"I think now caching is probably best understood as a tool for making software simpler" - that's cute. Caching might be beneficial for many cases, but if it doesn't do one thing then this is simplifying software. There's that famous quote "There are only two hard things in Computer Science: cache invalidation and naming things.", and, sure, it's a bit ironical, but there's some truth in there.
"Two programs could have similar behaviour but structured very differently, the difference being that one utilizes caching as an abstraction and one explicitly has the concept of different tiers of storage."
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.
> There's that famous quote "There are only two hard things in Computer Science: cache invalidation and naming things.", and, sure, it's a bit ironical, but there's some truth in there.
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.
Caching is simple, yes. The hard part is in the last word, invalidation. Even that is manageable for a single process. But as soon as you have multiple (threads / processes / nodes / data centers) updating the data, it does get quite complex, pretty fast.
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)
I'd say this is not the "naming things" that's hard. Beyond picking a common identifier format in the team, there are at least two dimensions that are much harder:
- 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 try to name things descriptively in simple terms and often end up with NamesAboutThisLong, once they get too long i know the thing is doing too much and some refactoring is needed for readability.
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.
Wow cool, you just summed up something I’ve found myself doing subconsciously in the past few years. Thanks!
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.
I figured the naming issue is deciding how much context. A name might begin inside an organization but need to endure a wider area. If you make all names so long and context-free that they can work in any context, they become unwieldy. Also it can be hard to realize some of the implicit context and what needs to be differentiated with the name. Where server used to suffice, now you need server-a and server-b.
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're not hard but I will say that when I was writing an app that was using both JavaScript and Julia, I kept getting off-by-one errors because Julia starts at 1 instead of 0.
Really the only time in my entire professional career that off-by-one errors have actually given me headaches.
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.
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.
You don’t support read-your-own-write and your cache data might be stale for arbitrarily long. These relaxed consistency constraints make caching a lot easier. If that’s acceptable to your use cases then you’re in a great place! If not… well, at scale you often need to find a way for it to be acceptable anyway
Does SQS guarantee delivery to all clients? If it does then that’s doing a lot of heavy lifting for you.
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 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.
> 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!
I think the idea is that if make clean allows the build to complete correctly, then the underlying issue is that the "mistake in the build script, incorrectly declared dependency, or similar" was causing the cached build results to not be invalidated when they should have.
In my experience, the larger the software you write, the truer these become. At some point all obvious names will have collisions, and getting caching right is crucial to do but difficult to achieve because it transcends the entire stack.
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 a little bit tongue in cheek; no one is seriously suggesting it's harder than P=NP or the problem of consciousness. But there's something a bit "death and taxes" to the inevitability that any large enough project is going to have some corner cases involving these old chestnuts.
Heck you can probably prove that any system for naming things is either inconsistent or incomplete.
Namings things is of course a bit tongue in cheek. But cache invalidation is hard. For example, allegedly MESI is one of the hardest things to validate in processor design.
For "only two hard problems," read "two candidates for among the hardest problems (but we feel strongly that these are indeed good candidates)," or something along those lines, more or less.
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.
Really? Have you tried building any substantial program that makes use of caching and succeeded in invalidating the cache both correctly and efficiently? It's not all about simple things like disk access, caching is also useful in software that models complex hardware where properties depend on multitudes of interconnected calculated values that are time consuming to calculate and where you cannot predict which ones the client will ask for next.
Your implementations are not cache systems: they are key-value store abstractions which can be used for caching. There is a simpler one in most languages (IIRC, named "hashes" in PHP).
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?
Yes, they are cache systems. And as you say, they can be used for caching.
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.
We can agree to disagree on this being a "caching system" and it "knowing when something needs updating" (from the API, I am guessing it's a "set" method).
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.
Simple. Use the principles I said above, for distributed and asynchronous systems.
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.
There is a secret technique, called content-addressing[1], which elegantly solves both of them at once.
