On Thu, Apr 18, 2013 at 05:12:40PM +0300, Christos Vasilakis wrote:
For the Reddit case and for others that may come in the future which
may have "unknown" paging mechanisms, I followed the approach Android
does and introduced a PageExtractor interface. It's only purpose is to
parse the appropriate content and return us with the "next" and
"previous" links. Existing header/body mechanisms were refactored to
separate classes that implement the PageExtractor interface. Further
an additional configuration option in the PipeConfig was added
[setPageExtractor] that set the appropriate extractor class.
Nice! has this landed
on master or is it on a PR?
Important to state is that this config option is entirely optional.
Internally if not set, we will still use the appropriate page/header
mechanism (still depending on the metadataLocation config option).
Having this option will allow different and "exotic" page extractors
to be plugged in by the users if necessary or if not still rely on our
defaults.
Perfect - being flexible enough to accomodate _weird_ endpoints while
keeping sane defaults.
PLUS PLUS
--
qmx