gocrawl is a polite, slim and concurrent web crawler written in Go.
For a simpler yet more flexible web crawler written in a more idiomatic Go style, you may want to take a look at fetchbot, a package that builds on the experience of gocrawl.
- Full control over the URLs to visit, inspect and query (using a pre-initialized goquery document)
- Crawl delays applied per host
- Obedience to robots.txt rules (using the robotstxt.go library)
- Concurrent execution using goroutines
- Configurable logging
- Open, customizable design providing hooks into the execution logic
gocrawl depends on the following userland libraries:
It requires Go1.1+ because of its indirect dependency on golang.org/x/net/html
. To install:
go get github.com/PuerkitoBio/gocrawl
To install a previous version, you have to git clone https://github.com/PuerkitoBio/gocrawl
into your $GOPATH/src/github.com/PuerkitoBio/gocrawl/
directory, and then run (for example) git checkout v0.3.2
to checkout a specific version, and go install
to build and install the Go package.
- 2016-05-24 : Set the
*URLContext.SourceURL()
and*URLContext.NormalizedSourceURL()
to the original URL on redirections (see #55). Thanks to github user @tmatsuo. - 2016-02-24 : Always use
Options.UserAgent
to make requests, useOptions.RobotUserAgent
only for robots.txt policy matching. Lint and vet the code a bit, better godoc documentation. - 2014-11-06 : Change import paths of net/html to golang.org/x/net/html (see https://groups.google.com/forum/#!topic/golang-nuts/eD8dh3T9yyA).
- v0.4.1 : now go-getable, since goquery is go-getable too.
- v0.4.0 : BREAKING CHANGES major refactor, API changes:
* Use an
*URLContext
structure as first argument to allExtender
interface functions that are called in the context of an URL, instead of a simple*url.URL
pointer that was sometimes normalized, sometimes not. * Remove theEnqueueOrigin
enumeration flag. It wasn't even used by gocrawl, and it is a kind of state associated with the URL, so this feature is now generalized with the next bullet... * Add aState
for each URL, so that the crawling process can associate arbitrary data with a given URL (for example, the ID of a record in the database). Fixes issue #14. * More idiomatic use of errors (ErrXxx
variables,Run()
now returns an error, removing the need for theEndReason
enum). * Much simplifiedFilter()
function. It now only returns abool
indicating if it should be visited or not. The HEAD request override feature is provided by the*URLContext
structure, and can be set anywhere. The priority feature was unimplemented and has been removed from the return values, if it gets implemented it will probably be via the*URLContext
structure too. *Start
,Run
,Visit
,Visited
and theEnqueueChan
all work with the empty interface type for the URL data. While this is an inconvenience for compile-time checks, it allows for more flexibility regarding the state feature. Instead of always forcing amap[string]interface{}
type even when no state is needed, gocrawl supports various types. * Some other internal changes, better tests. - v0.3,2 : Fix the high CPU usage when waiting for a crawl delay.
- v0.3.1 : Export the
HttpClient
variable used by the defaultFetch()
implementation (see issue #9). - v0.3.0 : BEHAVIOR CHANGE filter done with normalized URL, fetch done with original, non-normalized URL (see issue #10).
- v0.2.0 : BREAKING CHANGES rework extension/hooks.
- v0.1.0 : Initial release.
From example_test.go
:
// Only enqueue the root and paths beginning with an "a"
var rxOk = regexp.MustCompile(`http://duckduckgo\.com(/a.*)?$`)
// Create the Extender implementation, based on the gocrawl-provided DefaultExtender,
// because we don't want/need to override all methods.
type ExampleExtender struct {
gocrawl.DefaultExtender // Will use the default implementation of all but Visit and Filter
}
// Override Visit for our need.
func (x *ExampleExtender) Visit(ctx *gocrawl.URLContext, res *http.Response, doc *goquery.Document) (interface{}, bool) {
// Use the goquery document or res.Body to manipulate the data
// ...