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.
> caching often does simplify software though when done well
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.
Caching is a performance improvement. There is no software that requires caching, therefore it is always something being added on top of the business logic that is fundamentally required. As such, a cache is increasing complexity by nature of its existence.
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.
CPUs are still a great example for how caching simplifies things.
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.
I’m sure the CPU designers would love it if they didn’t need several different layers of cache. Or no cache at all. Imagine if memory IOPS were as fast as L1 cache, no need for all that dedicated SRAM on the chip or worry about side channel attacks.
That abstraction is another layer though. And additional layers are additional complexity. So, if you add another layer, the software is less simple than before.
You might need to have caching in your software. I don't doubt that. But there's simply no way it makes the software more simple except if you assume some unfortunate starting point where you could get rid of any high-complex performance optimizations in your existing code by replacing them with a more simple cache solution. But then the statement should be "refactoring makes your code simpler".
Getting cache keys or caching events wrong is easy and a nightmare.
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):
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.
On top of the other things mentioned (caching always introduces complexity with lifetime tracking, and thus can't make things simple), the article's got it the wrong way around.
When code has abstract interfaces for data access, introducing caching can be simpler (but not simple) by localizing it in the abstraction implementation which has or doesn't have caching.
But it is not an abstraction (you can perfectly well do caching without any abstractions, and it's frequently done exactly that way).
I think you and the article are referring to abstractions over different concerns.
The concern you're talking about is about the actual access to the data. My understanding of the article is that it's about how caching algorithms can abstract the concern of minimising retrieval cost.
So in some ways you're coming at it from opposite directions. You're talking about a prior of "disk by default" and saying that a good abstraction lets you insert cache layers above that, whereas for the author the base case is "manually managing the layers of storage".
Algorithms can't really abstract anything since they are, well, just algorithms (formal descriptions of how a computation should be done).
Looking at the author's examples again, I think most everybody would say that caching is used in both:
if data_id in fast_storage:
return fast_storage.get(data_id)
else:
data = slow_storage.get(data_id)
fast_storage.set(data_id, data)
return data
and
# Uses fast storage or slow storage just like above, but behind the get() method.
return storage.get(data_id)
The first one does not make an abstraction on storage, the second one does, but they are both "caching" data internally.
While there are genericimplementations of caching algorithms and we can consider those abstractions, "caching" is a wider term than those implementations, and is specifically not an abstraction (the fact that there is a caching implementation that abstracts something does not make all caching an abstraction).
Edit: Let me also point out that "abstract the concern of minimising retrieval cost" is not caching — I can say that eg. a simple interface with method FastGet(id) does the former, and it needs not use any caching if the underlying structure is fast enough and eg. directly in memory.
This is correct, I appreciate you for putting it so
coherently :). I think I didn’t make it clear enough in the piece that I’m coming from a stance of fast access being table stakes, and the question being about how that’s accomplished.
"Caching" is an idea of storing a result of an expensive computation in storage that is faster to get from than doing the original computation (in very generic computer terms, computation can be simply fetching from the network or slower local storage).
What you describe as "caching algorithms" are not really caching algorithms, but cached object lifetime management algorithms (LRU, LFU...).
It sounds like you are both trying to redefine what "caching" means (tying it to implementations of particular algorithms), but also what "abstraction" means.
We should be very deliberate with the language we use, and our main goal should be to make it simpler to understand, not harder — I believe you are doing the latter here.
Most optimizations require you to think about how your code is structured, so as a side-effect you make the code more understandable.
In this article, it's cache levels forcing you to separate different types of data because they're accessed at different frequencies. Another example is Rust's borrow checker, whose main purpose is arguably to facilitate both safe and efficient memory management, but which can also be used to enforce invariants that aren't clearly memory-related (e.g. builder pattern, temp files that auto-delete after they're dropped).
