DEV Community

Duncan Mackenzie
Duncan Mackenzie

Posted on • Originally published at duncanmackenzie.net on

Do as little work as you can

One of the keys to scalability in software systems is to reduce the amount of work you do per action. In a website, think of this as the amount of processing and data transfer done per request. The smaller that set of work is, the better. This applies to all sorts of systems and at all layers of the stack, but I am going to focus on the website scenario because it is where I spend most of my time.

For every bit of work you do as part of your web application, you should be thinking about how to do it less. You have two main approaches to do this, but you can end up using them both depending on your situation. Caching, where we hold onto the output of some work so that we can reuse it and avoid doing the work again, and pre-processing, where we do work in advance so that we don’t have to do it ‘in real time’.

Our sample site

As I go through each of these concepts, we’ll focus on a fictional site that gives information about vehicles. The site works by pulling the make, model, and year out of the URL and rendering the appropriate information to the user. The site has this set of pages:

  • A home page (that is a list of vehicle manufacturers, the make),
  • A page per manufacturer (with a list of models),
  • A page per model (with a list of years that model was produced), and
  • A per specific vehicle, such as a 2011 Toyota Rav4

Thinking of this as URLs, which I love to do, we have:

  • / (the home page)
  • /Toyota (and /Mazda, /Audi, /Nissan, etc.)
  • /Toyota/Rav4 (the model page)
  • /Toyota/Rav4/2011 (the specific vehicle page)

Let’s assume that behind the scenes, we have a database that has all the vehicle information in it, so the basic logic of the site would be a route for each of the four types of pages, the URL would provide the parameters and then the code would make an appropriate request to the database.

Something like the pseudo-code below

Home page (/)

  • SELECT DisplayName, UrlSafeName FROM Make ORDER BY DisplayName
  • Loop through and list out the manufacturers, with links to //

Manufacturer page (/Toyota/)

  • Pull name from URL
  • Check DB to see if it is a valid make, if not then 404
  • Pull list of models from the database: SELECT model.DisplayName, model.UrlSafeName FROM Model INNER JOIN Make ON Model.makeID=Make.ID WHERE Make.UrlSafeName= ORDER BY model.DisplayName
  • Display list of models, with each one being a link to the model page

And so on for the next two levels.

Caching in all its wonderful forms

In its simplest form, caching is simply hanging onto something so that you can re-use it, but this can happen at many different layers in a web application. We can do client caching, where we add headers to our response telling the client (the user’s browser for example) that the page is valid for the next hour. In that case, if a user visited the homepage of the site, clicked on Toyota, then Rav4, then 2011 the server would end up needing to make as many as ten database queries to return all these results. If the user then came back to the homepage twenty minutes later and clicked on Toyota again, their browser wouldn’t need to make any new requests as it (likely) still has the page cached. On the server, we’ve avoided making those ten extra database queries. Now if they came back again a few days later, the browser knows their local copy is too old and would make new requests resulting in new database requests on the server.

This is great, but it only addresses the idea of a single user visiting multiple pages. If we instead have thousands of users visiting a given page, every one of those requests is generating database requests. What you need to ask yourself as a developer is “are those requests necessary?”. In other words, do we think the results of those queries is changing so fast, that we need to retrieve the data on every request? Assuming, as would be the case in our example, this data is not changing that quickly, we are doing too much work. We have two ways to solve this with caching. We could cache the results from the database query on the server, changing our logic for the home page (as an example) to be more like:

  • Is my saved copy of the list of manufacturers still fresh enough (retrieved within the last hour, for example)?
  • If so, just use it, otherwise make a query to the database and then save it in memory.
  • Display the list on the page.

Caching has let us take our database transactions for the homepage from one call per visitor to one per hour. Assuming thousands of page views per hour, we’ve reduced the load on our database by a massive amount. We’ve also made the load ‘fixed’, in that it doesn’t increase as our traffic does.

There is still work on every request though, we still must check the cache and do our loop to render the list onto the page. These are both simple bits of work, but it is still work, and so our goal should be to get rid of it. Some webservers let us do ‘output caching’, where the server will hang onto the rendered HTML output and if someone request that page again (within a set time window) the server will just return the HTML without having to run our code.

That’s still work though, we’ve just moved it from our code to the web server. Just asking for the page and getting a response puts load on the server and we don’t want that.

This is where a CDN (Content Delivery Network) can come in. We point our domain name to a CDN provider like Akamai and tell it to use our server as the source (called the origin in the CDN world). Using the same type of cache headers we used for client caching, we tell the CDN that our homepage is valid for an hour. The CDN will make a request back to our server whenever its copy expires, but that should only be once an hour, so now we are down to one hit on our server per hour, per unique URL (because each page has different content, so the CDN caches each one independently). You might point out here that each request is still causing work, but it is to the CDN’s servers and those aren’t ours, so we don’t really care.

That’s awesome and about as far as caching is going to get us. We could play with the duration of the cache (should it be one hour or twelve?) but the basic reduction is there. To get to an even better state, we need to look at pre-processing.

Pre-processing

In our caching example, whenever the CDN needed a new copy of any given URL, it would request it from the source, which in turn would result in one or more database queries. It’s a minor amount of work, but it is still work. That once-per-hour request is putting load on our database and will be a tiny bit slower getting back to the user.

Our goal is to get rid of as much work as we can though, so that is where pre-processing would come in. What if, instead of ever hitting the database, we generated the HTML for a page ahead of time, whenever the underlying data changed? If we updated the information for a 2011 Toyota Rav4, we would generate the HTML that /Toyota/Rav4/2011 should return and push that updated content up to the webserver.

Now, when the CDN needs that page, it asks the web server for that page and the server just hands the HTML back with no database call. It’s not zero work, but we are getting close. These days, we use the term ‘static’ to describe a site like this, where you build the content only when needed and never in response to a user request. Combined with a CDN, this approach reduces the overall amount of work happening on every request down to as close to zero as you can get. Instead of running database queries whenever anyone asks for a page, we run it only once to generate the output ahead of time.

I used the example of a vehicle site because back in the early 2000s, the MSN Autos site used this exact technique. They pre-rendered all their vehicle info pages whenever they updated their data. It was a great idea at the time, and it still is!

There are other benefits beyond just performance and scalability with this approach, we’ve also simplified the system that is running in production. If our database is down, or slow, it won’t impact the live site. If the code that generates updated HTML fails, that will prevent updates (which could be important depending on your scenario), but it won’t cause any errors or outages on the site.

Putting it all together

All the various levels of optimization discussed above are not different paths you can take; you can do them all in any combination. Caching at the CDN is more impactful than caching on an individual user’s machine, but the local cache can still improve the performance for that user, so do both. If your page generation code, even though we only run it occasionally, is making the same database queries repeatedly, you should consider caching the results as you go to make the process faster.

I’ve written about this before in other contexts, but if you want a fast site, make it as static as you can, use a CDN, and avoid doing any more work than necessary.

Top comments (0)