// Return nil and true - let gocrawl find the links
return nil, true
}
// Override Filter for our need.
func (x *ExampleExtender) Filter(ctx *gocrawl.URLContext, isVisited bool) bool {
return !isVisited && rxOk.MatchString(ctx.NormalizedURL().String())
}
func ExampleCrawl() {
// Set custom options
opts := gocrawl.NewOptions(new(ExampleExtender))
// should always set your robot name so that it looks for the most
// specific rules possible in robots.txt.
opts.RobotUserAgent = "Example"
// and reflect that in the user-agent string used to make requests,
// ideally with a link so site owners can contact you if there's an issue
opts.UserAgent = "Mozilla/5.0 (compatible; Example/1.0; +http://example.com)"
opts.CrawlDelay = 1 * time.Second
opts.LogFlags = gocrawl.LogAll
// Play nice with ddgo when running the test!
opts.MaxVisits = 2
// Create crawler and start at root of duckduckgo
c := gocrawl.NewCrawlerWithOptions(opts)
c.Run("https://duckduckgo.com/")
// Remove "x" before Output: to activate the example (will run on go test)
// xOutput: voluntarily fail to see log output
}
Gocrawl can be described as a minimalist web crawler (hence the "slim" tag, at ~1000 sloc), providing the basic engine upon which to build a full-fledged indexing machine with caching, persistence and staleness detection logic, or to use as is for quick and easy crawling. Gocrawl itself does not attempt to detect staleness of a page, nor does it implement a caching mechanism. If an URL is enqueued to be processed, it will make a request to fetch it (provided it is allowed by robots.txt - hence the "polite" tag). And there is no prioritization among the URLs to process, it assumes that all enqueued URLs must be visited at some point, and that the order in which they are is unimportant.
However, it does provide plenty of hooks and customizations. Instead of trying to do everything and impose a way to do it, it offers ways to manipulate and adapt it to anyone's needs.
As usual, the complete godoc reference can be found here.
The major use-case behind gocrawl is to crawl some web pages while respecting the constraints of robots.txt
policies and while applying a good web citizen crawl delay between each request to a given host. Hence the following design decisions:
- Each host spawns its own worker (goroutine) : This makes sense since it must first read its robots.txt data, and only then proceed sequentially, one request at a time, with the specified delay between each fetch. There are no constraints inter-host, so each separate worker can crawl independently.
- The visitor function is called on the worker goroutine : Again, this is ok because the crawl delay is likely bigger than the time required to parse the document, so this processing will usually not penalize the performance.
- Edge cases with no crawl delay are supported, but not optimized : In the rare but possible event when a crawl with no delay is needed (e.g.: on your own servers, or with permission outside busy hours, etc.), gocrawl accepts a null (zero) delay, but doesn't provide optimizations. That is, there is no "special path" in the code where visitor functions are de-coupled from the worker, or where multiple workers can be launched concurrently on the same host. (In fact, if this case is your only use-case, I would recommend not to use a library at all - since there is little value in it -, and simply use Go's standard libs and fetch at will with as many goroutines as are necessary.)
- Extender interface provides the means to write a drop-in, fully encapsulated behaviour : An implementation of
Extender
can radically enhance the core library, with caching, persistence, different fetching strategies, etc. This is why theExtender.Start()
method is somewhat redundant with theCrawler.Run()
method,Run
allows calling the crawler as a library, whileStart
makes it possible to encapsulate the logic required to select the seed URLs into theExtender
. The same goes forExtender.End()
and the return value ofRun
.
Although it could probably be used to crawl a massive amount of web pages (after all, this is fetch, visit, enqueue, repeat ad nauseam!), most realistic (and um... tested!) use-cases should be based on a well-known, well-defined limited bucket of seeds. Distributed crawling is your friend, should you need to move past this reasonable use.
The Crawler type controls the whole execution. It spawns worker goroutines and manages the URL queue. There are two helper constructors:
- NewCrawler(Extender) : Creates a crawler with the specified
Extender
instance. - *NewCrawlerWithOptions(Options) : Creates a crawler with a pre-initialized
*Options
instance.
