The author is comparing "off-the-shelf" caching with custom caching. They're coming from the assumption that you must be caching somehow and arguing that the word "caching" should be understood to mean only particular approaches to the general idea of caching. And obviously the whole point of the general idea is to optimize things.
It's a rhetorical mess
The joke form of this quote goes along the lines of:
There are only two hard things in Computer Science: cache
invalidation, naming things, and off-by-one errors.
there's two hard problems in computer science: we only have one joke and it's not funny.
Apparently⁰ by Philip Scott Bowden¹
> I don't see what's so hard about DNS, it's just cache invalidation and naming things.
There are three hard problems in Computer Science:
1) Cache invalidation
2) Naming th3) Concurings
rency
4) Off-by-one errors
Likewise, naming things is simple as long as you alone, or a in a small team. But as soon as there are multiple organizations with all their own traditions, it gets tricky. Just witness the eternal flame wars about camelCase, PascalCase, snake_case, kebab-case, and UPPER_CASE. It is almost as hopeless culture clash as Emacs vs Vi vs PowerPoint...
(I leave the off-by-one errors as an exercise for the reader)
- The language dimension - choice of words, that are good enough for the purpose, and not confusing. For example, "Manager" is as ambiguous as it gets, it can mean many thing, except we've been using it long enough that there's a more specific shape of meaning[0] for that word in code/program architecture contexts - so you still would use it instead of, say "Coordinator", which would raise all kinds of questions that "Manager" no longer does.
- The epistemological dimension - whether the word you chose correctly names the concept you meant, and whether the concept you meant is actually the right one to describe the thing you're trying to describe. Ultimately, this is the hard thing at the root of philosophy. In practice, it manifests like e.g. choice between digging into some obscure branches of mathematics to correctly name the thing "endofunctor" or something, or calling it "Square" and saying "fuck it, we'll clarify the exceptions in the comments".
--
[0] - I mean "more specific" in the sense it's distinct from the other meanings and somewhat narrow - but still it's fuzzy as heck and you can't describe it fully in words; it's basically tacit knowledge.
I also avoid letting the reader make assumptions. HasPlayerJumpedRecently is bad. What does recently mean? HasPlayerJumpedInLastTenMs is better, even if it's a bit long...Which highlights that it should probably be refactored into a more flexible value; MsSincePlayerLastJumped.
If you arent assuming a time var wth Ms is milliseconds you aren't doing games dev so that one slides with me.
I use to be quite fond of short identifiers, especially ones the make the signs “line up”… until I worked with code long enough that I forgot what I did and had to read it again.
But that's what I meant by the epistemological aspect - what is "recently"? Still, "LastTenMs" is arguably even worse - why ten? Did you really mean "since last frame"? Then say "SinceLastFrame". Did you mean "since 2x update time delta"? Then maybe make a PlayerJumpCooldown constant, but then maybe it's not about cooldowns, so... here's when I'd probably say screw this, let's just track MsSincePlayerLastJumped and add a HasPlayerJumpedRecently() -> bool as a helper, with a fat interface comment explaining what "recently" means - and go back to talk with the game designers to give a more specific name for the "recently" thing here.
Point being, this is deceptively hard - doubly so because going all perfectionist about it is also bad.
If i were to helperise it with a bool I'd use the intent for the bool, CanPlayerJumpAgain or something (again...contrived example).
All software has to name things, and count. Caching (including invalidation) is best understood as a liability. If you can foist it off on your CPU and OS and DB, good for you. Programming whatever you're actually trying to get done is already hard enough.
They also tend not to be very hard.
Really the only time in my entire professional career that off-by-one errors have actually given me headaches.
A lot of languages have just settled on zero indexing, and many now have some variation of for/each or for/of that would eliminate a lot of potential ways to encounter this error.
Occasionally, though, I need to use the same index across multiple items, there’s not a trivial means in which to zip, and at that point I have to use an old school for loop. That’s when the 1-index vs 0-index bites me.
If you have a system with "slow storage" and "fast storage", caching is a way to abstract that away to just "storage".