These aren't abstractions though. An abstraction is the opposite, hiding structure when it's noisy and making it easier to change. For example, if you already have an architecture in mind and don't want to manually determine how frequently each type of data is accessed, it's better to use a compiler or library that automatically determines what to cache with little to no code or thought on your end; that's abstraction. Similarly, the abstract analogue to Rust's borrow checker is garbage collection, which allows programmers to not think about their data-structures' lifetimes at all. The cost is usually performance and you understand your application less in some ways (although you understand it more in other ways; abstraction hides details but too many details make it hard to see the big picture. Ideally, with abstractions in the right places, you hide only the "unimportant" details in ways that insignificantly affect performance).
As can be seen from other comments, people tend to focus on the consistency implications, but something not discussed often in the context of distributed systems is that caches tend to introduce bimodality and metastability [0] [1].
See e.g. DynamoDB for an example of design taking it into account [2].
Note: the author means that caching can be used as an implementation detail in an (abstracted) storage access system, as opposed to a baseline of having multiple storage systems (fast, medium, slow) and managing them directly.
This was confusing to me – the most obvious way to judge the purpose of a system is to compare with the baseline of not having that system at all, especially in the case of caching where the program is functionally complete and correct without a cache. Anyway, there may not be a right or wrong here. Just tripped me up.
Caching is an optimisation. Sometimes caching can be abstracted away, eg CPU cache or build cache are pretty much abstracted away for a usual web developer. But web page caching is very hard to abstract without any abstraction leaks and weird bugs. And even CPU cache is no longer an abstraction if you deal with very high performance code.
It sounds like they are arguing that when performance matters, you have to know more about caching. Fair enough, you have to know a lot more about things when optimizing. For a lot of cases you can ignore caching because it can be done transparently. You depend on it to some extent because if e.g. every instruction had to be fetched off rotating storage like the old days, it would play a big role in your design. It's just something solved in general for most software to not have to know much about it.
What? No, caching means a specific thing: keeping a copy of data away from the source of truth, closer to where you want to read it. Caching always makes systems more complex, it never makes things simpler, and it damn sure doesn't serve as any kind of abstraction unless you're redefining what words mean to indulge your technical philosophizing.
This article is talking about single-writer, single-reader storage. I think it's correct in that context. Most of the hairy problems with caches don't come up until you're multi-writer, multi-reader.
"fast storage" is about performance, your abstraction includes performance elements. If you go that down, then you are optimizing on your abstraction designs. What doesn't have to be wrong, but then don't say that is not optimization.
There is also an important (but often overlooked) detail that you/your application may not be the only user of the cache. At which point caching, indeed, is an optimization via abstraction: when you fetch an X, you are in no position to predict that the next fifty completely unrelated to you requests would also want to fetch the same X, so it should probably be cached to be readily served.
Which is why solving the "I want my data in fast storage as often as possible" problem may be counter-productive on the whole: you ain't the only client of the system; let it breath and server requests from others.
I believe this same language use is what makes this article confusing: Redis is not a cache, it is a key value store. Caching is usually implemented using key value stores, but it is not an abstraction (leaky or not).
In RDBMS contexts, index really is a caching mechanism (a cache) managed by the database system (query planner needs to decide when it's best to use one index or another).
But as you note yourself even in these cases where you've got cache management bundled with the database, having too many can slow down (even deadlock) writes so much as the database tries to ensure consistency between these redundant data storage elements.
Even more obvious if you think about the case of hardware-managed caches! The ISA typically exposes some simple cache control instructions (and I guess non-temporal loads/stores?), but apart from that, the actual choice of storage location is abstracted away from you (and your compiler).
Sometimes posts are so difficult to read they're hard to respond to. I think I get what they're saying. I think they're saying that they think caching should be simple, or at least, that it should be obvious how you should cache in your particular situation such that you don't need things like algorithms. But that argument is kind of nonsense, because really everything in software is an algorithm.
