Add "light speed" experiment code for SafeBrowsing service.

chrome/browser/safe_browsing/database_manager.cc

Index: chrome/browser/safe_browsing/database_manager.cc
diff --git a/chrome/browser/safe_browsing/database_manager.cc b/chrome/browser/safe_browsing/database_manager.cc
index 2f89931e50aebce92e15a1c38d6b39a257741cfc..dd247f65eca5c3b5c1432c97f3260b766e96443e 100644
--- a/chrome/browser/safe_browsing/database_manager.cc
+++ b/chrome/browser/safe_browsing/database_manager.cc
@@ -30,6 +30,7 @@
 #include "chrome/common/chrome_switches.h"
 #include "components/metrics/metrics_service.h"
 #include "components/startup_metric_utils/startup_metric_utils.h"
+#include "components/variations/variations_associated_data.h"
 #include "content/public/browser/browser_thread.h"
 #include "content/public/browser/notification_service.h"
 #include "url/url_constants.h"
@@ -650,6 +651,11 @@ void SafeBrowsingDatabaseManager::StartOnIOThread() {
   if (enabled_)
     return;
 
+  // TODO(asvitkine): Experimental code for measuring start up impact of SB.
+  // Remove when experimentation is complete. http://crbug.com/416219

[gab, 2015/02/12 16:57:00]

Wrong bug #, should be 450037, right?

[Alexei Svitkine (slow), 2015/02/12 17:27:12]

Done.

+  if (!variations::GetVariationParamValue("LightSpeed", "DisableSB").empty())

[gab, 2015/02/12 16:57:00]

So the experiment name is "LightSpeed" and it has a "DisableSB" group?

Shouldn't the experiment name be more specific to SB?

[Alexei Svitkine (slow), 2015/02/12 17:27:12]

This is using the parameters support in the variations framework.

The idea is a group can define parameters, so in this case, this is asking
whether the currently active group has a param "DisableSB". go/variations-params
for more details.

I figured we can use a single field trial for all of the "light speed"
experiments we're planning to do, with different params affecting different
settings.

This lets us not step on each others' toes when doing parallel experimentation
and also gives us flexibility of doing e.g. combination of different params.
WDYT?

[gab, 2015/02/12 17:35:42]

Oh very cool! SGTM

The one thing I'm worried about combining experiments (or otherwise having to
split 25/25/25/25 for two experiments) is that as we dilute the population into
smaller buckets, we'll lose statistical significance (and confidence intervals
may grow too big to prove anything).

Where as I was thinking that if we had two experiments E/F at 50/50, it's okay
for them to overlap as each group will get half of the effect of the other
experiment and it otherwise won't affect the significance of its individual
result (and we could theoretically dremel the combined impacts of {E}, {E,F},
{F}, {} though we wouldn't have those on the dashboards nor be able to easily
create special groups for a higher percentage of the population).

Lastly, do we get confidence intervals with params on the dashboard? (I assume
we do as it's just a group like any other, but just making sure as those will be
key)

[Alexei Svitkine (slow), 2015/02/12 17:47:56]

Right, it depends how many we want to run in parallel - indeed if too many, then
we will get smaller user populations. But we can see if that's actually a
problem and can always still do experiments in parallel if needed.

And yes, we get all the same confidence intervals with params on the dashboard -
as they're attached to regular groups.

[gab, 2015/02/12 18:26:58]

Okay SGTM.

+    return;
+
   DCHECK(!safe_browsing_thread_.get());
   safe_browsing_thread_.reset(new base::Thread("Chrome_SafeBrowsingThread"));
   if (!safe_browsing_thread_->Start())