A Web crawler, sometimes called a spider or spiderbot and often shortened to crawler, is an Internet bot that systematically browses the World Wide Web, typically for the purpose of Web indexing (web spidering). -- Wikipedia
- Architecture
- Considerations
- Code Comments
- Code Design
- Examples
- Structure
- Improvements
- TODO
- How long did it take?
![](/Integralist/go-web-crawler/raw/master/architecture.png)
Note: the implementation tracks already processed pages and so you'll find (over time) that the amount of work carried out is reduced. It's not logarithmic, or exponential, levels of reduction in workload (as the content is non-deterministic) but there is a degree of reduction.
Due to the potential unbounded nature of a web crawler, we want to ensure we don't overwhelm the CPU by creating a large number of goroutines (they still have a cost, especially with regards to the overhead of the scheduler having to context switch).
To try and avoid that issue we'll limit our concurrency by:
- Defining a bounded worker pool of goroutines.
- Providing tasks for the pool via a buffered channel.
Note: the worker pool defaults to 20, but will be set dynamically to a smaller number if there's less data to be processed.
This project was born from a 'take home' job interview code test, and so I've heavily commented the code to explain certain rationale/thought processes. I wouldn't normally have this many comments in my code as I prefer to move complicated logic into separate functions †
† this is a classic refactoring technique referred to as Extract Method, please refer to an old blog post of mine (2013) for more refactoring techniques.
To fulfil the design requirements of this project, we additionally have six components:
Note: dear lord having generics in Go would have helped make some of the repetitive tasks easier to design 🤦♂️
The coordinator package acquires the entry page for the given host, and then kick starts the crawling/parsing/mapping stages for each subsequent web page found. It does this by recursively calling an internal process function (meaning pages are processed in concurrent batches).
The requester is a simple wrapper around the net/http Get
function. It accepts a URL to request, and returns a struct consisting of the URL and the response body. It is used by both the Coordinator (in order to retrieve the entry page) and the Crawler (for requesting multiple URLs related to anchors found in each crawled page).
The crawler accepts a list of 'mapped' pages (the coordinator provides a single mapped page in order to kickstart the subsequent crawl), and it will loop over all the found anchors (nested linked URLs <a href="...">
) for each page and request them using the Requester.
The crawler will return its own list of requested pages, wrapped in a struct, to be further processed by the Parser package.
Once the crawler has returned a subset of pages, those will be passed over to the parser to tokenize. The parser will then return its own list of tokenized pages, wrapped in a struct, to be further processed by the Mapper package.
The parser has two exported functions:
Parse
: accepts arequester.Page
and tokenizes it.ParseCollection
: accepts a slice ofrequester.Page
and sends each page toParse
.
Once the parser has returned a set of tokenized pages, those will be passed over to the mapper to filter out any unwanted content. The mapper will then return its own list of pages, wrapped in a struct (with filtered fields), which are appended to a final results
slice within the coordinator package, and which is used to display what was crawled.
The mapper has two exported functions:
Map
: accepts aparser.Page
and filters it.MapCollection
: accepts a slice ofparser.Page
and sends each page toMap
.
The formatter is used when passing either the -json
or -dot
flags. It currently offers three exported functions:
Dot
: transforms the results data into dot format notation for use with generating a site map graph via graphviz.Pretty
: pretty prints any given data structure (for easier debugging/visualization).Standard
: the default output format used (number of URLs crawled/processed and the total time it took).
The Dot
output will be (for integralist.co.uk
) something like the following (albeit much longer):
digraph sitemap {
"https://www.integralist.co.uk/posts"
-> {
"https://www.integralist.co.uk/",
"https://www.integralist.co.uk/about/",
"https://www.integralist.co.uk/resume/",
"https://www.integralist.co.uk/search/",
"https://www.integralist.co.uk/posts",
"http://www.integralist.co.uk/posts/algorithmic-complexity-in-python/",
"http://www.integralist.co.uk/posts/data-types-and-data-structures/",
"http://www.integralist.co.uk/posts/python-code-design/",
"http://www.integralist.co.uk/posts/modern-js/",
"http://www.integralist.co.uk/posts/engineer-to-manager/",
"http://www.integralist.co.uk/posts/architecture-interview/",
"http://www.integralist.co.uk/posts/post-mortem-template/",
"http://www.integralist.co.uk/posts/opsbot/",
"http://www.integralist.co.uk/posts/go-interfaces/",
"http://www.integralist.co.uk/posts/multigrain-services/",
"https://www.integralist.co.uk/posts/page/2/",
"https://www.integralist.co.uk/posts/page/7/"
}
"https://www.integralist.co.uk/search/"
-> {
"https://www.integralist.co.uk/",
"https://www.integralist.co.uk/about/",
"https://www.integralist.co.uk/resume/",
"https://www.integralist.co.uk/search/",
"https://www.integralist.co.uk/posts",
}
...