Caching is storing a copy of data in a place or way that it is faster to retrieve than it would be otherwise. Caching is not an abstraction; it is a computer science technique to achieve improved performance.
Caching does not make software simpler. In fact, it always, by necessity, makes software more complex. For example, there are:
- Routines to look up data in a fast storage medium
- Routines to retrieve data from a slow storage medium and store them in a fast storage medium
- Routines to remove the cache if an expiration is reached
- Routines to remove cache entries if we run out of cache storage
- Routines to remove the oldest unused cache entry
- Routines to remove the newest cache entry
- Routines to store the age of each cache entry access
- Routines to remove cache entries which have been used the least
- Routines to remove specific cache entries regardless of age
- Routines to store data in the cache at the same time as slow storage
- Routines to store data in cache and only write to slow storage occasionally
- Routines to clear out the data and get it again on-demand/as necessary
- Routines to inform other systems about the state of your cache
- ...and many, many more
Each routine involves a calculation that determines whether the cache will be beneficial. A hit or miss can lead to operations which may add or remove latency, may or may not run into consistency problems, may or may not require remediation. The cache may need to be warmed up, or it may be fine starting cold. Clearing the cache (ex. restarts) may cause such a drastic cascading failure that the system cannot be started again. And there is often a large amount of statistics and analysis needed to optimize a caching strategy.
These are just a few of the considerations of caching. Caching is famously one of the hardest problems in computer science. How caching is implemented, and what it affects, can be very complex, and needs to be considered carefully. If you try to abstract it away, it usually leads to problems. Though if you don't try to abstract it away, it also leads to problems. Because of all of that, abstracting caching away into "general storage engine" is simply impossible in many cases.
Caching also isn't just having data in fast storage. Caching is cheating. You want to provide your data faster than actually works with your normal data storage (or transfer mechanism, etc). So you cheat, by copying it somewhere faster. And you cheat again, by trying to figure out how to look it up fast. And cheat again, by trying to figure out how to deal with its state being ultimately separate from the state of the "real" data in storage.
Basically caching is us trying to be really clever and work around our inherent limitations. But often we're not as smart as we think we are, and our clever cheat can bite us. So my advice is to design your system to work well without caching. You will thank yourself later, when you finally are dealing with the bug bites, and realize you dodged a bullet before.
I was discussing this with someone recently, caching is one of those things that people might do behind the scenes, thinking that it doesn't affect the API but in fact it can create all sorts of issues/complexity.
"I think now caching is probably best understood as a tool for making software simpler" - that's cute. Caching might be beneficial for many cases, but if it doesn't do one thing then this is simplifying software. There's that famous quote "There are only two hard things in Computer Science: cache invalidation and naming things.", and, sure, it's a bit ironical, but there's some truth in there.
"Two programs could have similar behaviour but structured very differently, the difference being that one utilizes caching as an abstraction and one explicitly has the concept of different tiers of storage."
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
> There's that famous quote "There are only two hard things in Computer Science: cache invalidation and naming things.", and, sure, it's a bit ironical, but there's some truth in there.
The joke form of this quote goes along the lines of:
:-DI rather like the snark of:
there's two hard problems in computer science: we only have one joke and it's not funny.
Apparently⁰ by Philip Scott Bowden¹
⁰ https://martinfowler.com/bliki/TwoHardThings.html
¹ https://x.com/pbowden/status/468855097879830528
Caching is simple, yes. The hard part is in the last word, invalidation. Even that is manageable for a single process. But as soon as you have multiple (threads / processes / nodes / data centers) updating the data, it does get quite complex, pretty fast.
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)
Don't bring a PowerPoint to a Vi/Emacs fight...
I'd say this is not the "naming things" that's hard. Beyond picking a common identifier format in the team, there are at least two dimensions that are much harder:
- 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 try to name things descriptively in simple terms and often end up with NamesAboutThisLong, once they get too long i know the thing is doing too much and some refactoring is needed for readability.
