Re: [PATCH JGIT] Minor : Make ObjectId, RemoteConfig Serializable

From: Shawn O. Pearce <spearce@spearce.org>
Date: Feb 8, 2009, 19:10 UTC
Message-ID: <20090208191024.GA30557@spearce.org>
In-Reply-To: <320075ff0902080526g2bee8188g395397b06a0c80ee@mail.gmail.com>

Nigel Magnay <nigel.magnay@gmail.com> wrote:

Show 20 quoted lines

> > A problem (big problem) with serialization is that it often leads to
> > fragile interfaces. One might want to have precise control over
> > the serialization so a change in the implementation doesn't affect
> > compatibility. Serializing AnyObjectId should not depend on the
> > implementation de jour. Second, how do we handle subclasses?
> >
> > But maybe leaving it this way would be our way of saying that
> > the interface may break at any time, promise.
> 
> Well, we can of course implement writeObject / readObject (or do so
> if/when compatibility breaks, and it's cared about)
> 
> That's how I tend to view it anyway (may break at any time) - you
> can't just update a jar library to a significantly new version and
> expect it all to stay compatible. Also for half my use, it's not for
> persistence, it's for transferring over the wire to a slave process.
> 
> Thinking a bit more clearly, I probably don't need AnyObjectId, just
> ObjectId - but I've also missed RefSpec and URIish as they're used in
> RemoteConfig..

Here's my two cents... we can do this, but only if we implement Externalizable and do the read and write ourselves so we have a stable format.

In the case of any of the types you are discussing there is an easy canonical form for them to be written on the wire, or to read back:

  ObjectId     - the 20 byte SHA-1
  RefSpec      - the string form as it appears in the config file
  URIish       - the string form as it appears in the config file
  RemoteConfig - a map of keys/values as it appears in the config

-- 
Shawn.