Spring Declarative Caching
Introduction
I had occasion at work recently to consider using Spring’s @Cacheable
annotation. I searched the web for some hints on how to use it, but wasn’t satisfied with any of the results. Even my favorite source1 of bite-size Java code samples had examples that I felt were too simple. Using “lookup” methods as examples of places where you may want to put @CacheEvict
and @CachePut
annotations seemed misleading to me.
So, at the very least, I wanted to provide more substantial examples. In particular, I wanted to show why Spring’s default cache implementation requires careful thought in a distributed application environment.
In the process, I ended up going off the rails a bit. I spent way more time setting up an environment in which I could run my examples than I did writing actual Java code. But, in a way, that’s great! If you came here to read Java code, the good new is: there’s not a lot of it.
Code
For this blog post, I bootstrapped a Spring project using the spring initializr and then put it on GitHub.
To run the examples with the least amount of fuss, you’ll need Docker
, Docker Compose
, and (gnu) make
, curl
, and jq
. If you have those tools, things should go smoothly. However, it is software… so, you know… good luck. If it breaks, you’ve got the source code!
For examples that show a command typed at a prompt, this character is my prompt: ✗
Big Picture
Broadly speaking, you’d like to cache return values of methods that are expensive to compute. You can measure “expensive” in different ways (e.g., CPU or memory used); but often, when we say “expensive,” we mean “it takes too much time.”
If we assume that all methods in question are “functions” in the mathematical sense – i.e., given the same inputs (to include the instance the method is operating on), they always give the same output – then caching is pretty simple. Any return value of such a method can be cached indefinitely, and the only decisions you need to make with respect to your cache are:
- How big can it get?
- Which eviction policy do you choose to ensure the size constraint is maintained?
If we do not assume that the method in question are functions, then we can consider much more interesting examples.
For instance, consider a service that, given an identifier, retrieves a record from a relational database. Assume that this database has multiple clients, each of which can update records. Should you expect that, given the same identifier, the service should return the same record now that it did six hours ago? Of course not; the record could have been updated in the intervening time.
Sample Application
The sample application includes a very simple service for managing a database of books. It uses Spring declarative caching to minimize database lookups. It lets you:
- Retrieve books
- Change books’ titles
Following is a stripped down version of the RestController class that implements the API endpoints for our service (see the source code for the complete class).
@RequestMapping("/books")
public class BookRestController {
private final BookRepository bookRepository;
@GetMapping("/clear")
@CacheEvict(cacheNames = "books", allEntries = true)
public void clearCache() {
}
@GetMapping
public Collection<Book> books() {
return bookRepository.findAll();
}
@GetMapping("/{isbn}")
@Cacheable(cacheNames = "books")
public Book bookByIsbn(@PathVariable("isbn") String isbn) {
return bookRepository.findByIsbn(isbn);
}
@GetMapping("/{isbn}/badUpdateTitle/{title}")
public Book badUpdateTitle(@PathVariable("isbn") String isbn,
@PathVariable("title") String title) {
return bookRepository.save(bookRepository.findByIsbn(isbn).withTitle(title));
}
@GetMapping("/{isbn}/betterUpdateTitle/{title}")
@CacheEvict(cacheNames = "books", key = "#isbn")
public Book betterUpdateTitle(@PathVariable("isbn") String isbn,
@PathVariable("title") String title) {
return bookRepository.save(bookRepository.findByIsbn(isbn).withTitle(title));
}
@GetMapping("/{isbn}/bestUpdateTitle/{title}")
@CachePut(cacheNames = "books", key = "#isbn")
public Book bestUpdateTitle(@PathVariable("isbn") String isbn,
@PathVariable("title") String title) {
return bookRepository.save(bookRepository.findByIsbn(isbn).withTitle(title));
}
}
Some notes on the methods and their use of Spring’s declarative caching.
clearCache
: This method is for testing purposes only! It uses@CacheEvict
to clear ALL items from the cache. It allows us to start from a known state (empty cache) when we are testing. In a production application, however, I can’t imagine a scenario when you’d want to clear the entire cache. This is one of the reasons I didn’t like examples I found on the web: this seemed to be the most popular example use of this annotation.books
: This method returns a list of all Books. It does not cache results. It’s a convenience method that lets us see all Books. There are interesting questions here that I chose to completely ignore (but they are worth further research):- If we were to cache the result, would it cache the list as a unit, or the items in the list individually?
- If it does cache the list as a unit (which I think is the case), how could we make it cache the results individually?