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.
Wow cool, you just summed up something I’ve found myself doing subconsciously in the past few years. Thanks!
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.
I figured the naming issue is deciding how much context. A name might begin inside an organization but need to endure a wider area. If you make all names so long and context-free that they can work in any context, they become unwieldy. Also it can be hard to realize some of the implicit context and what needs to be differentiated with the name. Where server used to suffice, now you need server-a and server-b.
If you have a system with "slow storage", caching is a way to optimize that to "storage that is sometimes fast".
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.
(You forgot off-by-1 errors.)
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.
Off by 1-errors is not part of the original quote, but is just a later addon to make it funny.
They also tend not to be very hard.
Except when they're part of some base assumptions in the domain or dozen of layers of abstractions below you. They are hard to prevent from happening.
They're not hard but I will say that when I was writing an app that was using both JavaScript and Julia, I kept getting off-by-one errors because Julia starts at 1 instead of 0.
Really the only time in my entire professional career that off-by-one errors have actually given me headaches.
I never understood this meme.
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?
> Where‘s the struggle?
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.
You don’t support read-your-own-write and your cache data might be stale for arbitrarily long. These relaxed consistency constraints make caching a lot easier. If that’s acceptable to your use cases then you’re in a great place! If not… well, at scale you often need to find a way for it to be acceptable anyway
Does SQS guarantee delivery to all clients? If it does then that’s doing a lot of heavy lifting for you.
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.
Eventually. The problem is that eventually delivering that message will result in clients assuming that it will always be the same, when it’s not.
> I never understood this meme.
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.
Here's one: everybody invalidating and refreshing their cache at the same time can cause a thundering herd problem.
I never understood about cache invalidation or naming things
Both are not that difficult, honestly.
Aren’t there a lot harder things out there
Cache invalidation isn't hard in theory. It's just one of those things that is very easy to get subtly wrong and difficult to test.
Think about all those times your program isn't building and `make clean` fixes it.
> 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!
I think the idea is that if make clean allows the build to complete correctly, then the underlying issue is that the "mistake in the build script, incorrectly declared dependency, or similar" was causing the cached build results to not be invalidated when they should have.
In my experience, the larger the software you write, the truer these become. At some point all obvious names will have collisions, and getting caching right is crucial to do but difficult to achieve because it transcends the entire stack.
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 a little bit tongue in cheek; no one is seriously suggesting it's harder than P=NP or the problem of consciousness. But there's something a bit "death and taxes" to the inevitability that any large enough project is going to have some corner cases involving these old chestnuts.
Heck you can probably prove that any system for naming things is either inconsistent or incomplete.
> no one is seriously suggesting it's harder than P=NP or the problem of consciousness.
Well, I for one feel that "naming things" ultimately boils down to the latter, which may or may not be harder than the former.
Namings things is of course a bit tongue in cheek. But cache invalidation is hard. For example, allegedly MESI is one of the hardest things to validate in processor design.
For "only two hard problems," read "two candidates for among the hardest problems (but we feel strongly that these are indeed good candidates)," or something along those lines, more or less.
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.
Really? Have you tried building any substantial program that makes use of caching and succeeded in invalidating the cache both correctly and efficiently? It's not all about simple things like disk access, caching is also useful in software that models complex hardware where properties depend on multitudes of interconnected calculated values that are time consuming to calculate and where you cannot predict which ones the client will ask for next.
Yes, I have. I've actually built a general-purpose cache system. Several, in fact:
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
You should especially start here: https://community.qbix.com/t/qbix-websites-loading-quickly/2...And then you can read about incremental updates: https://community.qbix.com/t/streams-plugin-messages-subscri...
Your implementations are not cache systems: they are key-value store abstractions which can be used for caching. There is a simpler one in most languages (IIRC, named "hashes" in PHP).
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?
Yes, they are cache systems. And as you say, they can be used for caching.