The author is arguing that the latter is the default way we should think about the concept of caching, which is a valid opinion to have.
We use caching a lot, anything that gets cached can only be written by one service each. The writing services emit cache invalidation messages via SNS that cache users must listen to via SQS, to clear/update their cache.
Alternatively we cache stuff with just a TTL, when immediate cache invalidation is not important.
Where‘s the struggle?
If it doesn’t guarantee delivery, then I believe you will at some point have a client that reads a cached value thinking it’s still valid because the invalidation message got lost in the network.
If there are no real consequences for reading stale data, and your writes are infrequent enough, then indeed you're lucky and have a relatively simple problem.
If you don't understand how and why and when eventual consistency is a problem, you will never understand why cache invalidation is hard.
By the sound of your example, you only handle scenarios where naive approaches to cache invalidation serve your needs, and you don't even have to deal with problems caused by spikes to origin servers. That's perfectly fine.
Others do. They understand the meme. You can too if you invest a fee minutes reading up on the topic.
It's used in DNS, which already was an example here. There is no way to be sure clients see an updated value before end of TTL. As a result, you have to use very conservative TTLs. It's very inefficient.
> anything that gets cached can only be written by one service each
How do you guarantee it's only written by one service each? Sounds like locking across network boundaries, which is not easy.
> The writing services emit cache invalidation messages via SNS that cache users must listen to via SQS
SNS and SQS are both nontrivial services (at least you don't have to build / maintain them I suppose) that require training to use effectively and avoid any possible footguns
I think you're underestimating the complexity in your own solution, and you're probably lucky that some of the harder problems have already been solved for you.
1. Be content with/resilient to the possibility of stale data.
2. Gatekeep all reads and writes (for some subset of the key space) through a single thread.
That's basically it.
Both are not that difficult, honestly.
Aren’t there a lot harder things out there
Heck you can probably prove that any system for naming things is either inconsistent or incomplete.
Well, I for one feel that "naming things" ultimately boils down to the latter, which may or may not be harder than the former.
You could group these two things into "getting the data model right" as the single hard thing, perhaps that rings more true to you :)
It's also possible that these used to be the only two hard problems at the time the aphorism was first recorded, but the underlying state of the world has changed since then and the aphorism, as recorded, is no longer current.
Think about all those times your program isn't building and `make clean` fixes it.
I don't think that's a good example. I've worked with more than 20 different build tools by now, and I cannot recall a single instance where the problem actually came down to cache invalidation. Every time I dug into it, the real cause was something else: a mistake in the build script, an incorrectly declared dependency, or something similar.
So when you say "think about all those times", not a single such time comes to my mind!
That literally IS cache invalidation. Incremental build systems are caching previous artefacts. They're supposed to invalidate the artefact if a dependency changes. But you forgot to declare the dependency... so it doesn't get invalidated when it should.
A lot of people haven't caught on, and try to cache things using ambiguous names, hence the struggle to invalidate their caches when the meaning changes.
[1] This can be applied even if you don't know the content yet; you just have to unambiguously name the inputs to the function that produces it. You might not know what all the inputs are, and then you have to start adding stuff like "unknown-unknown-2025-07-03T16", but it'll still basically work.
Here's one for PHP: https://github.com/Qbix/Platform/blob/main/platform/classes/...
And here is for the Web: https://github.com/Qbix/Platform/blob/dc95bd85fa386e45546b0b...
I also discuss caching in the context of HTTP: https://community.qbix.com/t/caching-and-sessions/192
WRITING ABOUT CACHING AND INVALIDATION
And then you can read about incremental updates: https://community.qbix.com/t/streams-plugin-messages-subscri...
Caching becomes hard when such a stores are used in distributed or concurrent contexts.
An example: imagine Qcache being used when fetching data from an SQL database. And data in the database changes with a direct SQL query from another process.
How will your key-value store know that the value in it is stale and needs refreshing?
Caching systems know when something needs updating several ways. One is pubsub: When the information is loaded from the cache, you want to set up a realtime websocket or webhook for example, to reflect any changes since the cached info was shown. If you can manage to have a server running, you can simply receive new data (deltas) and update your cached data — in that case you can even have it ready and never stale by the time it is shown.
