Re: Reproducibility of builds

[Claudius Heine <claudius.heine.ext@siemens.com>]

Hi,

On 09/05/2017 12:05 PM, Alexander Smirnov wrote:
> 
> 
> On 08/28/2017 02:27 PM, Claudius Heine wrote:
>> Hi,
>>
>> On 08/21/2017 01:23 PM, [ext] Claudius Heine wrote:
>>> Hi,
>>>
>>> On 08/03/2017 10:13 AM, Claudius Heine wrote:
>>>> Hi,
>>>>
>>>> am I right that Isar supports or should support reproducible root 
>>>> file system build?
>>>>
>>>> If I understand correctly, when multistrap is called, it fetches 
>>>> always the latest version of all packages from the debian repository 
>>>> mirrors. Am I mistaken or is this feature still on the roadmap?
>>>>
>>>> I that is on the roadmap, how are you thinking of solving this issue?
>>>>
>>>> The openembedded way would be to seperate the fetch and 'install' 
>>>> step and first download all packages into the DL_DIR and then use 
>>>> them from there. Maybe we could create this pipeline:
>>>>
>>>> dpkg-binary Recipe:
>>>>
>>>> fetch deb file into downloads -> insert into local repository
>>>>
>>>> dpkg-source Recipe:
>>>>
>>>> fetch sources into downloads -> build packages -> insert into local 
>>>> repository
>>>>
>>>> image Recipe:
>>>>
>>>> fetch all required packages into downloads -> insert all of them 
>>>> into the local repository -> create root fs using only the local 
>>>> repository
>>>>
>>>> Multistrap provides a '--source-dir DIR' parameter, that stores all 
>>>> installed packages into a directory. So if we would use that as a 
>>>> fetcher, then we would create a temporary rootfs just to get all 
>>>> required packages for the project.
>>>>
>>>> Are there other possible solutions for this?
>>>
>>> The problem with this solution is that its not possible to create 
>>> multiple images with different sets of packages that share the 
>>> version of the all the common packages.
>>>
>>> An alternative solution is to employ a repository cacher that caches 
>>> the 'Packages.gz' of the first request. This way it would also be 
>>> faster then running multistrap one additional time just to fetch all 
>>> required packages.
>>>
>>> Maybe apt-cacher-ng or something similar can be used for this.
>>> However I am currently not sure how this can be integrated into the 
>>> current build process. Some ideas? Maybe implementing a simple repo 
>>> caching proxy that is integrated into isar?
>>>
>>> The repository cacher is likely a daemon running in parallel to 
>>> multistrap and fetches everything to the DL_DIR that is requested by 
>>> it. Maybe provide a 'clean_package_cache' task, that deletes the 
>>> cached 'Packages.gz', causing the next root fs build to use new 
>>> package versions.
>>>
>>> I would really like to hear some feedback on this.
>>
>> In our meeting today, it was discussed that we should collect all 
>> requirements for this feature and discuss possible implementation 
>> ideas based on those requirements.
>>
>> Here are some requirements from my side:
>>
>>      1 If multiple different images with some common set of packages are
>>        build with one bitbake call, then all images should contain
>>        exactly the same version of every package that it has in common
>>        with any of the other images.
>>
>>      2 The resulting image should only depend on the build environment
>>        and isar metadata, not on the point in time it is build.
>>        This means if the environment, including the downloads directory,
>>        is complete (for instance by an earlier build of the image), every
>>        following build of this image recipe should result in exactly the
>>        same packages installed on this image.
>>
>>      3 Binary and source packages should be part of the archival process.
>>        Source packages are useful in case some package needs to be
>>        patched at a later date. Binary packages are useful, because
>>        building them from source packages is currently not 100%
>>        reproducible in Debian upstream. [1]
>>
>>      4 For development, it should be possible to easily reset the
>>        environment, triggering an upgrade of the packages on the next
>>        image build.
>>
>>      5 Deployable in CI environments. What those are exactly should be
>>        further discussed. Here are some:
>>
>>        5.1 Possibility to use a download cache, that is not bound to only
>>            one product/image/environment
>>
>>        5.2 More than one build at the same time in one environment should
>>            be possible
>>
>>      6 Efficiency: The reproducibility feature should be time and
>>        resource efficient as possible. E.g. Process should only fetch and
>>        store the required files.
>>
>>      7 Outputs a description file with the name and version of every
>>        package deployed/used in the image/environment.
>>
>> To 5: Since I don't have much experience with CI systems, requirements 
>> mentioned here might not be correct.
>>
>> Any comment or requirement additions are welcome.
> 
> Thank you for the requirements, they quite good describe your usecase. 
> Unfortunately, ATM I don't know all the capabilities of 
> multistrap/debootstrap, so could not propose too much.
> 
> In general, I think there could be following solutions:
> 
>   - Create local apt cache with specified packages versions.
>   - Patch multistrap to add capabilities to specify package versions.
>   - Add hook to multistrap hooks (for example, in configscript.sh), that 
> will re-install desired package versions via apt-get.

My solution is a bit different and does not require patching multistrap, 
should also work with other bootstraping mechanism. (AFAIK it should be 
possible to change the bootstraping mechanism at a later date, since the 
multistrap project is dead.)

I started implementing a http proxy in python, that caches all requests 
of '/pool/' and '/dists/' uris in seperate directories. 'dists' is part 
of the build environment while 'pool' contains all the packages and 
should be part of the download directory.

I am currently not actively working on this proxy, because of other 
tasks with higher priority, but I can give you access to it if you like.

On my TODO list for this is:

   - Port to asyncio (with a simple http implementation)
     This proxy is currently single threaded and can only handle one
     connection at a time. Porting to asyncio is possible, but since the
     python standard library does not provide a http implementation based
     on asyncio a small http implementation based on this has to be
     implemented as well.
   - Integrate into bitbake/isar as a scripts/lib and a bbclass
     To ease early development I implemented this proxy outside of
     bitbake, but with the idea to integrate it into bitbake at a later
     date. It should be easily doable to integrate this into bitbake via
     two tasks. One that starts the proxy, and one that shuts it down.
     Maybe add a shutdown via a bitbake event as well, so that it will be
     shut down regardless of the tasks handled. Or do it completely via
     bitbake events.

The current proxy limits repositories to the http protocol. But maybe 
its possible to have https proxies as well, but there its necessary to 
break the ssl chain.

Claudius

-- 
DENX Software Engineering GmbH,      Managing Director: Wolfgang Denk
HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany
Phone: (+49)-8142-66989-54 Fax: (+49)-8142-66989-80 Email: ch@denx.de