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.
We can agree to disagree on this being a "caching system" and it "knowing when something needs updating" (from the API, I am guessing it's a "set" method).
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.
Simple. Use the principles I said above, for distributed and asynchronous systems.
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...
There is a secret technique, called content-addressing[1], which elegantly solves both of them at once.
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.
[dead]
caching often does simplify software though when done well
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
> caching often does simplify software though when done well
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.
Caching is a performance improvement. There is no software that requires caching, therefore it is always something being added on top of the business logic that is fundamentally required. As such, a cache is increasing complexity by nature of its existence.
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.
If you hide caching away as an implementation detail behind an abstraction, it comes back and bites you as a leaky abstraction later.
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.
CPUs are still a great example for how caching simplifies things.
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.
I’m sure the CPU designers would love it if they didn’t need several different layers of cache. Or no cache at all. Imagine if memory IOPS were as fast as L1 cache, no need for all that dedicated SRAM on the chip or worry about side channel attacks.
That abstraction is another layer though. And additional layers are additional complexity. So, if you add another layer, the software is less simple than before. You might need to have caching in your software. I don't doubt that. But there's simply no way it makes the software more simple except if you assume some unfortunate starting point where you could get rid of any high-complex performance optimizations in your existing code by replacing them with a more simple cache solution. But then the statement should be "refactoring makes your code simpler".
additional layers (or software in general) are not inherently additional complexity
In some sense they are, since establishing an abstraction is strictly additive. Abstractions help manage complexity.
Getting cache keys or caching events wrong is easy and a nightmare.
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.
Trying some other way to explicitly manage multiple storage tiers could get pretty complicated.
I guess simplification needs to include "at what level" as a qualifier.
On top of the other things mentioned (caching always introduces complexity with lifetime tracking, and thus can't make things simple), the article's got it the wrong way around.
When code has abstract interfaces for data access, introducing caching can be simpler (but not simple) by localizing it in the abstraction implementation which has or doesn't have caching.
But it is not an abstraction (you can perfectly well do caching without any abstractions, and it's frequently done exactly that way).
I think you and the article are referring to abstractions over different concerns.
The concern you're talking about is about the actual access to the data. My understanding of the article is that it's about how caching algorithms can abstract the concern of minimising retrieval cost.
So in some ways you're coming at it from opposite directions. You're talking about a prior of "disk by default" and saying that a good abstraction lets you insert cache layers above that, whereas for the author the base case is "manually managing the layers of storage".
The language used is seriously confusing here.
Algorithms can't really abstract anything since they are, well, just algorithms (formal descriptions of how a computation should be done).
Looking at the author's examples again, I think most everybody would say that caching is used in both:
and The first one does not make an abstraction on storage, the second one does, but they are both "caching" data internally.While there are generic implementations of caching algorithms and we can consider those abstractions, "caching" is a wider term than those implementations, and is specifically not an abstraction (the fact that there is a caching implementation that abstracts something does not make all caching an abstraction).
Edit: Let me also point out that "abstract the concern of minimising retrieval cost" is not caching — I can say that eg. a simple interface with method FastGet(id) does the former, and it needs not use any caching if the underlying structure is fast enough and eg. directly in memory.
This is correct, I appreciate you for putting it so coherently :). I think I didn’t make it clear enough in the piece that I’m coming from a stance of fast access being table stakes, and the question being about how that’s accomplished.
"Caching" is an idea of storing a result of an expensive computation in storage that is faster to get from than doing the original computation (in very generic computer terms, computation can be simply fetching from the network or slower local storage).
What you describe as "caching algorithms" are not really caching algorithms, but cached object lifetime management algorithms (LRU, LFU...).
"Abstraction" is a higher level, simplified view of a set of concepts, yet caching is a single concept. See eg. https://en.wikipedia.org/wiki/Abstraction_(computer_science)
It sounds like you are both trying to redefine what "caching" means (tying it to implementations of particular algorithms), but also what "abstraction" means.
