The topic of MySQL utf8 vs utf8mb4 was discussed in depth just six days ago on HN: https://news.ycombinator.com/item?id=29793916
In fact, when the log4j thing came up, we audited all our existing COTS/internal and virtually all of it was still on the "EOL" log4j 1.x.
Since you're familiar with this. What happens with an existing upgrade - is strict mode enabled by a typical distro package upgrade or does it only impact new installs?
> What happens with an existing upgrade
sql_mode is a setting, so it depends on whether the default is being overridden: e.g. does the my.cnf config file specify a value explicitly? Does your application's connection logic set a session-level override for each connection? DBaaS vendors add an additional wrinkle since they tend to have custom UIs and APIs for global settings.
> typical distro package upgrade
Many Linux distros made it so that attempting to install "mysql" actually installs MariaDB instead, so when going this route, it's unfortunately always very hard to predict what you'll get or what will happen!
MariaDB is indeed pretty pleasant to use and for most use cases can be a drop in replacement for MySQL, should you not want to use MySQL, but also wouldn't want/need/be able to use PostgreSQL or something else: https://mariadb.com/kb/en/mariadb-vs-mysql-compatibility/
I might be in the minority here, but i actually rather enjoy MySQL/MariaDB, since MySQL Workbench is still the best DBA software that i've used, especially because of the ER modelling and schema sync functionality, and their approach to schemas/users within a DB feels more reasonable than that of Oracle/PostgreSQL.
That said, both PL/SQL and PL/pgSQL feel more powerful than whatever MySQL/MariaDB has. Also, in regards to MariaDB in particular, their versions have diverged from MySQL somewhat, so you might still want to look into the differences, if migrating over: https://mariadb.com/kb/en/mariadb-vs-mysql-compatibility/
As for predictability - nowadays i just run specific versions in containers (e.g. 10.7.1 or maybe 10.7 which corresponds to the patest PATCH release), to sidestep the distro differences altogether. It also lets me easily launch new instances, give resource limits to them, have multiple versions in parallel and mount the data directories wherever i need them, for easier backups etc.
I still bind mount directories though. Docker volumes seem really convenient but a few years back I dealt with helping out a coworker with some data loss (pre-volumes) in a container's filesystem so I have some trust issues with how they implement storage.
This mirrors my experience and despite how Docker seems to suggest that non-abstracted storage is a liability and we should use volumes instead in many cases, i disagree: https://docs.docker.com/storage/bind-mounts/ In my eyes, the current implementation for volumes is for data that i don't really care much about, but would still consider extracting/moving later and have it be persistent, like when trying out new software.
Personally, bind mounts also make it extremely easy to operate with whole installs of software - for example, should i want to upgrade something like Nextcloud (and have enough space), i can just stop the containers, make a copy at /docker/nextcloud_stack/data/nextcloud_mysql_old and run a new container version against /docker/nextcloud_stack/data/nextcloud_mysql and see if it all works.
If the update has corrupted everything, i can just delete the broken folder, rename my backup and proceed with restoring the older version. If i want to mess around with the broken version, i can do it with 0 worries about what else could break. It's amazing, way better than relying on the POSIX approach to a file system, where your software, its configuration and its persistent data is all over the place.
I once had to manage a Tomcat install which had the webapps, config, logs directories in their idiomatic locations rather than one folder, which was pretty horrible from a management perspective. I had to constantly jump around and wasn't entirely sure how it all fits together (probably lots of symlinks).
It's curious that some/many Kubernetes distros lock down the local storage driver, though - there's a preset data directory, where all of your mounted data ends up, which feels like a step backwards from a customization perspective. I had to change the data directory for the entire Kubernetes install to be able to store my data in a painfully obvious /app or /data mount point. Then again, their ability to limit the sizes of these persistent volumes is pretty cool.
