Journal tags: optimisation

A few years back, Craig wrote a great piece called Fast Software, the Best Software:

Speed in software is probably the most valuable, least valued asset. To me, speedy software is the difference between an application smoothly integrating into your life, and one called upon with great reluctance.

Nelson Elhage said much the same thing in his reflections on software performance:

I’ve really come to appreciate that performance isn’t just some property of a tool independent from its functionality or its feature set. Performance — in particular, being notably fast — is a feature in and of its own right, which fundamentally alters how a tool is used and perceived.

Or, as Robin put it:

I don’t think a website can be good until it’s fast.

Those sentiments underpin The Session. Speed is as much a priority as usability, accessibility, privacy, and security.

I’m fortunate in that the site doesn’t have an underlying business model at odds with these priorities. I’m under no pressure to add third-party code that would track users and slow down the website.

When it comes to making fast websites, most of the obstacles are put in place by front-end development, mostly JavaScript. I’ve been pretty ruthless in my pursuit of speed on The Session, removing as much JavaScript as possible. On the bigger pages, the bottleneck now is DOM size rather than parsing and excuting JavaScript. As bottlenecks go, it’s not the worst.

But even with all my core web vitals looking good, I still have an issue that can’t be solved with front-end optimisations. Time to first byte (or TTFB if you’d rather use an initialism that takes just as long to say as the words it’s replacing).

When it comes to reducing the time to first byte, there are plenty of factors that are out of my control. But in the case of The Session, something I do have control over is the server set-up, specifically the database.

Now I could probably solve a lot of my speed issues by throwing money at the problem. If I got a bigger better server with more RAM and CPUs, I’m pretty sure it would improve the time to first byte. But my wallet wouldn’t thank me.

(It’s still worth acknowledging that this is a perfectly valid approach when it comes to back-end optimisation that isn’t available on the front end; you can’t buy all your users new devices.)

So I’ve been spending some time really getting to grips with the MySQL database that underpins The Session. It was already normalised and indexed to the hilt. But perhaps there were server settings that could be tweaked.

This is where I have to give a shout-out to Releem, a service that is exactly what I needed. It monitors your database and then over time suggests configuration tweaks, explaining each one along the way. It’s a seriously good service that feels as empowering as it is useful.

I wish I could afford to use Releem on an ongoing basis, but luckily there’s a free trial period that I could avail of.

Thanks to Releem, I was also able to see which specific queries were taking the longest. There was one in particular that had always bothered me…

If you’re a member of The Session, then you can see any activity related to something you submitted in the past. Say, for example, that you added a tune or an event to the site a while back. If someone else comments on that, or bookmarks it, then that shows up in your “notifications” feed.

That’s all well and good but under the hood it was relying on a fairly convuluted database query to a very large table (a table that’s effectively a log of all user actions). I tried all sorts of query optimisations but there always seemed to be some combination of circumstances where the request would take ages.

For a while I even removed the notifications functionality from the site, hoping it wouldn’t be missed. But a couple of people wrote to ask where it had gone so I figured I ought to reinstate it.

After exhausting all the technical improvements, I took a step back and thought about the purpose of this particular feature. That’s when I realised that I had been thinking about the database query too literally.

The results are ordered in reverse chronological order, which makes sense. They’re also chunked into groups of ten, which also makes sense. But I had allowed for the possibility that you could navigate through your notifications back to the very start of your time on the site.

But that’s not really how we think of notifications in other settings. What would happen if I were to limit your notifications only to activity in, say, the last month?

Boom! Instant performance improvement by orders of magnitude.

I guess there’s a lesson there about switching off the over-analytical side of my brain and focusing on actual user needs.

Anyway, thanks to the time I’ve spent honing the database settings and optimising the longest queries, I’ve reduced the latency by quite a bit. I’m hoping that will result in an improvement to the time to first byte.

Time—and monitoring tools—will tell.

There’s a feature in service workers called navigation preloads. It’s relatively recent, so it isn’t supported in every browser, but it’s still well worth using.

