This weekend I spent some time trying to fix a conflict I was having between Travis CI and Heroku over Gemfile.lock. Heroku needed it to be checked in and Travis CI kept failing unless I let it build the file on it's own. The solution for that problem was to tell Travis to install Bundler. (This is documented, but I guess that bit is new or I ignored it for some reason. I certainly saw the next section about caching Bundler's output.)
So once I had that sorted out and the build wasn't failing on Heroku anymore, I opened a tab to the page for my Heroku app and saw:
"Ok," I thought, "It's been awhile since this was working. Maybe Heroku needs to allocate some resources or something." After refreshing the page, it looked right so I clicked on a link to one of my posts. That got me to a page that looked suspiciously familiar:
This made less sense. Jekyll builds the entire site as static pages
and caches them in the
_site directory. Once one page is rendered,
all the others should be rendered too. But maybe there is a layer of
caching somewhere that, I don't know, needs to regenerate the site for
each page? After reloading, the post page looked fine. Then I went
back to the front page and saw . . . well, you can probably guess.
Then I remembered that this was the issue that made me give up on Heroku in the past. It must not be a very reliable service if it couldn't serve up a static page. Or, I speculated, it works fine if you have enough traffic to keep the resources online, but if you don't have enough traffic, Heroku deprovisions the Linux containers that power the site. Then when the rare user (me) hits a page, the page isn't cached anymore but needs to be rebuilt. So I just, you know, gave up.
This time, I dug a little deeper and found a reference to Production
in turn introduced me to the JEKYLL_ENV environment
turns out GitHub Pages sets that to
production automatically, which is
why the site works out of the box with their platform. To fix the
problem on Heroku, all I needed was to run a simple command:
$ heroku config:set JEKYLL_ENV=production -a jlericson
Obviously you'll want to customize the
-a jlericson bit to point to
whatever your application might be. Another alternative would be to
add the environment variable to the app's
web: JEKYLL_ENV=production bundle exec puma -t 8:32 -w 3 -p $PORT
When I first set up Heroku it was because I wanted to see if it would work for hosting the Unicorn Meta Zoo podcast. I wanted to have an alternative to just putting MP3s on GitHub Pages. Their usage limits are generous, but I didn't want to abuse them. And it's always nice to have a fallback if GitHub no longer works for whatever reason. Even if I never use Heroku to host the site, making sure the blog works there means I'm not locked into my existing host.
Of course I'm not really worried about my personal blog. Worse case
scenario, I can run Jekyll locally and copy the static pages to some
other server. That's how I build my tag
and it could work for the entire site too, if needed. Being able to
git push is great and all, but it's really not hard to
script FTP transfers or whatnot.
- Our development team makes a change and tests it locally.
- When we are ready, we submit a request to move the changes to our QA site, which is hosted by the hosting company.
- We test the changes on the QA site and, when we are satisfied that works, we submit a request to push them to production.
- The hosting company vets the changes and they go live.
This is the best case scenario. The two review steps (#2 and #4) contractually take 3-5 days for a response.2 If something goes wrong, a change will need to go back to step #1 for more work and the whole process needs to start from scratch. In theory this could work as a long pipeline where projects might take a little longer to deploy. In practice any little misstep yanks developers back to half-forgotten code. It's not working.
So we're solving this by developing on our own site and asking the hosting company to redirect to our own pages. For instance, we built our a forum index page that looks like it's on the forum. It's more or less a polished version of the index I created for myself. As long as users don't look too carefully at the URL, they will get a seamless experience. Or at least that's the idea.
But the long term solution is to host the forums on our own. We're going to start by hosting our own QA environment to cut out that half of our development process. Using what we learn doing that, we'll be able to count the cost for canceling the hosting contract altogether. Even if we decide later not to self-host, the work we are doing now gives us flexibility to build what we need.
Just for fun, I put the various options in
UptimeRobot monitors to
see which is most reliable. At the time of publication, Heroku looks
bad because of me testing the
JEKYLL_ENV setting. I imagine it'll be
more reliable when I stop fiddling with it.
I'm not going to call the company out by name (though it's not hard to figure it out, if you know how) because we're still under contract with them. Probably they aren't petty enough to make things worse because of bad publicity, but you can't be too careful. ↩
These are business days, which take longer than normal days. Also . . . let's just say the contract is only as good as its enforcement. Sometimes the response time can be measured in weeks. ↩