The one and only public function is Run(seeds interface{}) error
which take a seeds argument (the base URLs used to start crawling) that can be expressed a number of different ways. It ends when there are no more URLs waiting to be visited, or when the Options.MaxVisit
number is reached. It returns an error, which is ErrMaxVisits
if this setting is what caused the crawling to stop.
string
: a single URL expressed as a string[]string
: a slice of URLs expressed as strings*url.URL
: a pointer to a parsed URL object[]*url.URL
: a slice of pointers to parsed URL objectsmap[string]interface{}
: a map of URLs expressed as strings (for the key) and their associated state datamap[*url.URL]interface{}
: a map of URLs expressed as parsed pointers to URL objects (for the key) and their associated state data
For convenience, the types gocrawl.S
and gocrawl.U
are provided as equivalent to the map of strings and map of URLs, respectively (so that, for example, the code can look like gocrawl.S{"http://site.com": "some state data"}
).
The Options type is detailed in the next section, and it offers a single constructor, NewOptions(Extender)
, which returns an initialized options object with defaults and the specified Extender
implementation.
The Options
type provides the hooks and customizations offered by gocrawl. All but Extender
are optional and have working defaults, but the UserAgent
and RobotUserAgent
options should be set to a custom value fitting for your project.
-
UserAgent : The user-agent string used to fetch the pages. Defaults to the Firefox 15 on Windows user-agent string. Should be changed to contain a reference to your robot's name and a contact link (see the example).
-
RobotUserAgent : The robot's user-agent string used to find a matching policy in the robots.txt file. Defaults to
Googlebot (gocrawl vM.m)
whereM.m
is the major and minor version of gocrawl. This should always be changed to a custom value such as the name of your project (see the example). See the robots exclusion protocol (full specification as interpreted by Google here) for details about the rule-matching based on the robot's user agent. It is good practice to include contact information in the user agent should the site owner need to contact you. -
MaxVisits : The maximum number of pages visited before stopping the crawl. Probably more useful for development purposes. Note that the Crawler will send its stop signal once this number of visits is reached, but workers may be in the process of visiting other pages, so when the crawling stops, the number of pages visited will be at least MaxVisits, possibly more (worst case is
MaxVisits + number of active workers
). Defaults to zero, no maximum. -
EnqueueChanBuffer : The size of the buffer for the Enqueue channel (the channel that allows the extender to arbitrarily enqueue new URLs in the crawler). Defaults to 100.
-
HostBufferFactor : The factor (multiplier) for the size of the workers map and the communication channel when
SameHostOnly
is set tofalse
. When SameHostOnly istrue
, the Crawler knows exactly the required size (the number of different hosts based on the seed URLs), but when it isfalse
, the size may grow exponentially. By default, a factor of 10 is used (size is set to 10 times the number of different hosts based on the seed URLs). -
CrawlDelay : The time to wait between each request to the same host. The delay starts as soon as the response is received from the host. This is a
time.Duration
type, so it can be specified with5 * time.Second
for example (which is the default value, 5 seconds). If a crawl delay is specified in the robots.txt file, in the group matching the robot's user-agent, by default this delay is used instead. Crawl delay can be customized further by implementing theComputeDelay
extender function. -
WorkerIdleTTL : The idle time-to-live allowed for a worker before it is cleared (its goroutine terminated). Defaults to 10 seconds. The crawl delay is not part of idle time, this is specifically the time when the worker is available, but there are no URLs to process.
-
SameHostOnly : Limit the URLs to enqueue only to those links targeting the same host, which is
true
by default. -
HeadBeforeGet : Asks the crawler to issue a HEAD request (and a subsequent
RequestGet()
extender method call) before making the eventual GET request. This is set tofalse
by default. See also theURLContext
structure explained below. -
URLNormalizationFlags : The flags to apply when normalizing the URL using the purell library. The URLs are normalized before being enqueued and passed around to the
Extender
methods in theURLContext
structure. Defaults to the most aggressive normalization allowed by purell,purell.FlagsAllGreedy
. -
LogFlags : The level of verbosity for logging. Defaults to errors only (
LogError
). Can be a set of flags (i.e.LogError | LogTrace
). -
Extender : The instance implementing the
Extender
interface. This implements the various callbacks offered by gocrawl. Must be specified when creating aCrawler
(or when creating anOptions
to pass toNewCrawlerWithOptions
constructor). A default extender is provided as a valid default implementation,DefaultExtender
. It can be used by embedding it as an anonymous field to implement a custom extender when not all methods need customization (see the example above).
