Increase the default cache time for failed host resolves from 1 second to 1 m...

Increase the default cache time for failed host resolves from 1 second to 1 minute.

BUG=1672
Committed: http://src.chromium.org/viewvc/chrome?view=rev&revision=43997

[darin (slow to review), 2010-04-07 05:44:56]

Wouldn't another solution to the PAC problem be to use a NXDOMAIN cache that
is not time based, but instead scope the lifetime of NXDOMAIN cache entries
to the PAC script invocation?

Before calling into the PAC script, set a flag that says to cache NXDOMAIN
results.  Upon exit from the PAC script, clear that flag and flush the
cache.  Alternatively, post a delayed task to flush the cache.  That way if
the PAC script gets invoked frequently, the NXDOMAIN cache would persist
over several PAC script invocations.

This just seems superior to increasing the DNS cache TTL for NXDOMAIN
results since the user has no way to clear that cache.  We could look into
making shift+reload bypass the DNS cache.

-Darin



On Tue, Apr 6, 2010 at 12:45 PM, <eroman@chromium.org> wrote:

> Reviewers: darin,
>
> Description:
> Increase the default cache time for failed host resolves from 1 second to 1
> minute.
>
> Details:
>
> This matches Firefox's default, and since our other cache numbers match
> Firefox
> seems reasonable to keep this one in sync.
>
> The intent of this change is to speed up PAC proxy execution. I had
> originally
> added the 1 second negative entries cache exactly for this case, and
> intended to
> revisit the exact policy later.... that time is now.
>
> One of the situations that can happen, is when you request a webpage with a
> bogus host, the PAC script may do repeated DNS resolves for this bogus
> host. And
> often, these resolves of non-existent hosts are very costly (multiple
> seconds).
>
> For for PAC scripts which do lots of DNS resolves (i.e. they call
> isInNet(hostname, ...) repeatedly, this quickly adds up.
>
> For example, for one user I saw it take 2+ seconds for each cache miss, and
> a
> total of 10 cache misses made this add up to 20+ seconds.
>
> With the larger cache expiration, this same example would just take 2
> seconds.
>
> Lastly, note that the problem became more pronounced the "universal
> intranet
> redirector" tests on startup (http://codereview.chromium.org/525079) were
> added,
> since now there are at least 3 bogus hosts that get requested shortly after
> each
> startup; and due to single-threadedness (issue 11079) , can stall the other
> requests for over a minute.
>
> BUG=1672
>
> Please review this at http://codereview.chromium.org/1576021
>
> SVN Base: svn://chrome-svn/chrome/trunk/src/
>
> Affected files:
>  M     net/base/host_resolver_impl.cc
>
>
> Index: net/base/host_resolver_impl.cc
> ===================================================================
> --- net/base/host_resolver_impl.cc      (revision 43642)
> +++ net/base/host_resolver_impl.cc      (working copy)
> @@ -38,7 +38,7 @@
>   HostCache* cache = new HostCache(
>       kMaxHostCacheEntries,
>       base::TimeDelta::FromMinutes(1),
> -      base::TimeDelta::FromSeconds(1));
> +      base::TimeDelta::FromMinutes(1));
>
>   return cache;
>  }
>
>
>

[eroman, 2010-04-07 05:49:56]

While I could do a solution which is specific to the PAC case, I think we will
be better served by simply increasing the timeout; we have a much lower
threshold than other browsers, and this sort of problem can be hit in other
places.

Note that we already will bypass the host cache when you hit refresh (although
not for the PAC case).

[darin (slow to review), 2010-04-08 04:14:30]

LGTM... I did not realize we bypassed the host cache when you hit refresh.
-Darin

On Tue, Apr 6, 2010 at 10:49 PM, <eroman@chromium.org> wrote:

> While I could do a solution which is specific to the PAC case, I think we
> will
> be better served by simply increasing the timeout; we have a much lower
> threshold than other browsers, and this sort of problem can be hit in other
> places.
>
> Note that we already will bypass the host cache when you hit refresh
> (although
> not for the PAC case).
>
>
> http://codereview.chromium.org/1576021
>

[heartforyourvalentine, 2010-04-14 10:56:56]

Error 105 (net::ERR_NAME_NOT_RESOLVED): browser messages are now cached for one
minute. Every time this dns error is displayed one need to wait >60 seconds
until reload/f5/shift-f5 is honored again.