Here’s the problem it solves…

If someone makes a return visit to your site, and the service worker you installed on their machine isn’t active yet, the service worker boots up, and then executes its instructions. If those instructions say “fetch the page from the network”, then you’re basically telling the browser to do what it would’ve done anyway if there were no service worker installed. The only difference is that there’s been a slight delay because the service worker had to boot up first.

1. The service worker activates.
2. The service worker fetches the file.
3. The service worker does something with the response.

It’s not a massive performance hit, but it’s still a bit annoying. It would be better if the service worker could boot up and still be requesting the page at the same time, like it would do if no service worker were present. That’s where navigation preloads come in.

1. The service worker activates while simultaneously requesting the file.
2. The service worker does something with the response.

Navigation preloads—like the name suggests—are only initiated when someone navigates to a URL on your site, either by following a link, or a bookmark, or by typing a URL directly into a browser. Navigation preloads don’t apply to requests made by a web page for things like images, style sheets, and scripts. By the time a request is made for one of those, the service worker is already up and running.

To enable navigation preloads, call the enable() method on registration.navigationPreload during the activate event in your service worker script. But first do a little feature detection to make sure registration.navigationPreload exists in this browser:

if (registration.navigationPreload) {
  addEventListener('activate', activateEvent => {
    activateEvent.waitUntil(
      registration.navigationPreload.enable()
    );
  });
}

If you’ve already got event listeners on the activate event, that’s absolutely fine: addEventListener isn’t exclusive—you can use it to assign multiple tasks to the same event.

Now you need to make use of navigation preloads when you’re responding to fetch events. So if your strategy is to look in the cache first, there’s probably no point enabling navigation preloads. But if your default strategy is to fetch a page from the network, this will help.

Let’s say your current strategy for handling page requests looks like this:

addEventListener('fetch', fetchEvent => {
  const request = fetchEvent.request;
  if (request.headers.get('Accept').includes('text/html')) {
    fetchEvent.respondWith(
      fetch(request)
      .then( responseFromFetch => {
        // maybe cache this response for later here.
        return responseFromFetch;
      })
      .catch( fetchError => {
        return caches.match(request)
        .then( responseFromCache => {
          return responseFromCache || caches.match('/offline');
        });
      })
    );
  }
});

That’s a fairly standard strategy: try the network first; if that doesn’t work, try the cache; as a last resort, show an offline page.

It’s that first step (“try the network first”) that can benefit from navigation preloads. If a preload request is already in flight, you’ll want to use that instead of firing off a new fetch request. Otherwise you’re making two requests for the same file.

To find out if a preload request is underway, you can check for the existence of the preloadResponse promise, which will be made available as a property of the fetch event you’re handling:

fetchEvent.preloadResponse

If that exists, you’ll want to use it instead of fetch(request).

if (fetchEvent.preloadResponse) {
  // do something with fetchEvent.preloadResponse
} else {
  // do something with fetch(request)
}

You could structure your code like this:

addEventListener('fetch', fetchEvent => {
  const request = fetchEvent.request;
  if (request.headers.get('Accept').includes('text/html')) {
    if (fetchEvent.preloadResponse) {
      fetchEvent.respondWith(
        fetchEvent.preloadResponse
        .then( responseFromPreload => {
          // maybe cache this response for later here.
          return responseFromPreload;
        })
        .catch( preloadError => {
          return caches.match(request)
          .then( responseFromCache => {
            return responseFromCache || caches.match('/offline');
          });
        })
      );
    } else {
      fetchEvent.respondWith(
        fetch(request)
        .then( responseFromFetch => {
          // maybe cache this response for later here.
          return responseFromFetch;
        })
        .catch( fetchError => {
          return caches.match(request)
          .then( responseFromCache => {
            return responseFromCache || caches.match('/offline');
          });
        })
      );
    }
  }
});

But that’s not very DRY. Your logic is identical, regardless of whether the response is coming from fetch(request) or from fetchEvent.preloadResponse. It would be better if you could minimise the amount of duplication.