- If the list were cached as a unit, how would we manage the cache when an individual Book was updated? Would we clear ALL cached lists of books (in case it appears in one of them)?
bookByIsbn
: Given an isbn, this method returns a single Book and caches the result (@Cacheable
).badUpdateTitle
: Given an isbn and a title, this method will update the title of the Book determined by the isbn. This is “bad” because it allows for records in the database to be updated without correpsonding records in the cache being updated.betterUpdateTitle
: Given an isbn and a title, this method will update the the title of the Book determined by the isbn and evict any instance of the book from the cache (@CacheEvict
). This is “better” because when a database record is updated, any corresponding record in the cache is removed.bestUpdateTitle
: Given an isbn and a title, this method will update the title of the book determined by the isbn and update any instance of the book from the cache (@CachePut
). This is “best” because when a database record is updated, any corresponding record in the cache is also updated.
A stripped down version of the Book entity follows (see the source for the complete class).
public class Book {
private String isbn;
private String title;
private String author;
private Boolean cached;
private String host;
}
This entity captures the isbn
, title
, and author
of the book – no surprises there.
However, it also has two additional fields: cached
and host
. These two fields aren’t actually stored in the database. They are managed by some clever AspectJ code. For a given instance of a book, they allow us to see:
- Whether it came from a cache.
- Which host provided it in response to our request (which is interesting when we deploy multiple instances of our application).
The same sample application can be deployed multiple ways, each having an impact on caching behavior.
Single JVM
The simplest of these architectures is a single instance of the application, using both an in-memory cache and an in-memory database (all in the same JVM):
In this configuration, all requests are answered by the same application instance. That instance has a single backend database. Its responses are saved in a single cache.
The following sequence diagram shows what happens when we attempt to retrieve a book. The Cache Interceptor intercepts the request and first checks to see if the result is already in the cache. If it is, the interceptor returns the result. If it’s not, the interceptor calls the Endpoint Method and adds the result to the cache before returning it to the Client.
To deploy the application in this configuration, run the following command (from the project’s root directory):
✗ make single-instance
Now let’s kick the tires! Open another shell (because the one in which you ran the previous command should now be occupied).
Note that if we clear the cache and then make 98 requests for the same book, 97 requests are served by the cache2:
✗ make clear-cache-for-all-replicas get-book
for i in {1..5};
do
curl -s http://localhost:8080/books/clear;
done
for i in {1..98};
do
curl -s http://localhost:8080/books/0130305529 | jq -c '. | {host, cached,title}';
done \
| sort | uniq -c
1 {"host":"72fca0a378f9","cached":false,"title":"On Lisp"}
97 {"host":"72fca0a378f9","cached":true,"title":"On Lisp"}
Then, if we (without clearing caches) “bad” update the title and then make 99 requests for the book… all 99 requests are served by the cache, and all 99 have the wrong (old) title:
✗ make bad-update-title get-book
curl -s http://localhost:8080/books/0130305529/badUpdateTitle/"HELLO%20WORLD"%20BAD | jq -c '.'
{"isbn":"0130305529","title":"HELLO WORLD BAD","author":"Paul Graham","cached":false,"host":"72fca0a378f9"}
for i in {1..98};
do
curl -s http://localhost:8080/books/0130305529 | jq -c '. | {host, cached,title}';
done \
| sort | uniq -c
98 {"host":"72fca0a378f9","cached":true,"title":"On Lisp"}
The lesson to be learned here is that you are responsible for keeping the cache up to date with the “source of truth.” In our example, the “source of truth” is our database.
Next, if we (again, without clearing caches) “better” update the title and then make 98 requests for the book… 97 requests are served by the cache, and one is not. However all 98 have the correct (new) title:
✗ make better-update-title get-book
curl -s http://localhost:8080/books/0130305529/betterUpdateTitle/"HELLO%20WORLD"%20BETTER | jq -c '.'
{"isbn":"0130305529","title":"HELLO WORLD BETTER","author":"Paul Graham","cached":false,"host":"72fca0a378f9"}
for i in {1..98};
do
curl -s http://localhost:8080/books/0130305529 | jq -c '. | {host, cached,title}';
done \
| sort | uniq -c
1 {"host":"72fca0a378f9","cached":false,"title":"HELLO WORLD BETTER"}
97 {"host":"72fca0a378f9","cached":true,"title":"HELLO WORLD BETTER"}
This is a much better example of how to use @CacheEvict
(in my opinion) than I was able to find online. Any time we update a record in the database, we evict any instances of it in the cache. This is much better than evicting the entire cache.