Wouldn't it be a nice world where you can just declare something EOL and it disappears and somehow magically it's not a problem anymore for anyone end-to-end? Thing's just don't work that way. Like, suddenly, a product that was working perfectly fine for who knows how long suddenly needs to modified because some ass-hats thinks that EOL isn't suitable for production.
There are a number of reasons to upgrade a dependency or a product; an EOL label shouldn't be a part of the calculus. EOL may imply certain conditions that may influence the decision to upgrade but for the sake of EOL you can eat a bag of dicks. Think harder why you're proposing an upgrade is needed and how it'll add value to the business. EOL software is perfectly fine especially because it's been battle tested.
We also use log4j 1.x all over the place and I feel zero need to change that.
Or do you mean in the abstract, fairy-tale ending sense?
And "EOL" is a convenient shorthand for those conditions.
Most use cases care about never getting patches for security problems ever again.
I dunno, a widly known RCE seems a pretty good reason to take something out of production.
Honestly, in that case I tend to agree with GP. As long as you recognize EOL means you are in charge of supporting it and possibly backporting security fixes and you're okay with that - why not use it?
In many cases it is not, it is purely a 'ticking time bomb' - if new CVEs or requirements arise which cause issues in EOL versions, you may not even become aware of the vuln.
In counterbalance, new versions of code have new, unknown CVEs, new bugs, etc.
The issue is that of support, the fallacy and process issue is that 'current' support is of greater quality than 'historical' support, or that old software has a disproportionately higher support costs than 'current' software.
The true issue is overdependence on FOSS software and unwillingess to contribute - all software has costs, major vendors are 'coasting' on work provided by free volunteers and panic when those volunteers decide to no longer work for free.
If vendors were doing the legwork necessary to properly inhouse their dependencies, I would wager that much less software would end up discarded in 'fear', for a lack of a better word than this - modern development is supposedly built on features and patches, any codebase should be maintainable for a much longer period than its supposed EOL.
If the end user of the product or their needs don't matter, sure.
What happens if you upgrade from older version? Are defaults from previous version used, or is it default only on fresh installs?
There are similarly dumb misconfigurations that can theoretically be made in most hosted open source software.
Also, I gotta say, now I have to look into this strict mode and what else it breaks, when in this particular case a "don't silently truncate strings" flag would be preferable and probably lower impact, since "strict" vs "security flaw mitigation" is not quite the same thing (does such a more granular flag exist?). ... unless that's all the strict mode does.
Hm, and since you brought this up elsewhere. Is this "strict by default" in MariaDB - which is what most (linux) people are (probably) thinking of.
That happened in 2010 - I'm not sure 12 years counts as "finally happening", or just how much quicker you would like than 12 years ago?
Other than a time machine, I'm not sure what else you would like them to do.
Do you actually use MySQL or are you just complaining from a theoretical POV?
If they were really conservative, they could wait until the next major release to repoint the name like that. Haven't there been four major releases since 2003?
In older versions of MySQL, InnoDB indexes were limited to 767 bytes. Meanwhile, VARCHAR(N) can store N characters. So with utf8mb3, that's a max of 3N bytes worst-case; but with utf8mb4 ("proper" utf8), it's now 4N. This meant that VARCHAR(255) CHARACTER SET utf8mb4 could not be indexed (in its entirety) in these older versions; the CREATE TABLE would fail if you try.
When you have a massive installed base of users, and those users have lots of dumps / install scripts containing "VARCHAR(255) CHARACTER SET utf8", you simply cannot change the meaning of "utf8" in a minor release.
I fully agree that MySQL could have made this transition faster/earlier than they did, but I disagree that it could be done trivially or terribly quickly.
I believe the limitation you mention was fixed in MySQL 5.7, which would have been a reasonable opportunity to make "utf8" live up to its name.
Anyway, I fully agree that the utf8 alias swap should have been done in MySQL 5.7, or perhaps even 5.6. I'm just disagreeing with your assertions that this would have been a trivial change.
