Caching in Hakyll: Long live Data.Binary!



Published on January 25, 2010 under the tag Some thoughts and findings on implementing caching in HakyllPublished on January 25, 2010 under the tag haskell

What is this about

Some experiences from trying to make Hakyll run faster. I explain some of the things I have tried, and some of the things that have failed, in the hope this could one day be helpful to other projects.

How much does speed matter in Hakyll

Before you implement any caching measures in any program, you try to estimate what the speed gain will be. This is a heap profile from me generating this site from scratch (so after a ./hakyll clean ).

A graph showing that pandoc takes most time

I am far from an expert at Haskell profiling, and this graph is a heap profile and not really a benchmark, but I think you can nonetheless see that the functions from Pandoc (seen here as the top yellow and pink blocks) are the “more heavy” ones. This makes sense, because rendering Pandoc markdown to html is not easy. The total time is under 2 seconds, but there are only a few blogposts on this site now. I’ve taken it to the test, copying a blogpost a hundred times, and the result was that the time taken rose linearly with the number of posts (hey, who would’ve expected that).

The thing is that we can’t simply “make pandoc faster”. Pandoc is a marvelous piece of software, and probably quite optimized. What we can do is try to call Pandoc less.

One “caching” technique used in Hakyll is simple timestamp checking. For example, the projects page on my website is created from two files: a projects.markdown file containing the content, and a default template templates/default.html containing the header and footer of this site. Now, Hakyll generates all of it’s files to the _site directory. So it will perform a trivial timestamp test, and if _site/projects.html exists, and is more recent than both projects.markdown and templates/default.html , we don’t have to generate it again.

good : If we edit projects.markdown , only that file needs to be built again.

: If we edit , only that file needs to be built again. bad: If we edit templates/default.html , all pages using this template need to be rebuilt. And as you can see, every page uses this template!

So the solution here would be to cache Pandoc results in the _cache directory. Our code to load a page would then be something like:

valid <- check that the page is in _cache and more recent than the original if valid then return the result from the cache else read and parse the page store it in the cache return it

This means that when we edit templates/default.html , we don’t have to render every markdown file anymore, we can just fetch it from the cache. Now, how will we implement this storing and fetching from the cache?

Rolling our own serializer (disclaimer: worst idea)

A simple markdown page has the following layout:

--- title: Foobar --- sidebar This is a _sidebar_. --- # Title - Item one. - Item two.

My first idea was to store this in the cache as

--- title: Foobar --- sidebar This is a <i>sidebar</i>. --- <h1>Title</h1> <ul> <li>Item one.</li> <li>Item two.</li> </ul>

So, we basically write it out in the same format, but with the rendered html, so we don’t need to call Pandoc anymore. This seemed like an okay idea to me, but I encountered some problems very quickly:

It is not that easy to write it out in the same format.

We need to watch out not to add --- lines in the html, or it will be read wrong when we fetch it from the cache.

lines in the html, or it will be read wrong when we fetch it from the cache. It is still pretty slow to parse, because we want to check for the different sections, like the sidebar in this example.

I eagerly ran the test with a hundred blogposts again, waited and… well, to be fair, I didn’t even finish implementing this (I implemented a part of it, but it was still buggy). So, in short, this was a pretty bad idea: not easy to implement, and it was slow.

deriving (Show, Read)

Haskell provides automatic serialization using the Show and Read typeclasses. This seemed like an appealing option to me, also because of the fact it is very trivial to implement. I just made the Page type an instance:

I eagerly ran the test with a hundred blogposts again, waited and… nothing. Well, nearly nothing. I knew the automatic Read class is not that fast for large datatypes, but I hadn’t expected it to be so slow in this particular case. From 60 seconds without caching, it now took 58 seconds. Great, just great. But my options weren’t exhausted yet. I had recently read about the supposedly great Data.Binary library, so I thought I’d give it a try.

Data.Binary

It turned out adding an instance for Binary Page was nearly as easy as adding the Show and Read instances. Since Data.Binary provides instances for lists, tuples and strings, we can also serialize maps using the toAscList and fromAscList functions. Put short, I was able to make the Page type serializable using three lines of code:

I eagerly ran the test with a hundred blogposts again, waited and… profit! From previously taking 60 seconds, it now took only 15 seconds!

Conclusions