Finally, if we (again, without clearing caches) “best” update the title and then make 99 requests for the book… all 99 requests are served by the cache, and all 99 have the correct (new) title:
✗ make best-update-title get-book
curl -s http://localhost:8080/books/0130305529/bestUpdateTitle/"HELLO%20WORLD"%20BEST | jq -c '.'
{"isbn":"0130305529","title":"HELLO WORLD BEST","author":"Paul Graham","cached":false,"host":"72fca0a378f9"}
for i in {1..98};
do
curl -s http://localhost:8080/books/0130305529 | jq -c '. | {host, cached,title}';
done \
| sort | uniq -c
98 {"host":"72fca0a378f9","cached":true,"title":"HELLO WORLD BEST"}
This example is arguably the best method for keeping a cache up to date. Any time a record is updated in the database, we add it to the cache (if it wasn’t there already) or update it in the cache (if it was already there). This allows us to serve as many lookup requests from the cache as possible.
If you want to keep playing, you can use the following command to restore the state to what it was before we started. However, if you plan to just move on to the examples in the next session, there is no need – because we are going to bring down the entire application and deploy it in a new configuration.
✗ make restore-title-for-all-replicas clear-cache-for-all-replicas
Replicas with Individual Caches, Shared Database
A misguided next step in the evolution of this application’s architecture (especially if you think the way I do) would be to create multiple replicas of the app that all store their data in a single database – without also centralizing the cache.
For Spring novices, this is where a huge “gotcha” lies; you are always free to put a @Cacheable
annotation on any component method. If you take no other action than that, you are using Spring’s default cache implementation which is an in-memory cache. This was fine for our single replica example, however, it will cause trouble here.
As you can see from the following sequence diagram, caches still belong to individual replicas, and are updated independently of one another. If a replica answers a given retrieval request, only its cache is populated with the value. If a given replica services an update request, only its cache is updated with the new value. Any other replica that has cached the value previously will now return stale data for lookup requests!
Okay, let’s give this new deployment scenario a spin. First things first. In a terminal where you have a prompt (i.e., not the one currently streaming docker logs from the previous scenario), run this:
✗ make clean
Give it a little time, then check your shell that had the docker logs. All containers should stop, and you should eventually be presented with a prompt again. At that prompt, type:
✗ make replicas-individual-cache-shared-db
Now, if we clear the cache and then make 98 requests for the same book, we’ll see that each replica responds with a single un-cached record, and then responds to successive requests from its own cache.
✗ make clear-cache-for-all-replicas get-book
for i in {1..5};
do
curl -s http://localhost:8080/books/clear;
done
for i in {1..98};
do
curl -s http://localhost:8080/books/0130305529 | jq -c '. | {host, cached,title}';
done \
| sort | uniq -c
1 {"host":"1de0e07a1854","cached":false,"title":"On Lisp"}
18 {"host":"1de0e07a1854","cached":true,"title":"On Lisp"}
1 {"host":"64434b3e44f2","cached":false,"title":"On Lisp"}
18 {"host":"64434b3e44f2","cached":true,"title":"On Lisp"}
1 {"host":"a7c9cadc9353","cached":false,"title":"On Lisp"}
19 {"host":"a7c9cadc9353","cached":true,"title":"On Lisp"}
1 {"host":"e4a74ccf237b","cached":false,"title":"On Lisp"}
19 {"host":"e4a74ccf237b","cached":true,"title":"On Lisp"}
1 {"host":"e5910004e389","cached":false,"title":"On Lisp"}
19 {"host":"e5910004e389","cached":true,"title":"On Lisp"}
Keep in mind that all replicas now have the book we are querying for cached. If we “bad” update the title, successive lookups will all be served from replicas’ caches. However, none of those caches will contain the proper, updated book.
✗ make bad-update-title get-book
curl -s http://localhost:8080/books/0130305529/badUpdateTitle/"HELLO%20WORLD"%20BAD | jq -c '.'
{"isbn":"0130305529","title":"HELLO WORLD BAD","author":"Paul Graham","cached":false,"host":"1de0e07a1854"}
for i in {1..98};
do
curl -s http://localhost:8080/books/0130305529 | jq -c '. | {host, cached,title}';
done \
| sort | uniq -c
19 {"host":"1de0e07a1854","cached":true,"title":"On Lisp"}
20 {"host":"64434b3e44f2","cached":true,"title":"On Lisp"}
19 {"host":"a7c9cadc9353","cached":true,"title":"On Lisp"}
20 {"host":"e4a74ccf237b","cached":true,"title":"On Lisp"}
20 {"host":"e5910004e389","cached":true,"title":"On Lisp"}
But, when we “better” update the title, only ONE stale cached value is removed (the one in the cache of the replica that serviced the update request). When we fetch books after that update, only the cache from which the stale value was removed bothers to repopulate its cache.
