On Thu, Dec 01, 2011 at 10:13:53PM -0500, Saggi Mizrahi wrote:
> I want to suggest a different way of handling constants.py and config.py
> the current way is clumsy and makes making convoluted code easy.
> GLOBAL VARIABLES ARE BAD. It's not something I invented. It's well 
> accepted that proper scoping yields better code.


> The way we use those two files shows how much of a mess globally 
> accessible variable can be.
> Not putting paths to executable in constants.py
> -----------------------------------------------
> Instead we should have a module for each utility executable. The module 
> should wrap the utility with a pythonic interface. That way there is no 
> reason for more then one module to have access to a utility's path.
> For example:
> cat.py
> ------
> def cat(files):
>     cmd = [_CAT_PATH] + files
>     out, err = exec(cmd)
>     return out
> Now there is no need to know where cat is. You have a proper interface 
> to use. Preventing people from re-wrapping the utility over and over. 
> This in turn means you have a canonical wrapper that is more likely to work.

I agree that our current behavior is clunky. Most external executable should be
wrapped by functions that exposes the functionality which we need from them.

I am not sure that having a module per executable makes the most sense; I'd
separate modularity based on Vdsm usage of them. Anyway, the only up side of our
current constants-aggregating approach, is that the number of files editted by
autoconf is limitted.

P.S. EXT_CAT is a horrible example. It exists only for letting Vdsm read
/etc/multipath.conf. This functionality must be wrapped properly and shoved into

> Every other constant
> --------------------
> Other constants should be moved to their proper modules and used only 
> when importing the module makes sense this will give you another step 
> where you make sure your decencies are correct.
> Configuration
> -------------
> Configuration files are just parameters to the program. They should be 
> treated as such. Their scope should not leave the main class of the 
> project. In our case VDSM is actually 2 projects meshed together like 
> Siamese twins so I think clientIf.py and hsm.py should be the only ones 
> with access to the config. Everything else should expect to get those 
> values in a regular programmatic fashion.
> In my opinion these changes will make VDSM a lot more flexible, will 
> make more parts independent and easily separated and at that end help us 
> make VDSM better for generations to come.

In my opinion our use of config.config is perfectly fine. The configuration file
is a single global object, and it makes sense to keep its Python representation
as a single global object. I'd like keeping it this way because if Vm migration
process needs a new parameter (say, when should we decide that migration is not
going anywhere) MigrationSourceThread can simply read it from the config. If we
take your approach, we would have to trickle the value all the way from clientIF
through Vm down to the specific function that needs it. It would clutter the
code and would not scale nicely with the number of configurational parameters.

> You can help!
> -------------
> * Start by taking you favorite utility and write a proper safe module 
> around it.
> * When writing new code try and not access the configuration directly.
> * Modules are free (really) don't just put stuff in misc\utils try and 
> put some order to things.
> _______________________________________________
> vdsm-devel mailing list
> vdsm-devel@lists.fedorahosted.org
> https://fedorahosted.org/mailman/listinfo/vdsm-devel
vdsm-devel mailing list

Reply via email to