Re: [PATCH v2 2/7] maintenance: store the "last run" time in config

From: Derrick Stolee <stolee@gmail.com>
Date: Aug 27, 2020, 13:02 UTC
Message-ID: <c30d745f-4d30-0404-420f-2009870da95b@gmail.com>
In-Reply-To: <xmqqzh6huyoi.fsf@gitster.c.googlers.com>

On 8/26/2020 1:03 PM, Junio C Hamano wrote:

Show 20 quoted lines

> Derrick Stolee <stolee@gmail.com> writes:
> 
>>>>  1. The time the task takes to execute should not contribute to the
>>>>     interval between running the tasks.
>>>
>>> ... as long as the run time is sufficiently shorter than the
>>> interval, that is.  If a task takes 10-30 minutes depending on how
>>> dirty the repository is, it does not make sense to even try to run
>>> it every 15 minutes.
>>
>> Definitely. The lock on the object database from earlier prevents these
>> longer-than-anticipated tasks from stacking.
> 
> Hmph, I actually was (anticipating|hoping) that you would give a
> good argument for having maintenance subsystem in change of
> scheduling rather than cron, as it can monitor how the already
> running job is goind and skip one cycle if needed.  The above is
> instead a good argument that independent cron jobs can still
> coordinate and there is no central and custom scheduler in the form
> of 'maintenance run'.

While the lock does prevent concurrent 'maintenance run' commands from colliding and causing unpredictable behavior as they both modify the object database, this does not help ensure that maintenance tasks actually happen if certain tasks are fired independently by cron and consistently collide.

This is the main motivation for me using a single crontab entry. More discussion of all of the tradeoffs is in [1].

[1] https://lore.kernel.org/git/bd4e18b7-6265-73e7-bc1a-a7d647eafd0a@gmail.com/

Show 14 quoted lines

>>>>  2. If the task fails for some unforseen reason, it would be good to
>>>>     indicate that we _attempted_ the task at a certain timestamp. This
>>>>     will avoid spamming a repository that is in a bad state.
>>>
>>> Absolutely.
> 
> Somebody already mentioned that using the configuration file for
> recordkeeping may not be a good idea, and I tend to agree, by the
> way.  I may want to periodically take a snapshot of my configuration
> to notice and remember changes I made myself intentionally
> (e.g. switched access method of a hosting site from ssh:// to
> https://, added a new branch that builds on something else, etc.) by
> comparing the snapshot with previous ones (and might even put it
> under version-control) and mechanical noise would interfere with it.

 
I will think of another way to handle this, then. If we cannot infer
that "this task was launched, therefore it is due to run" from an
optimal cron schedule, then I'll probably create a new file in the
.git repository that stores these values. That file would be in the
config format to make parsing easy.

Thanks, -Stolee