From: bwood (brass_at_[hidden])
Date: 2006-05-17 23:54:51
This is a MIME encoded message.
Peter Dimov writes:
> bwood wrote:
>> David Abrahams wrote:
>>> bwood <brass_at_[hidden]> writes:
>>>> Along the lines of what Lapin is thinking, I think B.Ser's
>>>> support is problematic. For example, consider wanting to move (or
>>>> delete) a data member out of a type between versions 1.1 and 1.2.
>>>> B.Ser needs that member to remain in the type in order to support
>>> No, you just read it and throw it away.
>>> if (version == 11)
>>> type_of_deleted_member ignored;
>>> ar & ignored;
>> Wait a minute. The code above looks like you may have focused on
>> deleting a member and not relocating it. "ignored" isn't coming
>> from anywhere else with a 1.1 client and it needs to be stored
>> somewhere. Also if a 1.2 server is sending to a 1.1 client,
>> it doesn't matter whether the member has been deleted or
>> relocated, it still has to send the right thing back.
> This is an inherent property of the communication between an 1.2
> an 1.1 client; any versioning system would have this problem. If you
> want to
> talk to 1.1 code, you have to be able to produce 1.1 data. This is
> if the underlying serialization layer doesn't support versioning at
> just makes it harder for you.
"Any versioning system would have this problem."
There's an alternative approach though that doesn't have the
1.1 1.2 1.3 1.4
class name Account Account Account_13 Account_13
The class name incorporates the release in which it was last
changed. As usual, the 1.3 version of the server is required
to support 1.1 clients. To do this the server is built with
both the Account and Account_13 definitions. Sorry for being
obvious, but it uses Accounts when talking to 1.1 clients and
Account_13s when working with 1.3 clients.
When only additions occur between releases, I think this
hierarchy would work
If members need to be relocated out of Account in 1.3, the
hierarchy might be
In either case, the 1.3 server may use polymorphism and
accomplish what is needed pretty simply. With a 1.1 client,
it produces 1.1 data by executing what originally was 1.1
code. This is helpful from a testing perspective. The only
way I can think of to introduce errors with 1.1 clients would
be to accidentally use Account_13 instead of Account.
I think this approach supports relocated members between
releases and could be used to build systems that are easier
to test than systems that use "versioning."
PS: I don't have Le Chaud Lapin's email address. I think he
might be interested in this discussion.
Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk