http-cache-semantics

Can I cache this?

CachePolicy tells when responses can be reused from a cache, taking into account HTTP RFC 7234 rules for user agents and shared caches. It also implements RFC 5861, implementing stale-if-error and stale-while-revalidate. It's aware of many tricky details such as the Vary header, proxy revalidation, and authenticated responses.

Usage

Cacheability of an HTTP response depends on how it was requested, so both request and response are required to create the policy.

const policy = new CachePolicy(request, response, options);

if (!policy.storable()) {
    // throw the response away, it's not usable at all
    return;
}

// Cache the data AND the policy object in your cache
// (this is pseudocode, roll your own cache (lru-cache package works))
letsPretendThisIsSomeCache.set(
    request.url,
    { policy, response },
    policy.timeToLive()
);

// And later, when you receive a new request:
const { policy, response } = letsPretendThisIsSomeCache.get(newRequest.url);

// It's not enough that it exists in the cache, it has to match the new request, too:
if (policy && policy.satisfiesWithoutRevalidation(newRequest)) {
    // OK, the previous response can be used to respond to the `newRequest`.
    // Response headers have to be updated, e.g. to add Age and remove uncacheable headers.
    response.headers = policy.responseHeaders();
    return response;
}

It may be surprising, but it's not enough for an HTTP response to be fresh to satisfy a request. It may need to match request headers specified in Vary. Even a matching fresh response may still not be usable if the new request restricted cacheability, etc.

The key method is satisfiesWithoutRevalidation(newRequest), which checks whether the newRequest is compatible with the original request and whether all caching conditions are met.

Constructor options

Request and response must have a headers property with all header names in lower case. url, status and method are optional (defaults are any URL, status 200, and GET method).

const request = {
    url: '/',
    method: 'GET',
    headers: {
        accept: '*/*',
    },
};

const response = {
    status: 200,
    headers: {
        'cache-control': 'public, max-age=7234',
    },
};

const options = {
    shared: true,
    cacheHeuristic: 0.1,
    immutableMinTimeToLive: 24 * 3600 * 1000, // 24h
    ignoreCargoCult: false,
};

If options.shared is true (default), then the response is evaluated from a perspective of a shared cache (i.e. private is not cacheable and s-maxage is respected). If options.shared is false, then the response is evaluated from a perspective of a single-user cache (i.e. private is cacheable and s-maxage is ignored). shared: true is recommended for HTTP clients.

options.cacheHeuristic is a fraction of response's age that is used as a fallback cache duration. The default is 0.1 (10%), e.g. if a file hasn't been modified for 100 days, it'll be cached for 100*0.1 = 10 days.

options.immutableMinTimeToLive is a number of milliseconds to assume as the default time to cache responses with Cache-Control: immutable. Note that per RFC these can become stale, so max-age still overrides the default.

If options.ignoreCargoCult is true, common anti-cache directives will be completely ignored if the non-standard pre-check and post-check directives are present. These two useless directives are most commonly found in bad StackOverflow answers and PHP's "session limiter" defaults.

storable()

Returns true if the response can be stored in a cache. If it's false then you MUST NOT store either the request or the response.

satisfiesWithoutRevalidation(newRequest)

This is the most important method. Use this method to check whether the cached response is still fresh in the context of the new request.

If it returns true, then the given request matches the original response this cache policy has been created with, and the response can be reused without contacting the server. Note that the old response can't be returned without being updated, see responseHeaders().

If it returns false, then the response may not be matching at all (e.g. it's for a different URL or method), or may require to be refreshed first (see revalidationHeaders()).

responseHeaders()

Returns updated, filtered set of response headers to return to clients receiving the cached response. This function is necessary, because proxies MUST always remove hop-by-hop headers (such as TE and Connection) and update response's Age to avoid doubling cache time.

cachedResponse.headers = cachePolicy.responseHeaders(cachedResponse);

timeToLive()

Returns approximate time in milliseconds until the response becomes stale (i.e. not fresh).

