Caching support

[bfg]

It would be totally bad-ass if AHC would support response cache, meaning that AHC would not fire up actual HTTP request if valid response would be already in the cache.

I think that it would be great to add such infrastructure to AHC code:

• AHC Cache interface with only no-op implementation
• Interface and default implementation that would evaluate whether response is cacheable
• Interface and default implementation that would map Request to cache key

Later we could add extra modules that would implement Cache interface using guava/caffeine/mapdb cache implementations.

I'm willing to contribute :-)

References:

• OkHttp cache support
• Apache HttpComponents cache support

[slandelle]

Here be dragons!

I'm willing to contribute :-)

Sure, go ahead! Feel free to ping here for discussing design.

Thanks!

[ben-manes]

Caffeine added support for variable expiration (see Expiry). This way entries can differ by their expiration times rather than relying on uniform, fixed timeout. If you are caching the responses then that would hold the expiration settings, which you would only evaluate on creation, and might look like:

LoadingCache<Request, Response> graphs = Caffeine.newBuilder()
    .expireAfter(new Expiry<Request, Response>() {
      public long expireAfterCreate(Request request, Response response, long currentTime) {
        long seconds = request.isCachable() ? request.getMaxAge() : 0L;
        return TimeUnit.SECONDS.toNanos(seconds);
      }
      public long expireAfterUpdate(Request request, Response response,
          long currentTime, long currentDuration) {
        return currentDuration;
      }
      public long expireAfterRead(Request request, Response response,
          long currentTime, long currentDuration) {
        return currentDuration;
      }
    })
    .build(request -> httpClient.call(request));

The feature is O(1) without relying on a maximum size, so you shouldn't observe much of a performance impact. Others use O(lg n) heaps or hide the entry and expect a maximum size be set (to lazily evict, e.g. redis/memcached). We do this using a timer wheel, like kernels do for timers.

[bfg]

@ben-manes yeah, that is nice, but i'm planning to use different design. I don't want to introduce parallel AHC api, but i want to add cache at AHC instance creation time, just okhttp and apache httpcomponents do.

And i want to make it cache implementation independent; and yes, cache implementation should be async as well in case somebody want's to plug in redis/hazecalst/whatever.

[ben-manes]

Right, I only meant it as an example that the feature was there. You could use a manual or async cache and compute using cache.get(key, func) or (a racy) cache.put(key, value). I didn't meant to imply that your strategy shouldn't be pluggable. Just that the capability is natively supported, whereas I think in Guava / MapDB you would need to emulate it using eviction, check the timestamp, and miss if expired.

[wsargent]

Also see playframework/playframework#7255

[slandelle]

@wsargent thanks for pinging. Where is EffectiveURIKey defined? I can't find it. I'm a bit skeptical if you key only account for the URI. You need to take the full request entity into account: method + URI + headers + request body.

[wsargent]

@slandelle https://github.com/playframework/play-ws/blob/master/play-ahc-ws-standalone/src/main/scala/play/api/libs/ws/ahc/cache/EffectiveURIKey.scala

You will also be interested in https://github.com/playframework/cachecontrol/blob/master/src/main/scala/com/typesafe/play/cachecontrol/CacheControl.scala

[wsargent]

You need to take the full request entity into account: method + URI + headers + request body.

the effective request uri is as defined in https://tools.ietf.org/html/rfc7230#section-5.5, from https://tools.ietf.org/html/rfc7234

[slandelle]

I won’t be able to check the other RFCs before a few days but this algorithm looks broken to me. It’s pretty common that servers response depends on request headers (Accept, Accept-Language, etc).

[wsargent]

Yes, a response has to be selected that matches the request header -- cached request and stored response is a huge part of the spec. That's covered in the ResponseSelectionCalculator.

Also the caching section of play-ws https://github.com/playframework/play-ws#caching and the cachecontrol README has more.

[domdorn]

any updates on this?
I'm currently implementing a Amazon SNS endpoint for a customer where we have to fetch certificates every time we receive a request (to verify its content) and being able to seamlessly cache this particular request (without spinning up an own caching solution) would be really nice.

[slandelle]

any updates on this?

Nope. Contributions welcome.