We should be very deliberate with the language we use, and our main goal should be to make it simpler to understand, not harder — I believe you are doing the latter here.
Most optimizations require you to think about how your code is structured, so as a side-effect you make the code more understandable.
In this article, it's cache levels forcing you to separate different types of data because they're accessed at different frequencies. Another example is Rust's borrow checker, whose main purpose is arguably to facilitate both safe and efficient memory management, but which can also be used to enforce invariants that aren't clearly memory-related (e.g. builder pattern, temp files that auto-delete after they're dropped).
These aren't abstractions though. An abstraction is the opposite, hiding structure when it's noisy and making it easier to change. For example, if you already have an architecture in mind and don't want to manually determine how frequently each type of data is accessed, it's better to use a compiler or library that automatically determines what to cache with little to no code or thought on your end; that's abstraction. Similarly, the abstract analogue to Rust's borrow checker is garbage collection, which allows programmers to not think about their data-structures' lifetimes at all. The cost is usually performance and you understand your application less in some ways (although you understand it more in other ways; abstraction hides details but too many details make it hard to see the big picture. Ideally, with abstractions in the right places, you hide only the "unimportant" details in ways that insignificantly affect performance).
As can be seen from other comments, people tend to focus on the consistency implications, but something not discussed often in the context of distributed systems is that caches tend to introduce bimodality and metastability [0] [1]. See e.g. DynamoDB for an example of design taking it into account [2].
[0] https://brooker.co.za/blog/2021/08/27/caches.html
[1] https://sigops.org/s/conferences/hotos/2021/papers/hotos21-s...
[2] https://brooker.co.za/blog/2022/07/12/dynamodb.html
Note: the author means that caching can be used as an implementation detail in an (abstracted) storage access system, as opposed to a baseline of having multiple storage systems (fast, medium, slow) and managing them directly.
This was confusing to me – the most obvious way to judge the purpose of a system is to compare with the baseline of not having that system at all, especially in the case of caching where the program is functionally complete and correct without a cache. Anyway, there may not be a right or wrong here. Just tripped me up.
Caching is an optimisation. Sometimes caching can be abstracted away, eg CPU cache or build cache are pretty much abstracted away for a usual web developer. But web page caching is very hard to abstract without any abstraction leaks and weird bugs. And even CPU cache is no longer an abstraction if you deal with very high performance code.
It sounds like they are arguing that when performance matters, you have to know more about caching. Fair enough, you have to know a lot more about things when optimizing. For a lot of cases you can ignore caching because it can be done transparently. You depend on it to some extent because if e.g. every instruction had to be fetched off rotating storage like the old days, it would play a big role in your design. It's just something solved in general for most software to not have to know much about it.
This is basically semantic argument, and I will not be engaging in it
You're right, but caching is an optimization.
What? No, caching means a specific thing: keeping a copy of data away from the source of truth, closer to where you want to read it. Caching always makes systems more complex, it never makes things simpler, and it damn sure doesn't serve as any kind of abstraction unless you're redefining what words mean to indulge your technical philosophizing.
This whole discussion on caching and abstraction was completely befuddling to me.
This article is talking about single-writer, single-reader storage. I think it's correct in that context. Most of the hairy problems with caches don't come up until you're multi-writer, multi-reader.
"fast storage" is about performance, your abstraction includes performance elements. If you go that down, then you are optimizing on your abstraction designs. What doesn't have to be wrong, but then don't say that is not optimization.
There is also an important (but often overlooked) detail that you/your application may not be the only user of the cache. At which point caching, indeed, is an optimization via abstraction: when you fetch an X, you are in no position to predict that the next fifty completely unrelated to you requests would also want to fetch the same X, so it should probably be cached to be readily served.
Which is why solving the "I want my data in fast storage as often as possible" problem may be counter-productive on the whole: you ain't the only client of the system; let it breath and server requests from others.
