As a follow on, I've already pushed an experiment using the <link>
method I discussed. We're currently considering other libraries to help
manage dependencies, too (cf., RequireJS).
On 2012-06-05 14:31 , David Karger wrote:
Ryan, of the scenarios you propose, I'd advocate the principle that the
user should never need to write js just to load extensions. I don't
know enough about labjs to be sure but I think this rules out direct use
of labjs, and possibly the exhibt-specific version. On the other hand,
if there were a way to use labjs "invisibly", embedding its invocations
inside the exhibit api and extensinos, it would be great.
LABjs has been in use within the Exhibit 3 codebase since it was
started. The problem with merely embedding it in Exhibit and then
allowing the specifying of extensions outside of that scheme is [see all
of previous message].
An option you didn't mention would be to incorporate a "just in time"
approach to the extensions. As you say, exhibit runs into problems when
if it starts running before all extensions have loaded. But more
specifically, the problem only occurs when it tries to use the
functionality offered by the extension. Canonical examples include
trying to generate a view provided by an extension that hasn't loaded
yet, or to import data using an importer that hasn't loaded yet. What
if exhibit were designed such that, if it got into that situation, it
simply registered a callback to be invoked when the necessary
functionality was available? e.g., view rendering could be gated by a
"require" method that wouldn't run its argument until a corresponding
"provide" method was called. I think this is the kind of functionality
offered by labjs, right? So as long as extension authors followed the
right design pattern, invoking labjs at the right moment, we'd be ok.
No, not at all. LABjs documentation is brief and readily available, I
suggest you read it before making suggestions on how it might be applied.
Making a UI seem responsive even if it isn't all there is a great way to
design things, but what was under discussion was load ordering - not
load transport delays. Comparing loading Exhibit, loading extensions,
loading data, and parsing data, network costs of loading extensions is
highly unlikely to be the reason the UI is blocked. It may interfere
due to indeterminate ordering; it is far less likely that it interferes
due to transport matters.
Still, if you want to redesign all of Exhibit along those lines, I'm
happy to look over a pull request. Making Exhibit initialization as
asynchronous as possible is a worthwhile goal.
Of course, you'd want to handle the case of the necessary element
*never* being provided---but I think the page's load event, which tells
you all scripts are loaded, could be used to trigger a "timeout" that
would signal an error if some necessary functionality were still missing
at that point.
This isn't related to the general topic, but as it pertains to Exhibit,
"the page's load event" does not tell you all the scripts are loaded.
Exhibit's capacity to load additional scripts means the page load event
(which specific one you're referring to is unclear) is definitely not a
reliable indicator of completion. Exhibit itself (now reliably) fires
its own scripts loaded event.
I agree with you that it's premature to rely on "async" loading, and
that inline-extensions would be come a parameter management nightmare.
An alternative to abusing link tags is abusing script tags. One could
add an attribute <script ex:role="extension" src=...>; exhibit could
find such tags and use them to detect scripts it should wait for (using
onload events) before running. On the other hand, for fastest response
it would be better to use the just in time approach discussed before,
since that would allow exhibit to quickly render its initial view even
if other views in the viewpanel require specific extensions.
Using script tags just comes back to the original problem that brought
us here, of browsers handling how they load scripts differently. If you
want Exhibit to get to the tags before the browsers themselves do, you
go back to the bad days of working with each browser's DOM ready event
and DOM manipulation inconsistencies to find them.
On 05/11/2012 03:08 PM, Ryan Lee wrote:
Hi all,
We've traced this down to a problem with the current extension loading
mechanism. In an attempt to keep things from shifting too much between
Exhibit 2 and Exhibit 3, I tried to allow for the form:
<script src="exhibit-api.js">
<script src="extensions/map-extension.js">
But, with the varied ways released browsers now handle the script tag,
this introduces a potential network delay in loading everything in the
right order; sometimes the map code gets loaded after Exhibit's started
to initialize itself, which is too late.
You probably weren't expecting design proposals to your question, but it
seems an opportune time to share what I've been kicking around the past
couple of days, each with its own strengths and weaknesses.
* Use LABjs
In this scenario, the only script src tag will be to load the LABjs
library, then use its loading framework inline to bring everything
related to Exhibit, including extensions, in in an orderly fashion. The
major drawback here is that it forces users who don't necessarily care
at all about code to pay attention to the actual scripting. For those
who do pay attention, it makes things a bit easier to incorporate
extra-Exhibit material into the loading process.
A related scenario would be to mask LABjs entirely with an
Exhibit-specific mechanism, but this only really adds a layer that might
allow an easier departure from LABjs if needed in the future.
* Use async / defer
HTML 5 introduces attributes to the script tag that allow the page
author to give hints to the browser about whether some scripts can load
without dependency on one another and whether some need to be run in the
order they appear in. But support is not universal across browsers.
* Abuse link rel
Instead of listing extensions (or other scripts) as actual script tags,
just point to them with
<link rel="exhibit-extension" href="...">
and let Exhibit pick out and load extensions itself during its loading
process. Other than the fact that I've never seen anybody use<link> in
this way, this might be the simplest and least obtrusive solution I
could think of. It does tend to make it even harder for those who don't
control the<head> of a document to get Exhibit in.
* Cram in more parameters
Instead of listing each extension separately, they could also be set as
additional parameters to the core of Exhibit
<script src="exhibit-api.js?extension=extensions/map-extension.js">
which (aside from taking some steps backwards into Exhibit's history) is
not as dubious as the prior proposal but makes that one line difficult
to read and assemble and additionally makes it difficult for the
extensions themselves to extract any of their own parameters; it would
probably require an ad hoc solution for adding extension-only parameters.
If you have any other ideas or feedback on the above, I'd love to hear
from you (or get a pull request from you on GitHub ;).
As for the Tile view issue, it had to do with paging and localization;
that's been fixed in trunk [1]. Perhaps we can make a new release
candidate in the near future bundling some of these fixes together.
1.
https://github.com/zepheira/exhibit3/commit/0814c0b3695bb96e79272c2b00de8b3c17d1f784