}
You can then either redirect this output to a standalone file (e.g. example.dot
) or pipe the output direct to the graphviz dot
command, in order to generate a graph of crawled web pages:
dot -Tpng example.dot -o example.png
![](/Integralist/go-web-crawler/raw/master/example.png)
Note: there is a known issue with this approach, which is that a large website with lots of interlinking pages will be impossibly slow to generate an image when using graphviz, simply because the permutations of cross-posting links (my site is one such example, and the above image is a tiny representation of cross linking).
To run the program, you can use the provided Makefile for simplicity:
make run
The above command is equivalent to:
go run -ldflags "-X main.version=71c00f0" cmd/crawler/main.go -hostname integralist.co.uk -subdomains "www,"
Note: we use the last repository commit for internal app versioning.
The final output from crawling integralist.co.uk
was (at the time of writing):
Number of URLs crawled and processed: 285
Time: 6.619973048s
If you want to see json output of the site structure (all crawled pages, and any links/static assets found) then we need to provide the go program with a -json
flag, which via Make is configured like so:
make run json=-json
The output will resemble something like the following (albeit much longer):
[
{
"Anchors": [
"https://www.integralist.co.uk/",
"https://www.integralist.co.uk/about/",
"https://www.integralist.co.uk/resume/",
"https://www.integralist.co.uk/search/",
"https://www.integralist.co.uk/posts",
"https://www.integralist.co.uk/posts/python-code-design/"
],
"Links": [
"https://www.integralist.co.uk/css/main.d02777fd.css",
"https://www.integralist.co.uk/css/custom.css",
"https://www.integralist.co.uk/css/highlight.github.css"
],
"Scripts": [
"https://www.integralist.co.uk/js/main.af838dd5.js",
"https://www.integralist.co.uk/js/custom.js",
"https://www.integralist.co.uk/js/highlight.pack.js",
"https://www.integralist.co.uk/js/highlight-trigger.js"
],
"URL": "https://www.integralist.co.uk/tags/asyncio"
}
]
The benefit of json output is that we can pipe it to other programs for further processing using other tools available to the calling environment.
For example, the below command will output (with integralist.co.uk
as the default -hostname
value) 290
(that's 290 crawled pages):
make run json=-json | jq .[].URL | sort | uniq -c | wc -l
To crawl a different website, let's say monzo.com
, we need to provide the go program with a -hostname
flag, which via Make is configured like so:
make run hostname=monzo.com
The final output from crawling monzo.com
was (at the time of writing):
Number of URLs crawled and processed: 1047
Time: 5m21.896589638s
The project follows the guidelines as defined by:
github.com/golang-standards/project-layout
.
├── Makefile
├── cmd
│ └── crawler
│ └── main.go
├── dist
├── go.mod
├── go.sum
└── internal
├── coordinator
│ └── coordinator.go
├── crawler
│ └── crawler.go
├── formatter
│ ├── formatter.go
│ └── formatter_test.go
├── instrumentator
│ └── instrumentator.go
├── mapper
│ ├── mapper.go
│ └── mapper_test.go
├── parser
│ ├── filters.go
│ └── parser.go
└── requester
├── http.go
└── http_test.go
Some things that we should consider...
-
Resilience: HTTP errors (e.g. timeouts) are quietly skipped over but we could utilize a package such as Hashicorp's go-retryablehttp which implements logic for handling retries and also exponential backoff.
-
Performance: profile hot spots, and tweak the 'bounded worker pool' pattern used.
-
Design:
- interfaces to make passing behaviour cleaner
- pass generic struct as dependency for each function instead of package level variables
- Tests (there are some, but coverage is weak).
- Might also require some code redesign to better support interfaces + dependency injection.
- Finish my refactoring of configuring instrumentation.
- i.e. removal of package level
Init
functions in favour of dependency injection with new instrumentation type. - also look at refactoring functions to avoid long signatures.
- i.e. removal of package level
- Think of different approach to rendering large/complex graph data (either json or dot format).
- Using graphviz didn't work out once the bidirection edges become large (as they do in my site).
- Ignore URLs based on pattern (e.g.
/tags/...
).- In the case of my blog, I don't necessarily care about crawling the tag endpoints.
- Fix duplicated URLs with trailing slashes.
- e.g.
https://www.integralist.co.uk/posts
vshttps://www.integralist.co.uk/posts/
- e.g.
The initial design plan I had was always to have a Parser and Mapper package, and to just use Wait Groups to control handling the concurrent requesting of pages within the main
package. That took a few hours to put together and was quite crude in implementation.
After that I implemented the bounded worker pool pattern, and could have just stopped there, but I was enjoying working on the project and so I spent a few more hours (over the coming days) tweaking the design and implementation, extra packages (coordinator & crawler) documentation etc.