Tiny, dirty, iffy, good enough, basic multi-threaded web crawler in Python | Hacker News Reader

Tiny, dirty, iffy, good enough, basic multi-threaded web crawler in Python

blog.webhose.io

14 pointsrangeva9 comments

The regex will break on

    <a href='actualLink' _href='spoofedLink'

And will return spoofedLink instead of actualLink, while browsers will follow actualLink. This is why you shouldn't be trying to parse xml/html with regexes.

Whilst it's true in principal that you should always parse xml/html with a proper parser, in practise it can be a little different. On the web you'll find just as much malformed html that will throw an error in any parser, as you will bad html that will throw a regex. As long as you have proper verification of the link at the end, you'll be safe enough.

Then when you factor in that constructing and extracting links from the DOM might take 500ms vs. the 5ms of a regex - a 100x increase in crawl speed in exchange for a little potential dirtiness is more than acceptable; particularly if you have a million pages to crawl. That's the difference between a crawl taking 1 week, vs. 2 years...

And anyway, all web crawls are inevitably going to end up dirty and scrappy, that's just the nature of the web. I once crawled a website that returned "_ontent-Type" instead of "Content-Type", making it impossible to tell what kind of file was about to be downloaded by my polite, "text only" crawler.

Crawling can be broken down to:

  1) fetching resources
  2) finding out what new resources to fetch

1) is an network bound problem, 2) is mostly disk/CPU bound. Realizing the difference between these two things and separating them is the key to building a good crawler.

Depending on how you find out what resources to fetch (parsing static documents vs. dynamic JS analysis with multiple dependencies on other resources (included JS &c)), "good-enough" crawlers are mostly bound to the network.

I've seen people running 1 crawl/process on their back-end and some management guy saying "we need to crawl faster, add more threads per crawl" when one crawl cycle spends 10x times more waiting on the network than it does parsing a document.

If you're crawling different origins in parallel then you could gain a boost in speed by threading. It's usually bound by the network/processing on the receiving side. Your server can probably handle several outgoing connections at once.

You can have several outgoing connections in one single thread, at the same time.

Threading is for CPU bound problems that requires a shared memory space.

I/O multiplexing and/or cooperative multitasking means you can have thousands concurrent connections per thread.