When should implementations serialize caches? #362

gavinp · 2014-07-07T14:41:24Z

Currently, cache.add() is described in the spec as being placed in the [[RequestToResponseMap]] when the response first resolves. There's no way to use cache.put() earlier than response resolution, since the spec doesn't permit passing a promise to the contructor for the response argument.

When the response resolves is somewhat implementation dependent; but it can be as early as when headers are received for the final, non-redirect response for the request; since the data is in a blob behind a promise, the response can be resolved before the full data is received.

When should a Cache API implementation be fully committed to disk?

A common pattern of use is likely to be that a ServiceWorker creates a new cache, loads a series of resources in to it, and moves the new cache on top of the existing cache as a replacement. Example:

var newCache = caches.add("newOfflineCache");
var toLoadRequests = ...;
Promise(toLoadRequests.map(newCache.add, newCache)).then(function() {
  return caches.rename("newOfflineCache", "OfflineCache");
});

But this rename isn't quite safe, is it? The requests are only guaranteed to have received headers, not bodies, and so if we do the rename, we risk partial/erroneous/missing responses in the renamed cache?

I am left with a few questions:

When should implementations finalize serializing an entry to disk, so it is likely to be found in subsequent uses of that Cache?
Should caches.rename() and others wait for all in-flight requests to reach this point before resolving?
Why isn't this mechanism exposed as part of the Cache API?

gavinp · 2014-07-07T14:51:39Z

(Is this a duplicate of issue #313 ?)

jakearchibald · 2014-07-10T14:35:25Z

If a server gives us <body> Hello wo, can we tell it was intending to send more? I was under the impression we couldn't, that's why "successful headers" as the resolve point.

A common pattern of use is likely to be that a ServiceWorker creates a new cache, loads a series of resources in to it, and moves the new cache on top of the existing cache as a replacement.

I think it's useful to be able to do that, but I'm not sure it's common. I've been going with:

During install, create caches named 'mysite-v2-stuff', 'mysite-v2-otherstuff' & populate
During activate, remove caches named 'mysite-v[^2][^-]'

…although I guess the rename way becomes more useful for full cache replacements outside of the serviceworker update process.

jakearchibald · 2014-07-15T09:56:31Z

Done a bit of research around this:

Loading a page with Content-Length < actual content length

Chrome: Truncated content - no indication of error
Firefox: Truncated content - no indication of error
Safari: Truncated content - no indication of error
IE: Truncated content - no indication of error

Loading a page with Content-Length > actual content length

Chrome: Rendered received content, spinner until timeout, ERR_CONTENT_LENGTH_MISMATCH & ERR_CACHE_MISS in console
Firefox: No in-progress render, spinner until timeout, rendered received content, no console errors
Safari: No in-progress render, spinner until timeout, no final render, "Failed to load resource: The network connection was lost." in console
IE: No in-progress render, spinner until timeout, rendered received content, no console errors

Loading a page with Content-Length > actual content length ending in abrupt server termination

Chrome: Received content rendered, ERR_CONTENT_LENGTH_MISMATCH in console
Firefox: Received content rendered, no errors
Safari: No content render "Safari can’t open the page “‎localhost:3000/app-name/” because the server unexpectedly dropped the connection."
IE: Received content rendered, no errors

Loading a page with no Content-Length, chunked encoding & abrupt server termination

Chrome: Received content rendered, ERR_INCOMPLETE_CHUNKED_ENCODING in console
Firefox: Received content rendered, no errors
Safari: Received content rendered, no errors
IE: Received content rendered, no errors

XHR a page with Content-Length < actual content length

Chrome: load event, truncated content
Firefox: load event, truncated content
Safari: load event, truncated content
IE: load event, truncated content

XHR a page with Content-Length > actual content length

Chrome: waits until timeout, error event, ERR_CONTENT_LENGTH_MISMATCH in console
Firefox: waits until timeout, load event, received content delivered
Safari: waits until timeout, error event, "Failed to load resource: The network connection was lost." in console
IE: waits until timeout, load event, received content delivered

XHR a page with Content-Length > actual content length ending in abrupt server termination

Chrome: error event, ERR_CONTENT_LENGTH_MISMATCH in console
Firefox: load event, received content delivered
Safari: error event, "Failed to load resource: The network connection was lost." in console
IE: load event, received content delivered

XHR a page with no Content-Length, chunked encoding & abrupt server termination

Chrome: error event, ERR_INCOMPLETE_CHUNKED_ENCODING in console
Firefox: load event, received content delivered
Safari: load event, received content delivered
IE: load event, received content delivered

astanley · 2014-07-16T18:10:21Z

Amazing summary, Jake.