This last option field, Extender
, is crucial in using gocrawl, so here are the details for each callback function required by the Extender
interface.
-
Start :
Start(seeds interface{}) interface{}
. Called whenRun
is called on the crawler, with the seeds passed toRun
as argument. It returns the data that will be used as actual seeds, so that this callback can control which seeds are processed by the crawler. See the various supported types for more information. By default, this is a passthrough, it returns the data received as argument. -
End :
End(err error)
. Called when the crawling ends, with the error or nil. This same error is also returned from theCrawler.Run()
function. By default, this method is a no-op. -
Error :
Error(err *CrawlError)
. Called when a crawling error occurs. Errors do not stop the crawling execution. ACrawlError
instance is passed as argument. This specialized error implementation includes - among other interesting fields - aKind
field that indicates the step where the error occurred, and an*URLContext
field identifying the processed URL that caused the error. By default, this method is a no-op. -
Log :
Log(logFlags LogFlags, msgLevel LogFlags, msg string)
. The logging function. By default, prints to the standard error (Stderr), and outputs only the messages with a level included in theLogFlags
option. If a customLog()
method is implemented, it is up to you to validate if the message should be considered, based on the level of verbosity requested (i.e.if logFlags&msgLevel == msgLevel ...
), since the method always gets called for all messages. -
ComputeDelay :
ComputeDelay(host string, di *DelayInfo, lastFetch *FetchInfo) time.Duration
. Called by a worker before requesting a URL. Arguments are the host's name (the normalized form of the*url.URL.Host
), the crawl delay information (includes delays from the Options struct, from the robots.txt, and the last used delay), and the last fetch information, so that it is possible to adapt to the current responsiveness of the host. It returns the delay to use.
The remaining extension functions are all called in the context of a given URL, so their first argument is always a pointer to an URLContext
structure. So before documenting these methods, here is an explanation of all URLContext
fields and methods:
HeadBeforeGet bool
: This field is initialized with the global setting from the crawler'sOptions
structure. It can be overridden at any time, though to be useful it should be done before the call toFetch
, where the decision to make a HEAD request or not is made.State interface{}
: This field holds the arbitrary state data associated with the URL. It can benil
or a value of any type.URL() *url.URL
: The getter method that returns the parsed URL in non-normalized form.NormalizedURL() *url.URL
: The getter method that returns the parsed URL in normalized form.SourceURL() *url.URL
: The getter method that returns the source URL in non-normalized form. Can benil
for seeds or URLs enqueued via theEnqueueChan
.NormalizedSourceURL() *url.URL
: The getter method that returns the source URL in normalized form. Can benil
for seeds or URLs enqueued via theEnqueueChan
.IsRobotsURL() bool
: Indicates if the current URL is a robots.txt URL.
With this out of the way, here are the other Extender
functions:
- Fetch :
Fetch(ctx *URLContext, userAgent string, headRequest bool) (*http.Response, error)
. Called by a worker to request the URL. TheDefaultExtender.Fetch()
implementation uses the publicHttpClient
variable (a customhttp.Client
) to fetch the pages without following redirections, instead returning a special error (ErrEnqueueRedirect
) so that the worker can enqueue the redirect-to URL. This enforces the whitelisting by theFilter()
of every URL fetched by the crawling process. IfheadRequest
istrue
, a HEAD request is made instead of a GET. Note that as of gocrawl v0.3, the defaultFetch
implementation uses the non-normalized URL.
Internally, gocrawl sets its http.Client's `CheckRedirect()` function field to a custom implementation that follows redirections for robots.txt URLs only (since a redirect on robots.txt still means that the site owner wants us to use these rules for this host). The worker is aware of the `ErrEnqueueRedirect` error, so if a non-robots.txt URL asks for a redirection, `CheckRedirect()` returns this error, and the worker recognizes this and enqueues the redirect-to URL, stopping the processing of the current URL. It is possible to provide a custom `Fetch()` implementation based on the same logic. Any `CheckRedirect()` implementation that returns a `ErrEnqueueRedirect` error will behave this way - that is, the worker will detect this error and will enqueue the redirect-to URL. See the source files ext.go and worker.go for details.