If you can’t run a server and don’t want to reveive pushes then another approach simply involves storing the latest ordinal or state for EACH cached item locally (e-tags does this) and then before rendering (or right after you do) bulk-sending the tags and seeing what has been updated, then pulling just that. The downside is that you may have a flash of old content if you show it optimistically.
If you combine the two methods, you can easily get an up-to-date syndication of a remote mutable data store.
My whole framework is built around such abstractions, to take care of them for people.
Phrases like "simply" and "never stale" are doing a lot of heavy lifting. Yet you haven't answered a very simple question I posted above: how would Q_Cache structure handle a direct SQL write query from another process on the database it is being used as a cache for?
SQL databases don't run websockets to external servers, nor do they fire webhooks. If you are running a server which you are hitting with every update instead of doing direct SQL on the DB, there's always a small amount of time where a cache user can see stale data before this server manages to trigger the update of the cache.
The SQL server should have a way to do pubsub. It can then notify your PHP webhook via HTTP, or run a script on the command line, when a row changes. It should have an interface to subscribe to these changes.
If your SQL data store lacks the ability to notify others that a row changed, then there’s your main problem. Add that functionality. Make a TRIGGER in SQL for instance and use an extension.
If MySQL really lacks this ability then just put node.js middleware in front of it that will do this for you. And make sure that all mysql transactions from all clients go through that middleware. Now your combined MySQL+Node server is adequate to help clients avoid stale data.
As I already said, if you for some reason refuse to handle push updates, then you have to store the latest row state (etags) in your PHP cache (apcu). And then when you access a bunch of cached items on a request, at the end collect all their ids and etags and simply bulk query node.js for any cached items that changed. You can use bloom filters or merkle trees or prolly trees to optimize the query.
Joel Gustafson had a great article on this and now uses it in gossiplog: https://joelgustafson.com/posts/2023-05-04/merklizing-the-ke...
Not trivial, really? Here, enjoy: https://github.com/krowinski/php-mysql-replication
Are you saying that “invalidating caches is hard… if you insist on using only systems with no way to push updates, and oh if they have a way to push updates then you’re using someone’s library to consume updates so it doesn’t count?”
What if an SQL update changes a million rows, and you had cached the sum of those million rows? Should it send a million notifications? Should it re-run the sum for you? What if it's a complex operation and it takes a while to re-compute, and another update arrives before the re-compute finishes?
And of course, you will always have some partitions, so you will occasionally need to re-query the data anyway and re-establish any kind of pubsub system. And the complexity in efficiently reconciling two views of a large amount of data is a major issue, that you are just brushing off as "you can do various things to maybe optimize the query".
All the difficulties you cite, including the “optimizations” I mentioned, stem from not using push, and insisting on periodically polling something.
If you step back, your whole problem is that you are using a system that only has pull and not push architecture. You’re trying to put a bandaid on that core fact. You chose a relational database system that you refuse to build extensions for (as opposed to pretty much any other system) in order to insist “nyeh nyeh you havent solved cache invalidation”.
Caching is just one part of sync. The part that stores a non-authoritative, slightly out of date copy locally. And sync, or replication, is just one part of an eventually consistent system. I mean, even MySQL supports replication, so just hook your client up to that protocol (encrypted in transit) and boom, now you can update your cache.
Here’s the thing. If you use any network protocol that is open, eg HTTP, then yea it’s solved. Because you can have webhooks and just spin up a server to handle them, and update your cache. A row is removed? Decrement your sum. A row is added? Increment it.
You are just insisting that “no, our system has to be clunky and we refuse to implement webhooks / websockets / sockets / push / mysql replication client / updates of any kind, now solve it for me to be as good as a system which has push capabilities.”
Mysql and Postgresql does in fact have push. It is called replication. The protocol is documented. So, much like a webhook with HTTP protocol, you can ingest changes on a table with a php long-running client and then simply update your apcu cache entries if they are still being stored, and matching that row. And all your PHP processes (managed by php-fpm or FrankenPHP or whatever) will benefit from that atomic apcu update.
