On 05/09/2008, at 7:00 AM, Oleg Gusakov wrote:
I tried to create mercury-based wagon provider and got overwhelmed
with Wagon's strange architecture decisions and APIs.
Yeah, as James said, it kind of grew with Maven (refactoring of Maven
1.x artifact code and an aborted pre-m2 attempt I believe) and the
user facing API is not great at all.
First of all - inheriting from AbstractWagon and implementing
required methods took 20 min - excellent! The rest two days (are
still being) spent trying to understand:
- wagon-tests - I assume it's integration tests every provider has
to pass?
More or less, but not that well separated from the implementation
itself.
- why do tests requite provider to fire events, especially TRANSFER
event? What happens when provider has much higher up APIs and does
not control the transfer process? What if provider does not want to
share events - why mandate them by the tests ??
This one of those things that is tied in too closely, and I added more
to really confirm the current behaviour is retained rather than
thinking about it from first principles.
I don't think the tests check anything more than at least one transfer
event at the end for the total transferred (Which Maven uses for it's
progress counter).
- why tests rely on the Resource object, while it's not manipulated
anywhere in the APIs? Necessity to create on per each put or get
represents an object leak for me. Especially highly artificial
sequence of populating this object with data - set content length
before Initiated event and tests fail.
No disagreement here, this is a poor abstraction.
- what is the exact combination of Resource manipulation in the get
operation? I tried several and am still missing what should be
filled in when: resource (local file, I assume?) is empty when
firing getInitialized(), but tests fail, saying they expected
resource with a content length and timestamp. ... any help is
greatly appreciated!
I tried to make them consistent in the last release such that if the
remote server supports it, they should set length and modified between
initiated and started events.
If you are not inheriting StreamWagon, I'd suggest grabbing that and
retaining the same boilerplate.
- where did getFileList() come from?? HTTP provider, for example,
cannot do it without scrapping a page, and if repository has
different indexing formats/options - it will not work. And tests
fail I don't implement one. After all - this is why maven-
metadata.xml exists!
- where did resourceExists() come from? In HTTP provider it is not
possible without a try-and-fail cycle, which slows everything down -
bad choice. This is why maven-metadata.xml exists!
- why such fascination with getIfNewer() in the APIs? This notion is
external to a dumb http server and the suggested unix-like timestamp
may yield wrong results by as much as 24 hours, depending on the
server and client locations. This is why we use UTC timestamps in
the metadata files .. and that is why metadata files exist :)
Wagon wasn't designed to take maven-metadata.xml into consideration
(it predates the creation of metadata in the repository by some time),
it's outside of maven-artifact.
- Brett
--
Brett Porter
[EMAIL PROTECTED]
http://blogs.exist.com/bporter/
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
