> What I think we're understanding now is that something that fails 1% of the time, might not fail for just 1% of queries, it might fail for 1% of the users and work perfectly for the rest, and that small set can be really unhappy as a result.
Excellent insight. Linux kernel developers have had a similar experience with the use of the likely() and unlikely() hints on conditionals; it took a while to realize that unlikely() has to mean unlikely for any user, not just "unlikely for most users, and likely for the rest". So, for example, it doesn't make sense to write "if (unlikely(feature_disabled))", because anyone with the feature disabled will meet that condition every time.
I've definitely had the experience that Google's automatic correction does the wrong thing for me more often than not; I almost always end up clicking the link to go to the search results I actually asked for.
Send over the examples when you come across them (or if you have it turned on, try www.google.com/searchhistory/ to find queries you've issued in the past.) I'll make sure they get to the right place.
Excellent insight. Linux kernel developers have had a similar experience with the use of the likely() and unlikely() hints on conditionals; it took a while to realize that unlikely() has to mean unlikely for any user, not just "unlikely for most users, and likely for the rest". So, for example, it doesn't make sense to write "if (unlikely(feature_disabled))", because anyone with the feature disabled will meet that condition every time.
I've definitely had the experience that Google's automatic correction does the wrong thing for me more often than not; I almost always end up clicking the link to go to the search results I actually asked for.