This reminds me of the use of materialized views as both a cache strategy and as an abstraction helper.
And they too can slow things down. Like all caches can. Like Redis can. Cache is a leaky abstraction.
(Although a materialised view is more like an index than a cache. The view won't expire requiring you to rebuild.)
I believe this same language use is what makes this article confusing: Redis is not a cache, it is a key value store. Caching is usually implemented using key value stores, but it is not an abstraction (leaky or not).
In RDBMS contexts, index really is a caching mechanism (a cache) managed by the database system (query planner needs to decide when it's best to use one index or another).
But as you note yourself even in these cases where you've got cache management bundled with the database, having too many can slow down (even deadlock) writes so much as the database tries to ensure consistency between these redundant data storage elements.
Even more obvious if you think about the case of hardware-managed caches! The ISA typically exposes some simple cache control instructions (and I guess non-temporal loads/stores?), but apart from that, the actual choice of storage location is abstracted away from you (and your compiler).
Use of a better abstraction is an optimization, though.
Did you hand-draw that graffiti? Never quite realized that graffiti of technical ideas looks really goated. Best part of the post, to be honest.
> looks really goated
Oof you're trying so hard you could cut diamond with that line.
I don't even understand what that means. Care to explain?
I think it's a drug reference‽
> I've always been told that caching is a tool to make software faster.
Who told you that?
> you don't have to go all the way back to some backend database or API server or SSD [...] Caching is thus a tool to improve performance.
That's called "latency." This is not at all the same as "performance."
> My feelings now are that that perspective on caching is wrong
I agree.
Everything is caching. Almost nothing operates on the target data directly.
Do you think that's a useful definition of the term?
If everything is caching, why even introduce the term: language should help us describe ideas, it should not be superfluous.
Anything can be an abstraction if designed carefully.
Sometimes posts are so difficult to read they're hard to respond to. I think I get what they're saying. I think they're saying that they think caching should be simple, or at least, that it should be obvious how you should cache in your particular situation such that you don't need things like algorithms. But that argument is kind of nonsense, because really everything in software is an algorithm.
Caching is storing a copy of data in a place or way that it is faster to retrieve than it would be otherwise. Caching is not an abstraction; it is a computer science technique to achieve improved performance.
Caching does not make software simpler. In fact, it always, by necessity, makes software more complex. For example, there are:
Each routine involves a calculation that determines whether the cache will be beneficial. A hit or miss can lead to operations which may add or remove latency, may or may not run into consistency problems, may or may not require remediation. The cache may need to be warmed up, or it may be fine starting cold. Clearing the cache (ex. restarts) may cause such a drastic cascading failure that the system cannot be started again. And there is often a large amount of statistics and analysis needed to optimize a caching strategy.These are just a few of the considerations of caching. Caching is famously one of the hardest problems in computer science. How caching is implemented, and what it affects, can be very complex, and needs to be considered carefully. If you try to abstract it away, it usually leads to problems. Though if you don't try to abstract it away, it also leads to problems. Because of all of that, abstracting caching away into "general storage engine" is simply impossible in many cases.
Caching also isn't just having data in fast storage. Caching is cheating. You want to provide your data faster than actually works with your normal data storage (or transfer mechanism, etc). So you cheat, by copying it somewhere faster. And you cheat again, by trying to figure out how to look it up fast. And cheat again, by trying to figure out how to deal with its state being ultimately separate from the state of the "real" data in storage.
Basically caching is us trying to be really clever and work around our inherent limitations. But often we're not as smart as we think we are, and our clever cheat can bite us. So my advice is to design your system to work well without caching. You will thank yourself later, when you finally are dealing with the bug bites, and realize you dodged a bullet before.
I was discussing this with someone recently, caching is one of those things that people might do behind the scenes, thinking that it doesn't affect the API but in fact it can create all sorts of issues/complexity.