✗ make better-update-title get-book
curl -s http://localhost:8080/books/0130305529/betterUpdateTitle/"HELLO%20WORLD"%20BETTER | jq -c '.'
{"isbn":"0130305529","title":"HELLO WORLD BETTER","author":"Paul Graham","cached":false,"host":"a7c9cadc9353"}
for i in {1..98};
do
curl -s http://localhost:8080/books/0130305529 | jq -c '. | {host, cached,title}';
done \
| sort | uniq -c
20 {"host":"1de0e07a1854","cached":true,"title":"On Lisp"}
20 {"host":"64434b3e44f2","cached":true,"title":"On Lisp"}
1 {"host":"a7c9cadc9353","cached":false,"title":"HELLO WORLD BETTER"}
18 {"host":"a7c9cadc9353","cached":true,"title":"HELLO WORLD BETTER"}
19 {"host":"e4a74ccf237b","cached":true,"title":"On Lisp"}
20 {"host":"e5910004e389","cached":true,"title":"On Lisp"}
Finally, the only thing that our “best” update improves is that it prevents a cache-miss from a single replica. It has no effect on data integrity. As you can see below, we now have TWO distinct stale values distributed across our replicas’ caches.
✗ make best-update-title get-book
curl -s http://localhost:8080/books/0130305529/bestUpdateTitle/"HELLO%20WORLD"%20BEST | jq -c '.'
{"isbn":"0130305529","title":"HELLO WORLD BEST","author":"Paul Graham","cached":false,"host":"e4a74ccf237b"}
for i in {1..98};
do
curl -s http://localhost:8080/books/0130305529 | jq -c '. | {host, cached,title}';
done \
| sort | uniq -c
20 {"host":"1de0e07a1854","cached":true,"title":"On Lisp"}
20 {"host":"64434b3e44f2","cached":true,"title":"On Lisp"}
20 {"host":"a7c9cadc9353","cached":true,"title":"HELLO WORLD BETTER"}
19 {"host":"e4a74ccf237b","cached":true,"title":"HELLO WORLD BEST"}
19 {"host":"e5910004e389","cached":true,"title":"On Lisp"}
Okay, that was fun. Remember, if you want to keep messing around with this deployment scenario, you can run the following command to “reset.” Otherwise, let’s move on to the next deployment scenario.
✗ make restore-title-for-all-replicas clear-cache-for-all-replicas
Replicas with Shared Caches, Shared Database
Okay, if it wasn’t immediately obvious that the previous scenario was a bad idea, it should definitely be obvious in hindsight. This scenario fixes the issues of the previous scenario by also using a centralized cache.
As you can see from the following sequence diagram, in this scenario all replicas read from and write to the same cache. Any update to the cache is seen by all replicas3.
Okay, time for our final deployment scenario. You know the drill…
In a shell with a prompt:
✗ make clean
In the shell that had logs and now (after you wait a little bit) has a prompt again:
✗ make replicas-shared-cache-shared-db
This time, when we fetch books, ONLY ONE request isn’t serviced from the cache, despite requests being distributed across five replicas.
✗ make clear-cache-for-all-replicas get-book
for i in {1..5};
do
curl -s http://localhost:8080/books/clear;
done
for i in {1..98};
do
curl -s http://localhost:8080/books/0130305529 | jq -c '. | {host, cached,title}';
done \
| sort | uniq -c
19 {"host":"2bb40d744e9c","cached":true,"title":"On Lisp"}
1 {"host":"380a19db4fcd","cached":false,"title":"On Lisp"}
19 {"host":"380a19db4fcd","cached":true,"title":"On Lisp"}
20 {"host":"4428dd122929","cached":true,"title":"On Lisp"}
19 {"host":"471359f51bd4","cached":true,"title":"On Lisp"}
20 {"host":"b7f5bbac15b7","cached":true,"title":"On Lisp"}
“Bad” update is, of course, still broken: all requests are serviced from the cache, however no response contains the actual book as it exists in the database.
✗ make bad-update-title get-book
curl -s http://localhost:8080/books/0130305529/badUpdateTitle/"HELLO%20WORLD"%20BAD | jq -c '.'