The `HttpClient` variable being public, it is possible to customize it so that it uses another `CheckRedirect()` function, or a different `Transport` object, etc. This customization should be done prior to starting the crawler. It will then be used by the default `Fetch()` implementation, or it can also be used by a custom `Fetch()` if required. Note that this client is shared by all crawlers in your application. Should you need different http clients per crawler in the same application, a custom `Fetch()` using a private `http.Client` instance should be provided.
-
RequestGet :
RequestGet(ctx *URLContext, headRes *http.Response) bool
. Indicates if the crawler should proceed with a GET request based on the HEAD request's response. This method is only called if a HEAD was requested (based on the*URLContext.HeadBeforeGet
field). The default implementation returnstrue
if the HEAD response status code was 2xx. -
RequestRobots :
RequestRobots(ctx *URLContext, robotAgent string) (data []byte, request bool)
. Asks whether the robots.txt URL should be fetched. Iffalse
is returned as second value, thedata
value is considered to be the robots.txt cached content, and is used as such (if it is empty, it behaves as if there was no robots.txt). TheDefaultExtender.RequestRobots
implementation returnsnil, true
. -
FetchedRobots :
FetchedRobots(ctx *URLContext, res *http.Response)
. Called when the robots.txt URL has been fetched from the host, so that it is possible to cache its content and feed it back to futureRequestRobots()
calls. By default, this is a no-op. -
Filter :
Filter(ctx *URLContext, isVisited bool) bool
. Called when deciding if a URL should be enqueued for visiting. It receives the*URLContext
and abool
"is visited" flag, indicating if this URL has already been visited in this crawling execution. It returns abool
flag ordering gocrawl to visit (true
) or ignore (false
) the URL. Even if the function returnstrue
to enqueue the URL for visiting, the normalized form of the URL must still comply to these rules:
-
It must be an absolute URL
-
It must have a
http/https
scheme -
It must have the same host if the
SameHostOnly
flag is setThe
DefaultExtender.Filter
implementation returnstrue
if the URL has not been visited yet (the visited flag is based on the normalized version of the URLs), false otherwise.
-
Enqueued :
Enqueued(ctx *URLContext)
. Called when a URL has been enqueued by the crawler. An enqueued URL may still be disallowed by a robots.txt policy, so it may end up not being fetched. By default, this method is a no-op. -
Visit :
Visit(ctx *URLContext, res *http.Response, doc *goquery.Document) (harvested interface{}, findLinks bool)
. Called when visiting a URL. It receives the URL context, a*http.Response
response object, along with a ready-to-use*goquery.Document
object (ornil
if the response body could not be parsed). It returns the links to process (see above for the possible types), and abool
flag indicating if gocrawl should find the links himself. When this flag istrue
, theharvested
return value is ignored and gocrawl searches the goquery document for links to enqueue. Whenfalse
, theharvested
data is enqueued, if any. TheDefaultExtender.Visit
implementation returnsnil, true
so that links from a visited page are automatically found and processed. -
Visited :
Visited(ctx *URLContext, harvested interface{})
. Called after a page has been visited. The URL context and the URLs found during the visit (either by theVisit
function or by gocrawl) are passed as argument. By default, this method is a no-op. -
Disallowed :
Disallowed(ctx *URLContext)
. Called when an enqueued URL gets denied acces by a robots.txt policy. By default, this method is a no-op.
Finally, by convention, if a field named EnqueueChan
with the very specific type of chan<- interface{}
exists and is accessible on the Extender
instance, this field will get set to the enqueue channel, which accepts the expected types as data for URLs to enqueue. This data will then be processed by the crawler as if it had been harvested from a visit. It will trigger calls to Filter()
and, if allowed, will get fetched and visited.
The DefaultExtender
structure has a valid EnqueueChan
field, so if it is embedded as an anonymous field in a custom Extender structure, this structure automatically gets the EnqueueChan
functionality.
This channel can be useful to arbitrarily enqueue URLs that would otherwise not be processed by the crawling process. For example, if an URL raises a server error (status code 5xx), it could be re-enqueued in the Error()
extender function, so that another fetch is attempted.
- Richard Penman
- Dmitry Bondarenko
- Markus Sonderegger
The BSD 3-Clause license.