But I do this from my own Pinboard account, which has
superpowers. Specifically, it lets me see private bookmarks
on all accounts. And since I already see everything, I don't
notice when everybody else starts to see everything, too.
I can't think of a way to store retrievable data, like a bookmark, on his servers without leaving some way for him to access the data if he wanted to.
1) It's trivial for him to inadvertently see something deeply personal to someone just by browsing the 'recent' list or doing a search.
UPDATE: I overstated this one - Maciej let me know by email that he can only access private data on the search / recent page if he intentionally masquerades a user. He can only inadvertently see private data when viewing individual user pages.
2) If his account's ever compromised (let's hope he's not reusing that password elsewhere!) then someone else gets that ability as well, accessible from any browser anywhere.
It's one thing when you have to ssh into a server somewhere and do a SQL query to access someone's private information. It's another thing to set up your admin account so you're casually exposed to it.
I like Pinboard's service too, but this isn't remotely cool.
The choice, really, is "do I look at my user's personal stuff and have a good idea that my service is working for them" vs. "do I not look at my user's stuff, and have much less of an idea if the service is working for them."
(yes, yes, good automated tests are the right way to solve the problem, but good automated test are difficult, especially when a broad range of behaviors could be 'correct')
In my own case, there are two places where I'm erroring on the "don't look at the user's data, even if it would allow you to provide better service" side. First? I don't mount a user's disk. it's a block device, as far as I am concerned.
As you might imagine, this makes backups way, way more difficult. Moves, too. It's possible that if I were to ignore this rule, I'd have enough spare disk bandwidth to do at least half-assed backups (I'm not doing any backups of customer data right now, which is pretty scary for me, because I'm in a business where if you lose the customer's data, you lose the customer in the worst way.)
Now, the right way to solve this problem is some kind of 'snapshot over the network' like zfs has or some other mechanisim that only transfers the change in block devices. Unfortunately, Linux has no such tools. (yes, yes, ZFS on linux is a possibility. So is a NAS. Rsync or something rsync like won't work because the problem is not network bandwidth but disk bandwidth. I mean, it's a problem that can be solved, a better way, but solving it the better way is more work.)
Next, from a technical support perspective? I could provide dramatically better service if I logged all the serial consoles. Dramatically better service. but last time I asked, people seemed uncomfortable with it, so I don't. (the thing is, it's only easy to log all consoles or log none of the consoles; I'd have to spend time and effort building a 'log consoles except for users that opt out' mechanisim, which is the right thing to do here, but due to time constraints that hasn't been done.) (to be clear, when a customer asks for help, or even if soemthing happens with a domain and I'm not sure it's working, I look at the serial console as is. My job would be nigh impossible without serial consoles. Also, by default I log the serial consoles on dedicated servers, though unlike Xen, it is easy for me to opt people out of that. Conserver is the tool I use there.)
Personally, I think I made the right choice when it comes to block devices (I have had good luck with my RAIDs, and treating disks as block devices means my customers can use whatever freaky filesystem they like.) but I think I made the wrong choice on the serial console; there really isn't that much personal stuff that comes up on the serial console, and it gives me a lot of clues.
I guess what I'm saying is that from a testing perspective? being able to see the user's bookmarks has a lot of advantages. It's a valid choice; and he's sharing that choice with the customers, which is the right thing to do.
(the interesting thing here, of course, was that if his account couldn't see the users bookmarks, while usually that would have given him much less diagnostic information, it would have given him more diagnostic information this time.)
I think the key really is communication; letting the user know what you are doing and reading the feedback. It wouldn't occur to me that some people would think serial console logs are super private things, but turns out? many people think they are.
I mean, from that perspective, it seems like pinboard ought to take this feedback into consideration and maybe change that policy. I'm just saying that there were reasonable reasons why that original choice was made, even if it sounds like the decision may have been the wrong one.