyone, which will always happen when the source data is unix epoch. Of course it also happens with correct time zone conversions, but to me that adds unnecessary complexity. Other examples given, like the opening hours of a store, are probably better to store in a "relative" way (09:00, Europe/Berlin) as the second part of that tuple will vary between UTC+1 and UTC+2.

Great question!

In my experience human readability is pretty important when you are debugging or working with SQL in a raw form (running random, ad-hoc queries). Storing unix epoch is fine and sometimes I do it but more recently I just realised that unless I am working with a database with billions of rows storing a human readable text is fine.

> In my experience human readability is pretty important when you are debugging or working with SQL in a raw form (running random, ad-hoc queries).

I agree, but most databases have functions for this. MySQL example:

> SELECT FROM_UNIXTIME(1196440219);

-> '2007-11-30 10:30:19'

I would claim this reinforces the benefit of unix timestamps because now you're getting it back in your local time (or whatever you choose to convert it to: SET time_zone), not whatever time it happened to be put into the database as.

For MySQL there's a more important point though:

* Storing timestamps in MySQL as pure text is simply wrong since MySQL has abstractions for timestamps (like the TIMESTAMP type)[1] and not using this is a completely unnecessary violation of good practice

* In the case we're discussing (no abstractions available, INT only) I would still say that unix timestamps brings the rather huge benefit of ensuring that all data is put in correctly: there is no way to sanitize inputs with a string column and ensure that the same timezone is always used, at least not without a bunch of extra an unnecessary code.

[1]: https://dev.mysql.com/doc/refman/8.0/en/datetime.html

You really don't want to store timestamps as text.

It's horrible for performance. A iso8601 timestamp takes at least double the amount of bytes (without a timezone), but more importantly you need to parse and validate the timestamp each time you use it. Filtering and sorting suffer a lot. Iso8601 without timezones can in theory be sorted trivially without parsing, but sorting strings is still a lot more expensive than sorting arrays of int64. And you also have to make sure to never write a different timezone. Why not just use an int?

You then have to parse the string into the native datetime type again in application code.

You also need to add a custom check constraint to prevent writing garbage to the column, which is easy to forget.

SQLite is not postgres. Often sqlite dbs span only a couple of MB, ie when used in an App, and performance might not be as important as readability during debugging/development.

The argument "this doesn't matter because all SQLite databases are small anyway" does not hold water. There are many large SQLite databases out there. And even if there wasn't, storing timestamps as text is simply and fundamentally A Bad Idea(tm).

True, but I think SQLite has a design goal of being an embedded DB.

As for your argument to store timestamps as text, look at DynamoDB who only allows storing dates as text, and it is designed to handle petabytes of data.

And more importantly, JSON.

Note that you don't have to have a JSON-typed column in SQLite in order to do JSON-y things. In SQLite, JSON is just "text that happens to contain JSON", and JSON functions are just functions that operate on "text that happens to contain JSON".

Reply on news.ycombinator.com