One way of doing that is to abstract away the promise you’re going to use into a variable. Let’s call it retrieve. If a preload is underway, we’ll assign it to that variable:

let retrieve;
if (fetchEvent.preloadResponse) {
  retrieve = fetchEvent.preloadResponse;
}

If there is no preload happening (or this browser doesn’t support it), assign a regular fetch request to the retrieve variable:

let retrieve;
if (fetchEvent.preloadResponse) {
  retrieve = fetchEvent.preloadResponse;
} else {
  retrieve = fetch(request);
}

If you like, you can squash that into a ternary operator:

const retrieve = fetchEvent.preloadResponse ? fetchEvent.preloadResponse : fetch(request);

Use whichever syntax you find more readable.

Now you can apply the same logic, regardless of whether retrieve is a preload navigation or a fetch request:

addEventListener('fetch', fetchEvent => {
  const request = fetchEvent.request;
  if (request.headers.get('Accept').includes('text/html')) {
    const retrieve = fetchEvent.preloadResponse ? fetchEvent.preloadResponse : fetch(request);
    fetchEvent.respondWith(
      retrieve
      .then( responseFromRetrieve => {
        // maybe cache this response for later here.
       return responseFromRetrieve;
      })
      .catch( fetchError => {
        return caches.match(request)
        .then( responseFromCache => {
          return responseFromCache || caches.match('/offline');
        });
      })
    );
  }
});

I think that’s the least invasive way to update your existing service worker script to take advantage of navigation preloads.

Like I said, preload navigations can give a bit of a performance boost if you’re using a network-first strategy. That’s what I’m doing here on adactio.com and on thesession.org so I’ve updated their service workers to take advantage of navigation preloads. But on Resilient Web Design, which uses a cache-first strategy, there wouldn’t be much point enabling navigation preloads.

Jeff Posnick made this point in his write-up of bringing service workers to Google search:

Adding a service worker to your web app means inserting an additional piece of JavaScript that needs to be loaded and executed before your web app gets responses to its requests. If those responses end up coming from a local cache rather than from the network, then the overhead of running the service worker is usually negligible in comparison to the performance win from going cache-first. But if you know that your service worker always has to consult the network when handling navigation requests, using navigation preload is a crucial performance win.

Oh, and those browsers that don’t yet support navigation preloads? No problem. It’s a progressive enhancement. Everything still works just like it did before. And having a service worker on your site in the first place is itself a progressive enhancement. So enabling navigation preloads is like a progressive enhancement within a progressive enhancement. It’s progressive enhancements all the way down!

By the way, if all of this service worker stuff sounds like gibberish, but you wish you understood it, I think my book, Going Offline, will prove quite valuable.

When I was helping Bevan with making the dConstruct site, I kept banging on to him about the importance of performance.

Don’t get me wrong: I wanted the site to look great, but I also very much wanted it to feel great …and nothing affects the feel of a site (the user’s experience, if you will) more than performance. As Jason wrote:

If you could only do one thing to prepare your desktop site for mobile and had to choose between employing media queries to make it look good on a mobile device or optimizing the site for performance, you would be better served by making the desktop site blazingly fast.

And yet this fundamental aspect of how performant a site is going to be is all too often left until the development phase. I’d really like to see it taken into account much earlier on, during the UX and visual design phases.

Anyway, as the dConstruct site came together, I just kept asking “What would Steve Souders do?”

For a start, that meant ripping out any boilerplate markup and CSS that was there “just in case.” I very much agree with Rachel when she says stop solving problems you don’t yet have. But one of the areas where the unfortunately-named HTML5 Boilerplate excels is in its suggestions for .htaccess rules so I made sure to rip off the best bits.

Initially jQuery was being included, but given how far browsers have come in their JavaScript support, I was able to ditch it and streamline the JavaScript a bit.