{"isbn":"0130305529","title":"HELLO WORLD BAD","author":"Paul Graham","cached":false,"host":"2bb40d744e9c"}
for i in {1..98};
do
curl -s http://localhost:8080/books/0130305529 | jq -c '. | {host, cached,title}';
done \
| sort | uniq -c
19 {"host":"2bb40d744e9c","cached":true,"title":"On Lisp"}
20 {"host":"380a19db4fcd","cached":true,"title":"On Lisp"}
20 {"host":"4428dd122929","cached":true,"title":"On Lisp"}
20 {"host":"471359f51bd4","cached":true,"title":"On Lisp"}
19 {"host":"b7f5bbac15b7","cached":true,"title":"On Lisp"}
“Better” update now evicts the stale book from the single cache. The first request after this update requires a database lookup, but after that, all replicas are able to service successive requests using the cache.
✗ make better-update-title get-book
curl -s http://localhost:8080/books/0130305529/betterUpdateTitle/"HELLO%20WORLD"%20BETTER | jq -c '.'
{"isbn":"0130305529","title":"HELLO WORLD BETTER","author":"Paul Graham","cached":false,"host":"b7f5bbac15b7"}
for i in {1..98};
do
curl -s http://localhost:8080/books/0130305529 | jq -c '. | {host, cached,title}';
done \
| sort | uniq -c
1 {"host":"2bb40d744e9c","cached":false,"title":"HELLO WORLD BETTER"}
19 {"host":"2bb40d744e9c","cached":true,"title":"HELLO WORLD BETTER"}
20 {"host":"380a19db4fcd","cached":true,"title":"HELLO WORLD BETTER"}
19 {"host":"4428dd122929","cached":true,"title":"HELLO WORLD BETTER"}
20 {"host":"471359f51bd4","cached":true,"title":"HELLO WORLD BETTER"}
19 {"host":"b7f5bbac15b7","cached":true,"title":"HELLO WORLD BETTER"}
And finally, “best” update updates the book without requiring any successive lookups to query the database!
✗ make best-update-title get-book
curl -s http://localhost:8080/books/0130305529/bestUpdateTitle/"HELLO%20WORLD"%20BEST | jq -c '.'
{"isbn":"0130305529","title":"HELLO WORLD BEST","author":"Paul Graham","cached":false,"host":"4428dd122929"}
for i in {1..98};
do
curl -s http://localhost:8080/books/0130305529 | jq -c '. | {host, cached,title}';
done \
| sort | uniq -c
20 {"host":"2bb40d744e9c","cached":true,"title":"HELLO WORLD BEST"}
19 {"host":"380a19db4fcd","cached":true,"title":"HELLO WORLD BEST"}
19 {"host":"4428dd122929","cached":true,"title":"HELLO WORLD BEST"}
20 {"host":"471359f51bd4","cached":true,"title":"HELLO WORLD BEST"}
20 {"host":"b7f5bbac15b7","cached":true,"title":"HELLO WORLD BEST"}
Again, if you want to keep playing around, you can reset the application state with this:
✗ make restore-title-for-all-replicas clear-cache-for-all-replicas
Summary
The big takeaway here is this: Spring makes easy things easy. To that end, it has a default cache implementation. If you’re not careful, you may be inclined to mark a component method as @Cacheable
– and then get on with your life. This could have severe ramifications depending on how your app is deployed now or in the future.
Like I hinted at earlier, I had a lot of fun writing this sample app. The actual “book database web service” was the boring part. Using AspectJ to decorate returned entities was really fun, and I think it made the examples much easier to read (you can tell from the payload which host it came from and whether or not it was cached). It was also a lot of fun using Docker Compose paired with Spring profiles to create wildly different deployment scenarios without having to change a single line of Java code. Seriously, have a look at the source – there’s not a lot of it. The only bit that I would say is perhaps “less than readable” is the AspectJ part – there’s not a lot to it, but it got complicated.
- Seriously, I’m not picking on Baeldung. It’s one of my favorite websites. Even though I didn’t love its examples in this instance, the article I referenced served as one of the starting points for this very blog post. ↩︎
- The Makefile contains targets that allow me to run commands to exercise the application more succinctly.
make
also has the wonderful feature of echoing commands that it runs. This allows me to type a short command and copy paste the output directly into this post – showing you both the commands that were actually run and the output that they generated. ↩︎ - This raises very interesting questions like “How does Spring’s Caching Interceptor handle cache updates from multiple threads?” That’s outside of the scope of this post, however, so I’m just going to “trust Spring” on this one. ↩︎