After that time (when timeToLive() <= 0) the response might not be usable without revalidation. However, there are exceptions, e.g. a client can explicitly allow stale responses, so always check with satisfiesWithoutRevalidation(). stale-if-error and stale-while-revalidate extend the time to live of the cache, that can still be used if stale.

toObject()/fromObject(json)

Chances are you'll want to store the CachePolicy object along with the cached response. obj = policy.toObject() gives a plain JSON-serializable object. policy = CachePolicy.fromObject(obj) creates an instance from it.

Refreshing stale cache (revalidation)

When a cached response has expired, it can be made fresh again by making a request to the origin server. The server may respond with status 304 (Not Modified) without sending the response body again, saving bandwidth.

The following methods help perform the update efficiently and correctly.

revalidationHeaders(newRequest)

Returns updated, filtered set of request headers to send to the origin server to check if the cached response can be reused. These headers allow the origin server to return status 304 indicating the response is still fresh. All headers unrelated to caching are passed through as-is.

Use this method when updating cache from the origin server.

updateRequest.headers = cachePolicy.revalidationHeaders(updateRequest);

revalidatedPolicy(revalidationRequest, revalidationResponse)

Use this method to update the cache after receiving a new response from the origin server. It returns an object with two keys:

• policy — A new CachePolicy with HTTP headers updated from revalidationResponse. You can always replace the old cached CachePolicy with the new one.
• modified — Boolean indicating whether the response body has changed.
  • If false, then a valid 304 Not Modified response has been received, and you can reuse the old cached response body. This is also affected by stale-if-error.
  • If true, you should use new response's body (if present), or make another request to the origin server without any conditional headers (i.e. don't use revalidationHeaders() this time) to get the new resource.

// When serving requests from cache:
const { oldPolicy, oldResponse } = letsPretendThisIsSomeCache.get(
    newRequest.url
);

if (!oldPolicy.satisfiesWithoutRevalidation(newRequest)) {
    // Change the request to ask the origin server if the cached response can be used
    newRequest.headers = oldPolicy.revalidationHeaders(newRequest);

    // Send request to the origin server. The server may respond with status 304
    const newResponse = await makeRequest(newRequest);

    // Create updated policy and combined response from the old and new data
    const { policy, modified } = oldPolicy.revalidatedPolicy(
        newRequest,
        newResponse
    );
    const response = modified ? newResponse : oldResponse;

    // Update the cache with the newer/fresher response
    letsPretendThisIsSomeCache.set(
        newRequest.url,
        { policy, response },
        policy.timeToLive()
    );

    // And proceed returning cached response as usual
    response.headers = policy.responseHeaders();
    return response;
}

Yo, FRESH

Used by

• ImageOptim API, make-fetch-happen, cacheable-request (got), npm/registry-fetch, etc.

Implemented

• Cache-Control response header with all the quirks.
• Expires with check for bad clocks.
• Pragma response header.
• Age response header.
• Vary response header.
• Default cacheability of statuses and methods.
• Requests for stale data.
• Filtering of hop-by-hop headers.
• Basic revalidation request
• stale-if-error

Unimplemented

• Merging of range requests, If-Range (but correctly supports them as non-cacheable)
• Revalidation of multiple representations

Trusting server Date

Per the RFC, the cache should take into account the time between server-supplied Date and the time it received the response. The RFC-mandated behavior creates two problems:

• Servers with incorrectly set timezone may add several hours to cache age (or more, if the clock is completely wrong).
• Even reasonably correct clocks may be off by a couple of seconds, breaking max-age=1 trick (which is useful for reverse proxies on high-traffic servers).

Previous versions of this library had an option to ignore the server date if it was "too inaccurate". To support the max-age=1 trick the library also has to ignore dates that pretty accurate. There's no point of having an option to trust dates that are only a bit inaccurate, so this library won't trust any server dates. max-age will be interpreted from the time the response has been received, not from when it has been sent. This will affect only RFC 1149 networks.