Per-Entity Metadata Working Group

["Cantor, Scott" <>]

> Ah, so the concern is that a node goes away in the middle of serving
> a metadata request to the client. And, the client would not recover or
> retry the request?

Defining "in the middle" as between the time the client is given a particular 
address and the request is made or completes, yes. Browsers now fetch all A 
records and auto-failover to all of them, so if you have aggressive timeouts, 
you get what appears to be a reliable system when you're serving largely 
static content. Non-browsers don't, and in some cases probably can't, even if 
they wanted to. The model just doesn't work for "services" very well.

It may be that REST clients are starting to evolve to compensate for this, 
but my initial research doesn't suggest to me that either Java nor libcurl 
are doing this, so that covers Shibboleth.

-- Scott