Re: Security patching and the branching workflow: a new security-updates branch

[zimoun]

Hi Léo,

On Fri, 26 Mar 2021 at 21:10, Léo Le Bouter <lle-bout@zaclys.net> wrote:

> For these reasons, I would like to propose a new branch called
> security-updates that would be based on master where we queue security
> fixes that introduce any arbitrary number of rebuilds without using
> grafts.
>
> We would merge the security-updates branch as soon as there is complete
> substitute availability for the branch and it's future merged version
> within master.

Basically, what you are proposing is more or less staging.  The issue is
that the cycles are sometimes longer than expected.  The cycle is
defined as 6 weeks for staging.  Is it really 6 weeks?  Well, It could
be shorter if there is more man power *and* if the build farm is working
better.

About the man power, it is already a good job.  Let plot roughly the
number of days between 2 merges between staging and master.

--8<---------------cut here---------------start------------->8---
$ git log --format="%cd|%s" --date=format:"%Y-%m-%d" --after=2019-01-01 \
      | grep Merge | grep staging | grep master                         \
      | cut -d'|' -f1                                                   \
      > /tmp/staging.txt 
--8<---------------cut here---------------end--------------->8---

then this tiny R script:

--8<---------------cut here---------------start------------->8---
$ cat /tmp/dates.R
library("ggplot2")

raw <- read.table("/tmp/staging.txt")
dat <- raw$V1
dates <- data.frame(start=strptime(dat[1:length(dat)-1], format="%Y-%m-%d"),
                    end=strptime(dat[2:length(dat)], format="%Y-%m-%d"))
dates$diff <- difftime(dates$start, dates$end, units="days")

p <- ggplot(dates, aes(x=start, y=diff)) + geom_point(color='red') + 
geom_smooth()
ggsave("/tmp/number-of-days-staging-merges.pdf")
--8<---------------cut here---------------end--------------->8---

shows the attached graph.  Well, it is variable and the variance and
standard deviation should be considered.  The mean is less then 8 days.

Obviously, I can have done a mistake and this should be taken with grain
of salt.  My point is: we should root the practise on what we are
already doing and see what are the incremental changes to improve.

For example, staging is not merged since a couple of months. Why?
Because none of us is taking the time to make it.  Hard to be close to
the oven and in the same to the mill. :-)

About the build farm, it is doing better and better.  Mathieu introduced
some plots [1] and again I think we should explore a bit the data to see
what is the average time to build staging, how many packages are built
on average, etc.

Moreover, priorities have recently been introduced.  I do know if we
have enough time elapse to see their impacts.  They are per branch,
IIRC, and maybe we could imagine a priority ’properties’ applied for a
package because it is a security update.  IIUC, it is not implemented
yet.  And as I said elsewhere, “to me, security is important. But it's
no less important than everything *else* that is also important!“, so
personally I am not convinced that security updates deserve a special
treatment compared to a regular update.  That’s my opinion. :-)

That’s said, from my point of view, staging or security-updates are only
names but somehow they cover the same idea: update a package with many
dependants and provide the substitutes as soon as it is available.

The key for a better scaling is, IMHO, to tweak what is going to master,
staging and core-updates; as I wrote in [2].

But the real hard part is to collectively make this tough work of
merging branches on the long run.


Cheers,
simon

1: <https://ci.guix.gnu.org/metrics>
2: <86mtv29erk.fsf@gmail.com">https://yhetil.org/guix/86mtv29erk.fsf@gmail.com>

number-of-days-staging-merges.pdf
Description: graph.pdf