Wherever possible, I made sure that background images in CSS were Base64 encoded as data URIs; icons, textures, and the like. That helped to reduce the number of HTTP requests—one of the easy wins for improving performance.

I’ve already mentioned the conditional loading that’s going on.

Then there’s the thorny issue of responsive images. The dConstruct 2012 site is similar to the dConstruct archive in that there is no correlation between browser width and image: quite often, a smaller image is required for wider screens than for narrower viewports because of the presence of a grid. So instead of trying to come up with some complex interplay of client and server cross-communication to figure out which size image would be appropriate to serve up, I instead took the same approach as I did for the archive site: optimise the hell out of images, regardless of whether they’re going to be viewed in a desktop or a mobile device.

Take a look at the original image of Kevin Slavin compared to the version that appears on the dConstruct archive.

See how everything except the face is so much blurrier in the final version? That isn’t just an attempt to introduce some cool bokeh. It makes for much smaller JPGs with fewer jaggy artefacts. And because human beings tend to focus on other human faces, the technique isn’t really consciously noticeable (although you’ll notice it now that I’ve pointed it out to you).

The design of the 2012 dConstruct site called for monochrome images with colour filters applied.

That turned out to be a fortunate boon for optimising the images. This time we were using PNGs rather than JPGs and we were able to get the number of colours down to 32 or even 16. Run them through Image Optim or Smushit and you can squeeze even more bytes out of them.

The funny thing is that sweating the file sizes of images used to be part and parcel of web development. Back in the nineties, there was something of an aesthetic that grew out of the need to optimise images with limited (web-safe!) colour palettes. That was because bandwidth was at a premium and you could be pretty sure that plenty of people were accessing your site on slow connections.

Well, here we are fifteen years later and thanks to the rise of mobile, bandwidth is once again at a premium and we can be pretty sure that plenty of people are accessing our sites on slow connections. Yet again, mobile is highlighting issues that were always there. When did we get so lazy and decide it was acceptable to send giant unoptimised images down the pipe to our long-suffering visitors?

Mathew Pennell recently wrote:

…it’s certainly true that the golden rule I grew up with – no page should ever be over 100Kb – has long since been mothballed.

But why? That seems like a perfectly good and still-relevant rule to me.

Alas, on the dConstruct site I wasn’t able to hit that target. With an unprimed cache, the home page comes in at around 300K (it’s 17K with a primed cache). By far the largest file is the CSS, weighing at 113K, followed by the web font, Futura bold oblique, at 32K.

By the way, when it comes to analysing performance in the browser, this missing manual for the Webkit inspector is really, really handy. I also ran the site through Google Page Speed but it seems that the user-agent chooses an arbitrary browser width (960? 1024?) so some of the advice about scaling images needs to be taken with a pinch of salt when applied to responsive designs.

I took a look at some other conference sites too. The beautiful site for the Build conference comes in at just under a megabyte for the homepage—it has quite a few fonts and images. It also has a monochrome aesthetic going on so I suspect quite a few of those images could be squeezed down (and some far-future expiry dates would help for repeat visitors).

Then there’s site for this year’s Mobilism conference which is blazingly fast. The combined file size on the homepage isn’t that different to the dConstruct site (although the CSS is significantly smaller) and I suspect there’s some server-side wizardry going on. I’ll have to corner Stephen at the conference next week and quiz him about it.

For now, server-side performance optimisation is something beyond my ken. I should really do something about that, especially as I’m expecting the dConstruct site to take a hammering the day that tickets go sale (May 29th—save the date).

In the meantime, there’s still plenty I can do on the front end. As Bruce put it:

It seems to me that old-fashioned, oh-so-dull techniques might not be ready for retirement yet. You know: well-crafted HTML, keeping JavaScript for progressive enhancement rather than a pre-requisite for the page even displaying, and testing across browsers.

All those optimisation techniques we learned in the 90s—and even wacky ideas like lowsrc—are back in fashion. Everything old is new again.