There is nothing that says only another mysql server can be a mysql replication client.
Boom, solved. Even in your case. Note that the SQL database did in fact expect you to open a socket, and then it writes its log to that socket. That’s all the streaming events you need.
But now you’re gonna say “but but no you cant have a long running PHP client, we’re talking about a shared hosting environment where the host disables your ability to run a long PHP service that opens a socket”. Well, look. Most of the difficultu is that you’re intentionally trying to use the most clunky architectures just to “prove” that “cache invalidation is really hard.”
Shall we continue with Naming next? :)
More importantly, you are pushing all the complex logic to this imaginary DB replication client: your "caching system" is still nothing more than a KV store.
It becomes way more difficult when you have N servers all of which could potentially serve the same data. Local caches could then, yes easily become stale, and even if you have a propagation mechanism, you couldn't guarantee against network failures or latency issues.
But suppose you have an HA redis instance that you store your cached data in. Even with a write-through cache, you basically need to implement a 2-phase commit to ensure consistency.
Bloom filters, Prolly trees etc can speed the batch polling up. But push architecture obviates the need for it prevents caches from being stale, it’s called eventual consistency.
(Well, of course unless there is a network partition and you can’t communicate with the machine holding the actual source of truth then yeah, your local cache will be stale. And to catch up on missed updates you will need to request all updates since a certain marker you saved. If the upstream had a ton of updates since your marker and can’t send the entire history since your marker then you might have to purge your cache for that specific table / collection, yeah.).
Pretty sure that when it comes to write transactions, such as bank transactions, you want them to be executed on a source of truth with enforces consistency, not eventual consistency. Any client software should not be trusted -- especially one written by a silly programmer who uses a cache as the source of truth.
and - as the OP suggests - it works best when the cache is a well-defined abstraction with properties and rules about how it works
just because "caching" is mentioned in a meme doesn't mean it can't be true that it can simplify software
I have to push back here, I think this is objectively untrue. By definition a system or piece of code on where you add a condition where something else happens (cache) that behaves differently than the uncached path increases complexity.
I'm not saying it's wrong to cache things or that they aren't useful, but I think they absolutely are an abstraction and an optimization at the cost of complexity. Good code bases hide complexity from the devs all the time, so it's not a question of whether you can code it away, but rather how difficult is it to troubleshoot the internals of the system.
Look at how CPU cache line behaviors radically change the performance of superficially similar algorithms.
Look at how query performance for a database server drops off a cliff the moment the working cache no longer fits in memory.
Hiding complexity can be a simplification, until you exceed the bounds of the simplification and the complexity you hid demands your attention anyway.
There's a long history in computer architecture of cores and accelerators that don't have a cache but instead rely on explicitly programmed local scratchpads. They are universally more difficult to program than general purpose CPUs because of that.
But getting them right can easily cross the boundary of purely optimizing performance towards simplifying public API of something. I think this is true.
I'd imagine an involved example where semantics and caching really start to offer a trade-off.
Imagine that somehow querying the actual meteorological data is quite expensive, and consider this badly written pseudocode (equals sign denoting default parameters):
- measureCurrentTemparature()
- retrieveAccurateTemperatureForNanoSecond(momentInTime)
-> cached abstractions which would access cached data:
- getTempearature(moment = now(), tolerance = 1min)
- getCurrentTemperature(tolerance = MIN_TOLERANCE)
I know, reality is much more complicated, and using time (seeing it as quasi-continuous) as a caching parameter is already stretching it so far.
Just a stupid example that came to my mind.
I've bitten myself in the ass with caching rasterized reprentations of images more than once, where the input were SVG images or limited formats that convert to SVG.
The only scenario where it would simplify software is if a bunch of complex (non-cache) things are being done to improve perf, and a cache would be the simpler solution. But in that case the simplifying step is not adding a cache, it is removing complex things that aren't actually required. After that you add a cache to improve performance (which increases complexity but is worth it for this imagined use-case). But maybe you remove the complex perf shenanigans, and realize that perf is still "good enough" even without a cache, keeping your software even simpler.