Re: [Ksummit-discuss] [CORE TOPIC] [nomination] Move Fast and Oops Things

[Dan Williams <dan.j.williams@intel.com>]

On Thu, May 22, 2014 at 9:31 AM, Dan Williams <dan.j.williams@intel.com> wrote:
> On Thu, May 22, 2014 at 8:48 AM, Theodore Ts'o <tytso@mit.edu> wrote:
>> On Wed, May 21, 2014 at 04:03:49PM -0700, Dan Williams wrote:
>>> Simply, if an end user knows how to override a "gatekeeper" that user
>>> can test features that we are otherwise still debating upstream.  They
>>> can of course also apply the patches directly, but I am proposing we
>>> formalize a mechanism to encourage more experimentation in-tree.
>>>
>>> I'm fully aware we do not have the tactical data nor operational
>>> control to run the kernel like a website, that's not my concern.  My
>>> concern is with expanding a maintainer's options for mitigating risk.
>>
>> Various maintainers are doing this sort of thing already.  For
>> example, file system developers stage new file system features in
>> precisely this way.  Both xfs and ext4 have done this sort of thing,
>> and certainly SuSE has used this technique with btrfs to only support
>> those file system features which they are prepared to support.
>>
>> The problem is using this sort of gatekeeper is something that a
>> maintainer has to use in combination with existing techniques, and it
>> doesn't necessarliy accelerate development by all that much.  In
>> particular, if it has any kind of kernel ABI or file system format
>> implications, we need to make sure the interfaces are set in stone
>> before we can let it into the mainline kernel, even if it is not
>> enabled by default.  (Consider the avidity that userspace application
>> developers can sometimes have for using even debugging interfaces such
>> as ftrace, and the "no userspace breakages" rule.  So not only do you
>> have to worry about userspace applicaitons not using a feature which
>> is protected by a gatekeeper, you also have to worry about premature
>> pervasive use of a feature such that you can't change the interface
>> any more.)
>
> I agree that something like this is prickly once it gets entangled
> with ABI concerns.  But, I disagree with the speed argument... unless
> you believe -staging has not increased the velocity of kernel
> development?
>
>> That by the way is the singular huge advangtage that centralized code
>> bases such as those found at Google and Facebook have --- if I need to
>> make a kernel change for some feature that hasn't made it upstream
>> yet, all of the users of some particular Google-specific kernel<->user
>> space interface is under a single source tree, and while I do need to
>> worry about staged deployments, I can be extremely confident that I
>> can identify all of the users of a particular interface, and put in
>> appropriate measures to update an interface.  It still might take
>> several release candences, but that's typically far shorter than what
>> it would take to obsolete a published upstream interface.
>
> Understood, but I'm not advocating that a system like this be used to
> support the Facebook/Google style kernel hacks to do things that only
> mega-datacenters care about.
>
>> As a result, I am much more willing to let a ugly, but operationally
>> necessary new feature (such as say a netlink interface to export
>> information about file system errors, for example) into an internal
>> Google kernel interface, but I'd be much less willing to let something
>> like that go upstream, because while it's annoying to have to forward
>> port such an out-of-tree patch, having to deal with fixing or
>> upgrading a published interface is at least an order or two more work.
>>
>> In addition, both Google and Facebook can afford to make changes that
>> only need to worry about their data center environment, where as an
>> upstream change has to work in a much larger variety of situations and
>> circumstances.
>>
>> The bottom line is just because you can do something at Facebook or
>> Google does not necessarily mean that the same technique will port
>> over easily into the upstream development model.
>
> Neil already disabused me of the idea that a "gatekeeper" could be
> used to beneficial effect in the core kernel, and I can see it's
> equally difficult to use this in filesystems that need to be careful
> of ABI changes.  However, nothing presented so far has swayed me from
> my top of mind concern which is the ability to ship pre-production
> driver features in the upstream kernel. I'm thinking of it as
> "-staging for otherwise established drivers".

Interesting quote / counterpoint from Dave Chinner that supports the
"don't do this for filesystems!" sentiment:

"The development of btrfs has shown that moving prototype filesystems
into the main kernel tree does not lead stability, performance or
production readiness any faster than if they stayed as an out-of-tree
module until most of the development was complete. If anything,
merging into mainline reduces the speed at which a filesystem can be
brought to being feature complete and production ready."

The care that must be taken with merging experiments is accidentally
leaking promises that you don't intend to keep to users.