dubravkoL · 2014-07-17T07:17:54Z

Thanks a lot! ;)

Spudley · 2014-07-17T08:10:32Z

Useful bit of research. Thank you. One question: Are these results consistent across browser versions or did you just test the latest version of each?

jakearchibald · 2014-07-17T08:11:18Z

Only looked at the latest stable versions of each

SKing7 · 2014-07-17T10:42:39Z

cool，how do you test these?

jakearchibald · 2014-08-04T10:01:57Z

@SKing7 just a little nodejs server sending the headers & data

jakearchibald · 2014-08-05T15:47:33Z

Update: Latest thinking is to resolve on full stream read, allowing for rejection on detectable errors

wanderview · 2014-09-23T20:38:36Z

Update: Latest thinking is to resolve on full stream read, allowing for rejection on detectable errors

Just to clarify, how does this effect a Match() while the Put() stream is still in process? Can a Match() for a request/response pair succeed prior to the Put() promise resolving?

// ignore returned promise
cache.put(request, fetchResponse);

// match request before put() would have resolved.  Should this reject or
// resolve with clone of incomplete fetchResponse above?
cache.match(request);

I think my first implementation will have the Match() reject in this case.

KenjiBaheux · 2014-10-28T19:08:32Z

Talked to Gavin, this issue covers rather convoluted edge cases and does not qualify as "impact MVP"

KenjiBaheux · 2014-10-28T19:10:01Z

We'll go with the current thinking. Tracked via crbug.com/392621 for Blink.

jungkees · 2015-07-07T06:53:22Z

I think we finally settled on this behavior: #573 (comment). (See also #361 (comment)). If there's no more concern, I'd be happy to close the related issues.

jungkees · 2015-07-08T01:47:23Z

Closing.

jakearchibald mentioned this issue Jul 31, 2014

Review FetchStores & Cache changes #372

Closed

KenjiBaheux added cache labels Aug 5, 2014

KenjiBaheux added decided and removed impacts MVP high risk labels Oct 28, 2014

jakearchibald mentioned this issue Dec 4, 2014

Can caches store and retrieve results that are still being written to the cache? #573

Closed

jakearchibald mentioned this issue Mar 27, 2015

Aborting a fetch whatwg/fetch#27

Closed

jungkees closed this as completed Jul 8, 2015

This was referenced Jul 8, 2015

Can Cache.add() committing at response resolution lead to duplicate requests? #361

Closed

what should Cache put operations do if the body errors out after resolving promise? #665

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

When should implementations serialize caches? #362

When should implementations serialize caches? #362

gavinp commented Jul 7, 2014

gavinp commented Jul 7, 2014

jakearchibald commented Jul 10, 2014

jakearchibald commented Jul 15, 2014

astanley commented Jul 16, 2014

dubravkoL commented Jul 17, 2014

Spudley commented Jul 17, 2014

jakearchibald commented Jul 17, 2014

SKing7 commented Jul 17, 2014

jakearchibald commented Aug 4, 2014

jakearchibald commented Aug 5, 2014

wanderview commented Sep 23, 2014

KenjiBaheux commented Oct 28, 2014

KenjiBaheux commented Oct 28, 2014

jungkees commented Jul 7, 2015

jungkees commented Jul 8, 2015

When should implementations serialize caches? #362

When should implementations serialize caches? #362

Comments

gavinp commented Jul 7, 2014

gavinp commented Jul 7, 2014

jakearchibald commented Jul 10, 2014

jakearchibald commented Jul 15, 2014

Loading a page with Content-Length < actual content length

Loading a page with Content-Length > actual content length

Loading a page with Content-Length > actual content length ending in abrupt server termination

Loading a page with no Content-Length, chunked encoding & abrupt server termination

XHR a page with Content-Length < actual content length

XHR a page with Content-Length > actual content length

XHR a page with Content-Length > actual content length ending in abrupt server termination

XHR a page with no Content-Length, chunked encoding & abrupt server termination

astanley commented Jul 16, 2014

dubravkoL commented Jul 17, 2014

Spudley commented Jul 17, 2014

jakearchibald commented Jul 17, 2014

SKing7 commented Jul 17, 2014

jakearchibald commented Aug 4, 2014

jakearchibald commented Aug 5, 2014

wanderview commented Sep 23, 2014

KenjiBaheux commented Oct 28, 2014

KenjiBaheux commented Oct 28, 2014

jungkees commented Jul 7, 2015

jungkees commented Jul 8, 2015