SQLite Release 3.37.0
sqlite.org
sqlite.org
I would expect the additional checks on writes to be negligible in most cases (dwarfed by other bookkeeping and I/O). I imagine that strict mode could enable some optimizations for read queries, but I can't see those being huge wins.
I agree from an algorithmic perspective but in the real world, it really depends. The size of a non-text/binary cell within a column is now fixed, meaning guaranteed to never need to resize/reallocate when streaming results. That could translate to non-negligible improvements (multiple percentage point speed up) if it’s implemented separately from the existing code path.
I think this is not true since the on-disk format didn't change, ie, ints are still variable-size whether strict or not.
I wonder if SQLite didn't store the schema for a table as a "CREATE TABLE..." string, but instead something a bit more... structured, then such things could be added in a more backwards compatible way?
And yes, it'd be very nice to have a relational metaschema that doesn't suck for SQL schema. So far all such metaschemas I've seen leave a lot to be desired.
What would be nice is if SQLite had separate “compatible with” and “read-only compatible with” fields, since the read-only behavior should not have changed at all (although I suppose describing the table would fail depending on how the presence of the STRICT modifier is codified).
Something like an object-relational mapper, e.g. SQLAlchemy?
Some people seem to object to the Any type, but it's incredibly useful since a column may naturally contain a different type per row and it avoids ugly hacks to accommodate them.
> The behavior of ANY is slightly different in a STRICT table versus an ordinary non-strict table. In a STRICT table, a column of type ANY always preserves the data exactly as it is received. For an ordinary non-strict table, a column of type ANY will attempt to convert strings that look like numbers into a numeric value, and if successful will store the numeric value rather than the original string.
So it has some utility in a strict table that you can't get otherwise.
[1] https://www.sqlite.org/stricttables.html#the_any_datatype
As an embedded SQL engine I think this is the most rational way to use it, making it very predictable and very flexible.
As they say, if you can't beat 'em, join 'em!
Of course, migrating current databases is another story. They have plenty of CHECK constraints and app-level validation already so the value in migrating those is small (although I'd still do it because the ANY data type columns will not coerce / convert anything when the table is STRICT which to me is still a win).
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.
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.
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.
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.")
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."
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?
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.
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)
- 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.
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.
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.
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. :)
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.
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.
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.
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.
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.
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.
> What about TIMESTAMP? DATETIME?
I tried out DuckDB which is more my style but the .NET libs are poor and I can't shave that yak right now.
> In a STRICT table, a column of type ANY always preserves the data exactly as it is received. For an ordinary non-strict table, a column of type ANY will attempt to convert strings that look like numbers into a numeric value, and if successful will store the numeric value rather than the original string
There's still much less data types than in most SQL databases though, e.g. float4, boolean, timestamp etc are missing.
Is that part a real problem? (The NaN thing certainly is.)
Doubles are a strict superset of floats. You don't risk losing any precision on the float to double conversion, and if you then convert the result back to float, you won't lose any precision either.
Is it hard to work around? No. But every time I use SQLite it requires half a dozen little hacks to work around it's strangeness and I would prefer to not have to do that.
Anyway, I've never touched a serialization library that doesn't mess with the data types it gets. There is always some conversion needed, and this one is on the easy end of the scale. The fact that composed types (records and arrays) aren't on the list is a much larger troublemaker.
(Regardless, I'm curious about the applications of 13M inserts per minute on a synchronous single-computer database. Pushing a machine to that extreme looks very interesting. Is that application something you can share?)
It's not a particularly demanding use-case, especially with respect to querying. Could just use some serialization format, parquet maybe, but I know SQLite works reliably and has good libs. It's nice to have standard tools available, and I know no matter what changes I'll be able to read these databases in 20 years. Didn't have a lot of time to work on this project so I didn't spend much time evaluating alternatives.
> You still can't store a float in the database and get it back out unmodified. SQLite will expand it to a double,
> and it can't store NaNs.
I would like an even stricter mode with no automatic type coercion.
I'm chiding SQLite, a bit unfairly, that they only implemented the strict feature half right, still leaving some type coercion in.
I wonder if this will provide more adoption of this client server database.
EDIT: for this not aware, Expensify is a paid sponsor of SQLITE and there’s an official branch (Bedrock) of SQLite that has client / server capabilities.
I love this level of attention to detail and kindness in the SQLite documentation. Very few projects do this.
At last.
That's fair and allowed by the standard, but it will break some things.
An almost Clintonian phrase! :-)
If it's what I'm thinking of, this concern was carefully considered:
If a subquery in the FROM clause has an ORDER BY, that ORDER BY is omitted if all of the following conditions are also true:
1. There is no LIMIT clause in the subquery
2. The subquery was not one of the virtual subqueries added internally by SQLite for window-function processing
3. The subquery is not part of the FROM clause in an UPDATE-FROM statement
4. The outer query does not use any aggregate functions other than the built-in count(), min(), and/or max() functions.
5. Either the outer query has its own ORDER BY clause or else the subquery is one term of a join.
Personally i'm more helped by good examples than these syntax diagrams.
However, it does make handling more "advanced" types less standard. For example, there isn't a standard way to store a date, especially if we want to preserve high-precision (nanoseconds).
Why not just by convention as ints, nanoseconds UTC post epoch?
This also is an interesting tidbit:
"The query planner now omits ORDER BY clauses on subqueries and views if removing those clauses does not change the semantics of the query."
Curious how it might improve queries -- I wonder how many folks were doing potentially redundant order by's in subqueries
1: Start out lean and simple. Often replacing an old complex one
2: Become more complex over time
3: Be replaced by a new lean and simple one
If you can take a developer using the old version that was lean and simple, give them the latest version full of extra features, and they can use that version without any need to change anything, then I think it's fine.
And, so far, I think that SQLite has done a good job of keeping their core usage lean and simple.