Nor do I think it should have been a "stop the world" high-priority change prior to emoji existing in unicode (~late 2010). Non-BMP characters were very rare before that point. Hard to justify a business need for a complex migration just to store archaic hieroglyphs. For another point of comparison, MS SQL Server only added support for non-BMP characters in 2012, and only added support for utf8 encoding in 2019!
https://blog.koehntopp.info/2022/01/12/utf8mb4.html#so-what-...
I definitely agree!
> now I have to look into this strict mode and what else it breaks
It's a mode to ensure nothing happens silently: illegal data type conversions, attempting to insert invalid values for the data type (too large, too long, invalid characters, etc), attempting to do an INSERT which doesn't specify some columns that lack defaults, etc. With strict mode enabled, all of these things throw hard errors, like in most other relational databases. Generally, it covers the entire class of things that are typically the most common complaints about MySQL on HN.
Interesting wording re: "breaks" -- for practical purposes this is a real concern for existing applications, which may be doing bad things which previously didn't throw errors with strict mode off. Do you want your DB to suddenly fail loudly for these things (strict mode), or do you want your existing application to continue working in these cases? Can't have it both ways.
> Is this "strict by default" in MariaDB - which is what most (linux) people are (probably) thinking of.
MariaDB made it default in 10.2, ~5 years ago.
I don't agree that most Linux people conflate MySQL and MariaDB though. At large US tech companies that use MySQL, they're almost all literally using Oracle MySQL, or a patch-set like Percona Server which directly tracks MySQL, or an internal patch-set that does the same.
In recent years, MySQL and MariaDB have diverged a bit on feature set and focus; each has some interesting features that the other lacks. They are not drop-in replacements for each other.
It seems to be that there's a world of difference between "silently truncate string" and "type coercion" even if I readily agree both are sources of security vulnerabilities.
Thanks for the replies though, even though I clearly annoyed at least 2 people due to the downvotes.
And for sake of comparison -- if I understand correctly, prior to just 2 months ago, sqlite didn't even have the option for strict type enforcement [1]. Do you judge sqlite as harshly, or think its creator is out of touch with the database world?
Sqlite is more of an embedded database, but if it silently truncated character data under some conditions that would make it equally unusable. Better to have no type enforcement than defective type enforcement.
Regarding "silent", MySQL at least emits a warning when this happens. Still not great / obviously wrong in a modern context, but just for clarity's sake it wasn't completely silent.
As for "defective type enforcement", that's a subjective opinion. In C, atoi("123hello456") returns 123 -- once you hit an invalid character, everything from that point is truncated. Is C atoi defective? Perhaps, but it's an opinion.
With strict mode disabled, four-byte characters worked the same way in MySQL utf8mb3: four-byte characters are invalid in this data type, and everything gets truncated at the first invalid character. Would simply stripping the invalid characters, but keeping valid chars after them, necessarily be more "correct"? No, it's an implementation decision, especially considering that emoji did not exist at the time and barely any non-BMP (4-byte) utf8 characters were used anywhere.
atoi is deficient by modern standards, but at least it clearly documents what the behavior is and always has. It doesn't truncate anything. For a database, truncating data is unforgivable. That is what errors are for, so your database doesn't act as a black hole where your data may silently disappear forever.
And yes, if you can't return an error for some reason, you should preserve as much data as possible. Throwing away hundreds or thousands of characters after an initial anomaly is insane. But a database should never do that anyway, it should either store what you give it, or return an error. Same with a filesystem or any other trustworthy means of data storage. Sorry we were too lazy to tell you we threw your data away isn't acceptable, nor is relying on anyone to check for warnings that should be hard errors.
A QA walks into a bar and orders "123hello456" beers. The correct answer is not for the barkeep to give them 123 beers, its to ask "I beg your pardon?" So the truncation in general is the problem. If the operation fails - too many bytes, wrong encoding, string too long - the operation should fail. It's not the db's job, that is the business logic's job.
The world is moving towards strict typing, and since this is what my brain is wired for I'm very happy.